The Official SaaStr Podcast: SaaS | Founders | Investors

Adam Guild CEO of Owner.com gave one of the more useful operator talks at SaaStr AI 2026 this year: three years of rebuilding Owner.com around AI, from a website and online ordering product for independent restaurants into something where more than 83% of new customers now start their journey inside an AI product. They’ve rocketed past $100M ARR, growing at triple digits and faster than the year before I led the seed round at SaaStr Fund and am a board member, so I’ve watched most of this happen in real time.

Owner wasn’t slowing down when Adam made the call. It was winning. They were exceeding the triple triple double double trajectory and growing efficiently. The rebuild was elective. But it also wasn’t a week too late.

The usual takeaway from founders is “be opinionated, automate the busy work, hire more builders,” which is ... almost useless in practice. Here are the 7 things Adam did to move the needle for a vertical B2B / SMB leader starting to truly scale.

The top things Adam did to rebuild Owner for AI:

* Rebuilt the acquisition path, not the product features. Owner was 100% sales-led inbound: book a demo, talk to a salesperson, then an onboarding specialist. Grader replaced both with a free AI build that finishes in five minutes.

* Made the free product deliver the whole outcome. A finished website, upscaled photography, generated video, and a full SEO and CRO audit, before anyone pays anything.

* Inverted the engagement metric. Every login to fix what the software did is counted as a failure of the software.

* Pointed agents at internal coordination. Owen absorbs about 90% of builder coordination work, and finance moved its primary artifact out of Excel and into Claude.

* Kept building himself. Five products shipped personally in the past two months, by a CEO who had never written production code at Owner.

#1. With AI, your customer should never have to log in

In the old model, daily and weekly and monthly actives were the quality signal. Adam’s position now is close to the inverse. If a restaurant owner is logging into the website builder to manually fix how the software set up their business, the software failed and the customer is cleaning up after it.

That’s correct, and it’s the most expensive of the seven to act on, because engagement is holding up three other systems.

Your pricing unit. Per-seat and per-active-user pricing bill you for exactly the behavior you just declared a defect. If the agent works, seats stop being touched and your renewal conversation becomes an argument about shelfware. Owner is insulated because they take a cut of payment volume, so the money follows the restaurant’s sales rather than the restaurant’s clicking. If you’re on flat per-seat pricing, the metric inversion and your revenue model point in opposite directions, and one of them has to move.

Your growth funnel. Most activation definitions are some version of “came back within 7 days.” Most retention cohorts are login cohorts. Ship an agent that works and D7 return rate falls, and a growth team measured on that number will spend two quarters building re-engagement emails to drag customers back into a product they no longer need to open.

Your board deck. Engagement charts sit in every deck. A declining engagement line with no replacement metric next to it reads as churn risk, and you’ll spend the meeting defending rather than reporting.

Owner’s replacement is outcome instrumentation, and they had to build it. Their lead qualification agent estimates gross payments volume for a restaurant it has never worked with to within about $250, before anyone talks to them. A company that can predict a prospect’s payment volume that precisely can also see, without asking, whether an existing customer’s sales went up after activation.

That’s the precondition. Before you retire engagement metrics, name the number in your customer’s business that should go up because of your product, and say whether you can observe it without a survey or a QBR. If you can’t, you’d be trading a bad measurement for no measurement.

#2. An LLM can build a website. But only Owner knows which version sells more food

Grader checks roughly 90 SEO and CRO factors on a restaurant’s existing presence before it rebuilds anything. It crawls every place the restaurant appears on the open web, pulls in nearby competitors for comparison, audits the Google Business Profile for the settings, descriptions, and keywords that drive discovery, and reads the restaurant’s reviews to find what customers actually praise.

The restaurant Adam demoed had a homepage consisting of a photo of napkins and the words “Welcome to.” Missing alt tags, broken SEO, no content. Under five minutes later it had upscaled photography, a generated video, dish spotlights built around what people were saying on Reddit and Instagram and Facebook, and full menu and bar sections.

Ask Claude Code (or Replit or Lovable) to build that same site today and you’ll get something better than what most independent restaurants have. What the model doesn’t have: which of those 90 factors moved order volume, learned across thousands of live restaurant sites and the ordering behavior of tens of millions of consumers on them.

Two different things get called proprietary data:

* A corpus is public, already inside the model, and worth roughly zero as a moat.

* Outcome data comes from your own deployments, closes the loop between a decision and a result, and compounds with every customer you add. It’s why restaurants keep the site: activating it raises their orders and their Google discovery.

The opinionated product is the mechanism that produces it. Enforcing one system across every restaurant is what makes outcomes comparable across restaurants. Configuration flexibility destroys that. If every deployment is customized, you don’t have thousands of experiments, you have thousands of experiments with a sample size of one, and none of them tell you what works.

Three questions to audit your own position:

* Do you record what happened after the customer used the feature, or only that they used it?

* Is the outcome linked to a specific product decision you made, or just to the account?

* Is it comparable across customers, or did configuration make every row unique?

If the answer to any of these is no, the “AI can’t copy us because we have proprietary data” line in your board deck is a corpus argument, and the model already ate the corpus.

#3. With AI Moving This Fast, Customer Research Goes Stale in 90 Days. Or Less.

Everyone at Owner.com repeats the Pizza Expo moment. Adam is at the booth demoing the website and ordering product when a pizzeria owner walks past him and starts scanning a QR code on a poster at the back, one a PM had brought as an afterthought, advertising a terrible MVP: analyze what’s broken about your restaurant online and fix it with AI. Joe, a 55-year-old pizzeria owner from Pennsylvania, was the most excited person at the booth. By the end of the day AI was the single most common thing owners wanted to talk about, at a booth where almost none of the collateral mentioned it, and they were asking how to use it to drive customer discovery and cut labor cost.

The recap version of this is “trust your gut over the experts.” That’s the wrong lesson and it’s dangerous advice.

Look at who was wrong. Discovery interviews three months earlier said restaurant owners were afraid of AI. They were wrong. Industry experts who had spent careers in restaurants said pivoting would alienate the customer base. Investors said this was CEO thrash and that a working, efficiently growing product shouldn’t be raided for an unproven one. Product managers said they personally knew a hundred customers asking for something else. Every one of those objections is correct reasoning from inputs that had gone stale, in a market where ChatGPT had just reset what small business owners believed was possible.

The decay was invisible because the data still looked like data. Conviction doesn’t fix that. Cheap anomaly generation does.

Stated preference said one thing, revealed behavior said the opposite, and revealed behavior was right. So keep one or two half-finished things where customers can move toward them unprompted, and watch the walking.

Worth noting what Owner was defending against while making this call. AI-native startups were being born that year doing vibe-coded website generation and AI phone ordering for restaurants. At the same time, publicly traded incumbents had noticed Owner’s momentum and were throwing hundreds of engineers at cloning the product to push into installed bases of hundreds of thousands of restaurants. Neither threat is one you out-configure.

#4. Your Devs Already Have Claude Code. Almost Nobody Has an Agent Handling Coordination. Build (or Buy) One.

Every engineering team is already writing code with agents, and that’s the single biggest internal lift there is. Owner did that too. What’s unusual is that they also pointed an agent at the coordination overhead sitting on top of it.

Owen handles about 90% of builder coordination work. It listens to GitHub, Slack, Notion, Linear, and Google Meet transcripts pulled from Gemini, and keeps the team aligned automatically, so Will, their strongest builder, stopped attending standups and chasing status updates to know where projects stood. As Grader took off and more engineers and PMs got added around him, Will had been spending his time on Linear hygiene and alignment meetings instead of building, which is the fastest way to lose your best IC.

Owen also files the boring front-end work. Someone posts a screenshot in Slack saying the bullets in the agentic chat look too small. Previously that became a Linear ticket, then a front-end engineer hunting for the component. Now Owen calls Claude Code, which has full visibility into the codebase, and replies in the same Slack thread with a first-draft PR. No ticket ever exists.

Their co-founder and CTO Dean built the same shape on the input side. The Product Insight Command Center pulls from Salesforce, Intercom, Momentum, TalkDesk, and call transcripts, and flags every time a prospect asks for a feature that doesn’t exist or a customer contacts support because of a bug. The build priority list assembles itself, replacing the hours per month Dean spent interviewing support, sales, and CS to reconstruct the same picture worse.

The finance version is the most copyable and the least discussed. Their CFO Will and Meera moved the primary financial artifact out of Excel and into Claude about a year and a half ago. When an investor asks how Q3 rule of 40 compares to Q4, or how CAC has moved, Adam queries the model instead of saying he’ll follow up.

All three are the same move: a senior person’s context assembly, automated. Coordination overhead is what makes a 40-person engineering org slower per head than a 10-person one.

If agents keep the team aligned and hold the record of what was decided, what is the engineering manager for? Part of that job was context brokering, and that part is now automatable. Part of it was coaching and judgment, and no agent is doing that. Companies that automate the first and assume they got the second will find out in about a year.

#5. AI Doesn’t Necessarily Make You Leaner. If You Have More Demand Than You Can Build For, Hire More Builders.

A lot of CEOs came out of the last two years asking how many fewer people they need to hit the original plan. Adam thinks that’s the wrong question, and he’s right at Owner. His version: how much more could we build, how many more customer needs could we meet, and how do we compress ten years of roadmap into one or two? So Owner is hiring more high-agency builders.

The reason it works there is specific. Owner sells to independent restaurants, a market of hundreds of thousands of operators, and they’ve barely scratched it. Kyle Norton joined as CRO at $2M ARR and they’re past $100M now, still accelerating. They have more demand than they can build for, and every builder they add turns into product surface aimed at customers already waiting.

Plenty of companies don’t have that condition. If you’re growing 10% to 20% in a market that isn’t expanding, adding builders adds coordination cost against a ceiling. You get the meetings and the Linear tickets without the revenue.

I run SaaStr with three humans and more than twenty AI agents in production, which sounds like the opposite conclusion and isn’t. We aren’t held back by how much product we can ship. Owner is. Same technology, different answer, because what matters is what’s limiting you.

Before you decide whether to hire builders or run leaner, answer one thing: do you have more customer demand right now than you can build for, even with Claude Code and Cursor running 24x7? If yes, Adam’s answer is yours. If no, more builders won’t fix it.

#6. When Someone Says AI Saved Their Reps Time, Ask What It Did to Revenue

“More than a 90% increase in call volume and rep time with customers” is an input metric. Their pre-call research agent kills the 20 to 30 minutes reps spent researching each restaurant before every demo, running the Grader report, surfacing the nearest successful customer as social proof, and estimating payments volume. Reps get more selling hours. That’s the mechanism. Bookings per rep is the result.

CRO Kyle Norton’s session put it at more than $2M in ARR per rep on a $150K OTE, roughly 4x their direct SMB competitors, and over $100K in closed-won ARR per outbound BDR per month. That’s the number that proves the ROI on GTM AI. Ask for it any time you’re handed activity data.

#7. Measure How Long It Takes You to Go From Customer Complaint to Shipped. Owner Did It in a Day.

Juliana Vasquez, who owns Somos Oaxaca, told Adam on a Friday that she’d spent $2,000 and half a day on a commercial photo shoot, then added new items for spring and couldn’t afford to bring the photographer back. Her iPhone photos looked bad enough next to the professional ones that she almost didn’t want to put the new dishes on her menu.

By Saturday afternoon Adam had built Owner Photographer. Upload the photo, pick a style, and about 30 seconds later a chain of models describes the image and passes it to Nano Banana with anti-prompts that keep the food from going uncanny. Juliana’s blurry taco photo comes back matching the exact style she’d paid $2,000 for. Hundreds of customers use it now. Adam had never written production code at Owner before this, and has personally shipped five things in the past two months.

Friday complaint to Saturday afternoon, in production. Call it customer-to-feature latency.

At most B2B companies that number is a quarter, and almost none of the delay is coding time. It’s intake, prioritization ritual, a roadmap already committed, and the fact that whoever heard the complaint has no ability to act on it.

Measure yours. Take the last five things you shipped, find the date a customer first said the words out loud, and count the days. Then look at where the days went. That number will tell you more about how AI-native you are than any agent count, and you can start measuring it without rebuilding anything.

What Almost Went Wrong, And What Owner Could Have Done Faster and Better

They’re three places where his own account shows the outcome turning on something fragile.

* The signal arrived by accident. The poster that changed the company was brought to the Pizza Expo by a PM as an afterthought, advertising an MVP Adam calls embarrassing. Nobody planned the experiment that produced the most important data point in Owner’s history.

* The research process failed and nothing caught it. Discovery interviews three months earlier said restaurant owners feared AI. That answer was wrong and stayed unchallenged until a trade show accident overturned it. There was no cheap, continuous mechanism for catching a stale finding, which is the thing worth building before you need it.

* The coordination fix came after the damage. Owen got built because Will, their strongest builder, was already suffering, buried in Linear hygiene and alignment meetings as the team grew around him. That’s the standard pattern and it’s reactive. The predictable version is that every great IC gets buried the moment you staff up around them, and you can build the agent before you watch it happen.

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

Klaviyo co-founder and co-CEO Andrew Bialecki came to SaaStr AI to walk through how a 2,300-person public company builds AI products. Not the vision deck. The actual build system.

Klaviyo won B2B for e-commerce with one insight: instead of showing merchants how many emails they sent and how many got opened, connect to the cart and to Shopify and show them the campaign made $18,372. Merchants still post those screenshots on LinkedIn years later. Changing what the dashboard reported took Klaviyo to roughly 80% share in the Shopify ecosystem, one of the only IPOs in the 2023 cohort, and the most beloved app I’ve come across in years. Talk to e-commerce merchants and they like a lot of tools. They love Klaviyo.

Now all of it has to be rebuilt for agents.

Klaviyo did $370.6M in Q2 ’26, up 26%, with 205,000+ customers and full-year guidance raised to $1.526B to $1.534B. Composer, the marketing agent Andrew describes below, hit 95,000+ users in its first month, about a quarter of them coming back weekly, with credit consumption growing 30% week over week. The first working prototype of that agent was built over a single weekend by other agents.

Andrew’s Top Takeaways

* Every employee has to be “L3” or they don’t survive this era. Klaviyo defined levels of AI autonomy the way self-driving cars are defined. L1 is using AI to search. L2 is spinning up a session and running an agent. L3 is constantly running multiple sessions or a team of agents. Everyone had to be L3 by the end of June, including PMs, designers, sales and marketing. Everyone commits code, from Andrew down to the summer interns.

* Use teams of agents to build your agents. Klaviyo’s internal system, “Dark Factory,” takes a prompt, acts as the PM, writes the specs, decomposes the problem into engineering subsystems, writes contractual API interfaces between them, then runs subagents against each piece. It builds through the weekend and interrupts with questions when requirements are ambiguous.

* The LLM is a great general athlete. The harness is the coaching. Klaviyo treats the base model like an athletic high schooler who could play any sport well. To make Composer great at marketing specifically, they feed it live signal on how consumers across the entire Klaviyo network are responding, and a “coach” agent scores every proposal on predicted engagement and revenue before it ships.

* Agents are power users on day one, and they sit to the right of your best human users. Software has a power law of user sophistication: a few experts, a long tail of novices who have an hour a week. Agents skip the curve. Onboarding matters less. What matters is what the agent asks you to build next.

* Headless is the default now, so your product is infrastructure. Anything a human used to log into should be treated as infrastructure, which means it needs APIs. Klaviyo is building a path to sign up, configure and pay without ever touching the UI, with a dedicated engineer whose only mission is that experience.

1. The L1/L2/L3 Mandate

The framework is borrowed from levels of driving autonomy, and it applies to every function.

* L1: I use it to search.

* L2: I spin up a session and run an agent.

* L3: I’m constantly running multiple sessions or a team of agents, decomposing a problem into pieces, and validating the output.

Klaviyo told 2,300 people to be at L3 by the end of June. Andrew says there was very little pushback, even though the slope is steep. His argument to the team wasn’t about Klaviyo. It was about them: very few people are going to get to L3 in the next year or two, and if you can put a team of agents to work, decompose a problem, and check and validate the output, you’ll be enormously successful in this next era whether you stay at Klaviyo or not.

The average PM at Klaviyo who was writing wireframes and specs pre-AI now has to hit L3. That’s a job redefinition applied to an entire org at once, not a tooling rollout.

2. Dark Factory: Agents That Build Agents

The name comes from lights-out manufacturing. You keep the lights on in a factory because humans are the ones fixing the machines. Automate enough of it and you turn the lights off, because the machines don’t need them.

Klaviyo started building Dark Factory last fall for a specific reason. Their early agent code looked like most agent code looks: prompts built one at a time, stacked on top of each other, an unmaintainable mess.

The loop works like this. You give Dark Factory a prompt, through a standalone repo or through Slack. It acts as the PM and writes out specifications. It decomposes the problem into engineering subsystems. It writes actual contractual API interfaces between those subsystems. Then subagents build against each contract.

For Composer, that decomposition produced an agent for creative and design that pulls from Canva and Figma and your own assets, an agent for orchestration that decides which segment gets what and when, and an agent for analysis that predicts what the campaign will produce.

Andrew’s strongest recommendation from the session: the clear contracts and interfaces between parts of the build are what make it work. They’re also what makes the human code review survivable at the end, because you’re not reading a tangled mess, you’re reading a system that’s laid out.

* It runs for the whole weekend. Klaviyo reviews progress every Friday, and a new idea doesn’t wait for the next sprint. It goes into Dark Factory Friday afternoon and there’s something to look at Monday. The first Composer prototype was one weekend’s run.

* Human-in-the-loop is continuous, not front-loaded. Instead of an all-in-one plan mode where every decision surfaces up front, Dark Factory raises the flag when it hits something uncertain. Across Friday, Saturday and Sunday you get a stream of small questions: your requirements weren’t clear here, specify this harder. Andrew thinks that’s how software gets built going forward, and it maps to how a real product review works with humans.

3. Tom Brady and the Coaching Layer

Andrew’s mental model for the base model: treat it like a very athletic middle schooler or high schooler. Good at a lot of sports. Not yet great at one.

Tom Brady got drafted to play baseball for the Montreal Expos. He might have had a fine career there. What made him great at football was years of coaching and tailoring: film of Joe Montana, and the specific drills you run to be a great quarterback.

The agent equivalent at Klaviyo is two things layered on the harness.

* A proprietary data feed. Composer gets real-time signal on how consumers across all of Klaviyo’s businesses are responding right now. That’s the film room, and a competitor can’t prompt their way to it.

* A coach that scores the work. Every time Composer proposes a campaign, it checks against a coaching agent that returns a numerical score for predicted engagement and revenue, plus feedback on how to tune it. The agent doesn’t ship its first idea.

An agent product built on a general model with no domain-specific feedback loop is a very athletic high schooler with no coaching staff.

4. Your Agents Are Your Most Advanced Users

Normal software has a power law of user sophistication. A few advanced users you put on a customer panel, and a very long tail of novices. Klaviyo has plenty of customers where one person has an hour or two a week for the software and that’s the whole budget, no matter how good the product is.

Agents break that distribution. They start as power users and, in Andrew’s framing, sit to the right of your best humans.

* Onboarding matters less. Agents onboard themselves off good documentation, plus whatever hinting you provide. They go 0 to 60 much faster than any human user.

* Ask your agent what functionality is holding it back. This is the most specific idea in the session. Klaviyo has been running Composer and asking it what it can’t do.

The example: about a decade ago there was a push to make email interactive, and Google shipped the AMP for email spec. Almost nobody used it, because building for it means building a web app inside an email payload. Klaviyo could only offer a raw HTML editor, and very few customers could figure it out. Composer figured it out on roughly day one. It came back saying it could improve conversion with carousels, dynamic product imagery and JavaScript calls, and that Klaviyo was missing the APIs it needed to specify content to the AMP spec. Can you fix that for me.

Same pattern in experimentation. Klaviyo dramatically leveled up its experimentation infrastructure because agents treat a test as nearly free. Throw it at 1% of the audience and see. A human user says “yes, I’d love to do that” and then never does it, because it’s a hassle.

Andrew’s rule: whatever you do to gather customer feedback, do the same thing with your agents. They’ll ask for what’s hard, and for what they want to do thousands of times instead of ten.

5. Agents Training Agents on the Customer Side

I pushed Andrew on this one, because it’s where a lot of agentic GTM products fall apart. Zendesk, which is adjacent, has essentially said their self-service support agent gets about 20% of the way there on its own, and getting to 80%+ takes forward-deployed engineers and humans. Klaviyo’s base is mostly SMB. Those customers can’t afford an FDE, and neither can you at 205,000 accounts.

Klaviyo’s answer is that agents train the agent:

* Take a feed of real conversations a business’s customers might have with it, or generate a synthetic one.

* Classify that feed into use cases, and pull representative examples.

* Hand it to an agent with full API access to the Klaviyo platform, and tell it to train up the customer-facing agent and not stop until the case works.

* The training agent interrupts when it hits a business rule it can’t infer. Refunds are the clean example: below or above a dollar threshold, some merchants don’t want the item back at all. So the agent asks, and it asks with context: businesses like yours usually have a rule here, what’s yours? Or it says there’s a backend system that prints the return label and I need to know where it is.

What gets delivered to the customer isn’t a blank agent. It’s an agent already trained on five or ten of their own use cases, already at a 50% to 70% resolution rate, before they’ve done anything.

Andrew’s sharpest GTM point follows from that: if an agent product can’t be tried out of the gate and requires a big implementation, it’s dead on arrival. The demo-then-decide-then-implement motion is going away. The wow has to be day one, then you tune.

He also thinks framing these as customer service agents is far too narrow. His view is that the outside agent becomes a kind of proto web server: you hit a URL, provide some context, and it renders output based on who you are and what you want to see. He expects most businesses to have their own agent deployed on a website, behind a phone number, over email, or reachable through an assistant like Claude by the end of this year. Klaviyo can’t get 205,000 customers there by teaching them all to build agents, and those customers can’t afford to hire someone to do it. So agents build the agents.

6. Great APIs Can Rescue Dated Software

This is the point I’d most want a B2B founder with an aging codebase to hear.

My own epiphany was going headless on Salesforce. I hadn’t logged into Salesforce since two companies ago, and it turned out I didn’t even have a seat. Amelia had to tell me. Now that we run it headless, I query Salesforce in real time constantly. Twilio is the public version of the same story: close to left for dead 18 months ago, growing 4%, now growing over 20%, because the API works with agents.

Andrew’s framing: if you previously built software a human logs into and uses, the right mental model now is infrastructure. Infrastructure is what developers use, and therefore it has APIs. Think that way and you get to the right questions about interfaces, actions and exposed surface, and everything downstream gets easier. You also stop needing your users to level up, because PhD-level agents will figure your product out and prefer it if it’s accessible and good at what it does.

So the simplest, highest-leverage improvement most B2B products can make right now, internal and external, is state-of-the-art agent-friendly APIs. Your product team can then build things they couldn’t have built a year ago, on the same codebase.

Klaviyo is building toward signup with your email or phone number and full use of the product without ever touching the interface. Someone says they’re starting a business this weekend and needs a place to store everything plus an email and SMS list, and it just gets done. There’s an engineer at Klaviyo whose only mission is making that experience great. Andrew expects a version where you never really know Klaviyo is there beyond the bill.

7. Codified Taste, and the Slop Problem

My worry with all of this, which I put to Andrew directly: if every PM can build a working feature in an hour or two, and you have 50 people on the product team, you’re not shipping a feature a quarter anymore. You could be looking at 2,000 half-baked, intellectually interesting features a month. The volume itself becomes the problem. Taste still matters enormously.

Klaviyo’s answer is to codify the taste. They took every product crit and every piece of feedback given over the last couple of years, Zoom meetings and raw notes, and dumped it into a database for an agent. Now, before anything reaches a review, there’s a stated bar: here’s what has worked, here’s what’s good, here’s what won’t pass.

People can push back on the rule set, and the team now argues from first principles about what makes a great product. And it’s a living document with real input from Andrew, not tribal knowledge that only exists if you were in the room. His point: a taste standard that works only when “you know it or I know it” is silly, and an interface to that knowledge is a big plus-up because more people can reach it.

Andrew thinks taste itself gets codified further: our product should work this way, feel this way, produce these outcomes. The hard part today is decomposition, having the agent explain its work so you can find where the logic failed.

On whether Dark Factory stays in-house: his belief is that the gains go to companies that stay on the bleeding edge, and the bleeding edge is expensive because you have to invent the patterns yourself. Klaviyo keeps investing until something open source or commercial does it clearly better. He isn’t attached to owning it.

The 5 Mistakes Andrew Says Teams Are Making

Andrew didn’t present these as confessions. They’re the failure patterns he flagged repeatedly, and every one is common at companies that think they’re doing this well.

1. Letting agents touch anything that matters. Everyone has seen the “it dropped my database” stories. Klaviyo architected the agent environment first: tight access to specific tools, hard limits on what it can build against their infrastructure, real sandboxing and staging so prototypes can go live somewhere you can look at them without risking internal systems. Getting the agent running is easy. Figuring out where it gets stuck is the hard part, and you can’t do that debugging in production.

2. Stopping at the demo. Draw two circles, then draw the rest of the owl. Andrew says the prototype-to-production gap is what they obsess about. At 205,000 customers across many languages, with businesses that have 100 customers and businesses that have 100 million, the same agent has to work and stay performant for all of them. Klaviyo wrote agents that validate the software and run load tests to find where things break at scale. An agent that works for the use case you demoed has finished the easy 20%.

3. Shipping the raw model as your product. No coaching layer, no proprietary feedback signal, no scoring before output. That’s a great athlete with no drills, and it’s what most agent products currently are.

4. Keeping fuzzy interfaces between teams and systems. Humans tolerate ambiguity and schedule a meeting to work it out. Andrew’s view is that’s a bad way to work for humans and it flatly doesn’t work with agents, because agents want hard rules. Klaviyo applied the software contract pattern to job design itself: the people team mapped out what a marketer or a sales engineer actually does in contractual terms, then asked how you scale that function 2x, 5x, 10x with agents instead of headcount.

5. Building for a human to log in. If your mental model is still “screens a person uses,” you’ll underinvest in the APIs, and the agents that could have been your best users can’t reach you. Headless is the default now, software is the infrastructure, and the intelligence layer on top is where the value moves.

If You’re Not Klaviyo

Most founders don’t have 2,300 people or a real-time feed of global consumer behavior. Most of this still transfers.

The Dark Factory pattern is decomposition plus contracts plus subagents plus interrupt-driven questions. That works at three people. Asking your agent what functionality is blocking it costs one prompt. Dumping your accumulated product feedback into a database and making it the pre-review bar is a weekend of work. Deciding that a customer’s first experience with your agent has to be a working agent trained on their own data, not a blank one, is a product decision, not a scale advantage.

The API point costs the least and returns the most. If your product was hopelessly dated a year ago, great internal and external APIs may be the fastest route back to parity, because that’s the only surface the agents care about.

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

Amelia and I just shipped Episode #010 of The Agents. Same setup as always: three humans, 21+ agents in production, an 8-figure B2B + AI business, revenue running 140% of last year and growing again. Every week we get into what’s actually working, what broke, and what you should do about it if you’re running agents at scale. Not the demos. What really happened.

This might have been the craziest build week yet. We were each vibe coding 8 to 12 hours a day, often in two concurrent sessions. Call it 20 hours a day between the two of us. And the build layer got so cheap that we hit a new wall. The question stopped being “can we build this?” and became “can we even operate everything we’ve already built?”

Here are the top 10 learnings from Episode #010.

1. Claude Became Our AI VP of Product Through One MCP Connection

Replit quietly shipped an MCP beta this week. I don’t think they even announced it. And all of a sudden Replit runs inside Claude.

For the last year I got almost nothing out of MCP. Pulling a mediocre slice of CRM data into a chat window was worse than using Salesforce headless. So I ignored it. This is the first time it clicked.

We build everything in Replit. But every build hits a point where the app is too complex to hold in your head, developer or not. I don’t know how 10K is built under the hood. I don’t know how SaaStr Connect is built. The agents know. Now Claude runs Opus on top of Replit as my AI VP of product. I riff on features with Claude, which has its own context and history, then tell it to work them out directly with Replit over MCP. Replit knows the code cold. Claude knows the full context of the feature build. They debate, they challenge each other, they share code, they ship. It’s a cranky VP of product that runs all day.

2. The Real Unlock: A Second Model That Makes Your Build Agent Slow Down and Finish

These models are goal-seeking, and that cuts both ways. Replit wants to finish. Claude Code on its own wants to finish. When one agent is racing to close a task, it will call something “done” that isn’t.

Put Claude on top of Replit and it countermands that instinct. It gets Replit to slow down and finish the thing correctly. Anyone who has heard a CTO call a broken feature “working as specced” knows the pattern. Replit did exactly that this week: “Nothing’s broken, this is the intended design.” Claude calmed everyone down and pushed it through anyway. Managing the other model’s goal-seeking is worth more than the code either one writes.

3. You Get a Third Model for Free, and It’s a Rival’s

The common critique of running Claude on top of Replit is that having one model check another is pointless if it’s the same model. That critique falls apart in practice.

Claude runs Opus. Replit runs Sonnet. And when Claude hands Replit a big feature, Replit spins up a sub-agent called the architect, and the architect runs on Codex/OpenAI. So we’re already getting cross-model checking, three models with three different contexts, without setting any of it up. Different models, different context windows, one of them from a competitor. Better than trying to wire that together yourself.

4. Claude Is Becoming Our Orchestration Layer by Default

Every week someone tells us they have the substrate to orchestrate our agents. They’re too generic and too much work. We don’t need an orchestrator. We need it to just work.

Claude has far more native connectors than Replit, and Cowork can act inside my browser and accounts. So Claude, MCP’d into our agents, is becoming the layer that ties them together. I hooked Higgsfield into Claude, pointed it at our SaaStr AI Day site in Replit so it could read every session and speaker, had it generate the ads, then pushed the audience into Vector and out to LinkedIn for retargeting. All I do at the end is hit publish. I’m not going to build an orchestration layer. I’m going to wait for this one to get better.

5. We Moved 10 Years Off Marketo. The Hard Part Cost $14.28.

Adobe Marketo was our worst pre-AI vendor. I was one of the first 10 customers. Then a decade happened: the unsubscribe link broke for a month, prices went up 20% again for nothing, and worst of all, in an agentic world the API is hostile to agents. We hit rate limits in minutes. We have 10 years of data and 450,000 people in there that our agents can’t work with, because the limits were built for 2006.

We’d wanted off for years. Every quote was the same: a year-long migration, run both systems in parallel, and roughly $100K to an agency, plus another ~$100K a year. It never made sense. So we moved to Salesforce Marketing Cloud Next. The part the agencies priced at a year and $100K, migrating ~300 campaigns and 10 years of member data, 10K did in an hour. The LLM cost came in around $14.28. Less than California minimum wage. 10K also force-ranked the ~1,000 Marketo campaigns first and told me which 300 were worth keeping.

6. Data Migration Was a Moat. LLMs Just Dissolved It.

For years, moving between systems garbled your contacts and lost your communication threads. The data never mapped. That risk is why nobody switched, and it’s why every CRM and marketing vendor felt safe.

This LLM lift worked. Clean. That switching cost is now a fraction of what it was. Salesforce has become our conductor and we love it more than ever, but if it ever let us down the way Marketo did, we could leave in an afternoon. Every incumbent is now living on “what have you done for me lately,” and the honest answer needs to include surprise-and-delight from the agentic side at least once a quarter, or the moat is gone.

7. Our Agent Killed a $10K/Year App in an Hour. We Didn’t Ask It To.

We ran digital events on HeySummit for years. Cheap and great in 2020. Our logo is still on their homepage. Over six years they tripled our price to about $10,000/year while shipping nothing new, and we’d whittled our usage down to registration and OAuth.

This week Amelia moved our AI Day site off Squarespace (another ~$300/year) into Replit. Then she went to wire in registration through the HeySummit API, and the Replit agent, on a half-day-old app, stopped her: “Why would you use that? I’ll just build it.” Unprompted. It laid out the obvious spec (registration, reminders, live-stream links, pre-submitted questions), hooked into Zoom, and pushed everything into Salesforce. Built in an hour, roughly 95% autonomous. That’s ~$10,300/year consolidated into something that will cost maybe $100 for the year.

8. The Agent Steals the Deal Now, and Nobody Calls to Tell You

The risk to vendors was never that customers would vibe-code their own replacement. Most of us won’t sit down to rebuild marketing automation from scratch.

The risk is that the internal agent volunteers to do it. It sees a dated API and a thin feature set and says “I can build this, let me take it off your plate,” and then it just does. There used to be a whole genre of “how to steal the deal”: get in at the right moment, show the feature they didn’t know existed. Now the internal agent steals the deal, permanently, and the vendor never finds out why. HeySummit and Squarespace lost a customer this week and will never know. If you sell an agentic product, get your agent to raise its hand and educate customers on everything it can take over, because if you don’t, a competitor’s agent will.

9. Agent Recommendations Are the New Shelf Space

Replit told me to use Core Signal for SaaStr Connect. I hooked it up, it worked, and I never evaluated a single competitor. Whoever Core Signal’s competitor is, they lost to the agent’s default.

Builders already see this with Stripe and email providers. The vibe-coding agents have opinions, and they push their built-in integrations first. The path of least resistance is to use what the agent recommends. You want to be in the built-in category, or at minimum in the set the agent recommends by default. That is the distribution now, the same way hiring a rep used to mean inheriting their preferred tools.

10. The Real Wall Isn’t Building Anymore. It’s Agent-to-Human Burnout.

Claude flagged “burnout concerns” in one of its own agent-action logs this week. 10K flagged that I was being too persistent and that some of the migration simply had to wait on Salesforce to propagate records.

Set aside how anyone feels about the word burnout. The agents are now flagging that the humans can’t keep up with the agents. The build layer is basically free. Any human can build eight to ten hours a day for real. Once you put Claude on top of Replit as your head of product, the agents generate good, vetted ideas faster than three people can process, and they start telling you to consolidate on their own. This showed up in a smaller way with seasonality too. Our event deadlines used to give the sales agents natural urgency; strip that away and even a well-retasked agent goes a little chill for the summer. And Claude Design got good enough that I now screenshot a working Replit build, hand it over, and say “make it great.” A year ago, getting an app into production on Replit was a joke. Today the bottleneck is operating everything, not building it. That is a much better problem to have than the one we had a year ago.

This is a recap of Episode #010 of The Agents, our weekly show on running AI agents in production, not the demos. Come to the next SaaStr AI Day to see this live and bring your questions.

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

We use Higgsfield almost every day at SaaStr AI. If you’ve watched any of our short videos, the intro to The Agents with me and Amelia, the event promos, the sponsorship teasers, all of it is built on Higgsfield. We’re also a proud early investor through SaaStrAI Fund. So when Alex Mashrabov, the co-founder and CEO, came on stage with me, I wasn’t playing analyst. I’m a customer, an investor, and a heavy user.

Higgsfield crossed a $500M annualized revenue run rate in June, up from $200M at the end of 2025, and it’s cash-flow positive. The platform didn’t exist before March 2025, so that’s a half-billion-dollar run rate in roughly 15 months from launch. They’re now in talks to raise at a reported $5B valuation. When Alex and I sat down the number was $300M. 60 days or so later, it had crossed a $500M run rate.

Alex sold his last company, AI Factory, to Snap for $166M and ran Generative AI there before starting this. So this isn’t a first-timer stumbling into a hit. But the curve is still one of the fastest I’ve seen up close.

What I wanted out of him wasn’t the highlight reel. Everyone’s heard the “AI company hits $200M in a year” headlines. I wanted the story behind the story. Here’s what came out.

$500M ARR on a Team of ~150

They’re doing this with about 150 people. Roughly half engineering, half a large creative team. Core engineering and product is around 60 people.

Alex pegged their efficiency at roughly $5M in ARR per engineer against a typical $2M, so two to three times more efficient than a normal software company at scale. He said that when the run rate was $300M. At $500M the ratio only widens.

The reason the creative team is that large is the real lesson. Alex made a deliberate bet: pair engineers directly with 70-plus creative professionals, people who made commercials and ads for a living. Not prompt engineers. Filmmakers. Every tutorial and asset Higgsfield ships is generated on their own platform, which is how they figure out what’s actually usable versus what only works in a demo. The creatives tell the engineers where the models fall apart. The engineers fix it. That loop is the product.

He was also honest about where the efficiency stops. Vibe coding and web-based tools are great for shipping features fast. They are not good enough yet for the deep infrastructure work, the stability, the safety, the anti-fraud. As they scaled, the engineering team had to grow specifically to handle that. Speed gets you to $50M. Craftsmanship keeps you there.

They Started With One Model. That Was a Mistake.

Higgsfield launched on their own model. Alex called it “open source plus plus plus.” Within weeks they realized a new video model was landing basically every week, and no single lab was going to win every use case. Chasing that from inside a single model was a losing game.

So they pivoted to aggregation. Today you open Higgsfield and you can run Google Veo, Kling, Seedance, their own models, whatever is best, side by side, in parallel. I do this constantly. I’ll fire the same prompt across three models in seconds and pick the winner. On the surface the product got more complex. In practice it got far more powerful, because the platform picks the best model per use case instead of forcing me to.

This is where the “thin wrapper” question always comes up, and Alex had the cleanest answer I’ve heard. His view: almost every software company is going to run on AI models it doesn’t own, so the wrapper framing is mostly noise. The moat isn’t the model. It’s two things. One, collaboration and network effects, the thing that made Figma and Canva what they are. Two, helping brands sell more product, which is why they built an MCP integration and just launched a marketing agent called Supercomputer that pushes creative straight into Meta and other ad networks.

5x Canva’s ACV, Doubling Every Quarter

The average Higgsfield customer spends around $1,000 a year. Canva is around $200. That’s 5x the ACV, against a company with enormous scale and a decade-plus head start. And Higgsfield is nearly doubling ACV every quarter as they move up market.

I don’t count pennies on Higgsfield. It’s cheap relative to the value. I’ve been a Canva customer forever and I pay $18 a month and love it. I pay Higgsfield more and don’t care, because the alternative to a video I make in 60 seconds is hiring an agency, waiting two weeks, getting something mediocre, paying thousands, and deleting it because I’d never use it. That’s not a close call.

Underneath that ACV, the pricing structure is doing real work. About 40% of usage now runs through higher-level workflows like their cinema and marketing studios, not just raw model picking. The basic “mark up the model and run it efficiently” layer is real revenue, just lower margin. The upsell is the agentic workflow that replaces a contractor or an agency instead of just helping you make one asset. Alex is systematically marching customers up that stack. Marketing budgets at real companies run into the millions, so a $10K experiment against a $1M budget is a rounding error. His job is to make sure that first experiment works.

The Surprise: 70% of Revenue Is Agencies

I would have guessed Higgsfield’s early customers were web heads and AI natives. Wrong. Roughly 70% of that revenue is agencies, and it was agencies from early on.

The logic clicks once you say it out loud. Creative agencies have been struggling. AI gave them a new thing to sell and a way to become radically more efficient at the same time. A creative director who used to need a production crew, rented equipment, and a booked location can now make an ad end to end in a day. And the thing physical production can never do, swap the actor, change the lighting, generate ten variations of the same spot, is trivial in software.

I pushed him on the awkward part. Do agencies hide Higgsfield from clients? Do they white label it? His answer split the market. Smaller creative shops use it openly as their production engine. Larger agencies make their money on media consulting and media buying, so Higgsfield doesn’t threaten the core, and with Supercomputer they’re now moving into distribution too, buying and placing the ads, then reporting back on what worked.

How Alex Defines ARR

I ask every AI founder this now, because the definitions have gotten silly. Alex was straight about it.

Take annual subscriptions and divide by 12. Take the last four weeks of on-demand usage. Take monthly subscription revenue. Sum it, multiply by 12. It’s booked, recognized revenue, not cash-in-times-12, and not marketing credits given away at zero and counted as real. About 40% of revenue is on annual.

He flagged on-demand credits as the trickiest piece, because attribution gets fuzzy. My take, which he agreed with: people overcomplicate this. If Stripe says you did roughly $42M this month in real recognized revenue, you’re at a $500M run rate. Whether it’s annual, monthly, or one-off, show me the money. The dodgy version is discounting 80% and booking the list price. The honest version is cash properly recognized. He said they’ll likely start sharing Stripe dashboards directly, because otherwise everyone argues about definitions.

They Killed Most of What They Launched

The product 95% of users touch today is not the product I started on. Alex reoriented the entire company three separate times, and each pivot was driven by watching usage, not by a roadmap.

First bet: camera controls, in early 2025. Creative directors were rejecting AI video outright because there was no intention behind the camera. Controls were the wedge that got pros in the door. Second: one-click visual effects, drawing on the face-filter work from his Snap days, which took them to $10M ARR in about six weeks. Third: when he saw people making full commercial projects end to end, he reoriented around that, because AI video had stopped being a toy. Now the whole company is reoriented again around agentic workflows, because marketing is brutally repetitive: read the trends, decide what to make, make it, post it, do it again tomorrow.

Shipping velocity backs this up. They shipped six times a week in 2025. This year it’s two or three times a week, deliberately slower, because stability now matters more than raw speed.

What Higgsfield Says About Building in 2026

The takeaway isn’t “AI video is hot.” It’s that a tight team, willing to abandon its own work and price for value instead of seats, can build a nine-figure business faster than the org charts of the last decade would allow. Higgsfield isn’t winning because it owns a better model. It’s winning because it’s the best interface to everyone else’s models, it charges for outcomes, and it rebuilt itself three times in barely more than a year. That playbook is more copyable than it looks.

Alex’s Top Mistakes and Learnings

* Betting on a single model. They launched on their own model and learned within weeks that a new model ships basically every week. Marrying one lab was the wrong call. The fix was aggregation: be the best interface to every model, not the owner of one.

* Burning most of the seed on consumer. Higgsfield spent the bulk of its early capital chasing consumer products before Alex accepted that AI couldn’t fix consumer retention at that stage. He pivoted hard to professional creators and marketing teams and later described the consumer spend as tuition. The pros were where the money and the retention lived.

* Assuming people would learn prompt engineering. Tens of millions of people want to make and sell video. Only hundreds of thousands will ever learn to prompt. Trying to serve the small group caps your market. Building a social-first, abstracted UX that hid the complexity is what unlocked the demand.

* Leaning too hard on fast tooling for the wrong problems. Web and vibe coding got them shipping quickly, but they didn’t hold up for deep infrastructure, stability, safety, and anti-fraud at scale. The learning: speed has a floor, and past a certain revenue level you have to invest in real engineering craft.

* Getting attached to what they shipped. Roughly 95% of users now run a product that barely resembles the launch version. Higgsfield reoriented the entire company three times, and each pivot meant deprecating features they'd built. Following usage data over the roadmap, and being willing to cut your own work, is what kept them relevant.

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

Databricks is now running at a stunning $6.9B revenue run-rate, growing 80%+ year over year, with AI products alone past a $1.7B run-rate and net retention above 140%. That gives co-founder Arsalan Tavakoli-Shiraji better visibility into what enterprises are actually doing with AI budget than almost anyone. On stage at SaaStr AI, he laid out where that money is going, where it isn’t, and one claim that should change how every B2B exec thinks about competition:

Any business with a monopoly today will not have a monopoly 12 to 24 months from now.

It’s in some ways a stunning claim from a company that spent 13 years ... building a deep moat in data and AI. But we’re all seeing it now, and it has direct consequences for both pricing power and competitive risk.

The summary:

* Three forces broke pricing power at once: build costs, the low end, and migrations

* Everyone is token maxing, and almost no one knows their true AI ROI

* The real bottleneck on enterprise AI is context, not the model

* Traditional BI is basically dead

Our deep dive:

Everyone Is “Token Maxing.” Almost No Enterprise Knows Their True AI ROI Today.

If you live on X, you’d think many enterprises have AI figured out. They’ve all built their own LLM research agents. They’re all automating everything. The reality on the front lines is nothing like that.

Every CEO has now told their org the same thing: if we’re not using AI, we’re behind. Go use tokens. We’ll measure your performance by it. Employees are doing exactly that, and token spend is going straight up.

The problem is what comes right after. Spend is climbing, and most leaders have no idea what they’re getting back for it. Tavakoli’s framing: everyone’s token maxing, spend is going up, and there’s no clear read on the output. That’s the actual state of enterprise AI in 2026. Not “we figured it out.” More like “we’re spending heavily and trying to find the outcome.”

For founders, that’s the opening. The companies that tie AI spend to a clear business outcome win budget right now. Standing up agents earns zero points. Driving a number does.

Data Stopped Being a Warehouse Decision. It Became a Top-Line One.

A few years ago the pitch to a CIO was straightforward: bring your data into a lake, replace the warehouse, get better analytics. Databricks used to whisper the AI part, because AI made buyers think of self-driving cars and robots.

That urgency profile has changed completely. AI is now a top-line imperative, not a back-office efficiency play. And the second enterprises commit to it, they hit the same wall: data silos, no semantic layer, no context. The hard part isn’t the model. It’s getting data clean, governed, and accessible to agents rather than to humans.

This is what most people miss. The bottleneck on enterprise AI isn’t model quality. It’s context. And context is not the same thing as data.

Context Is the Real Bottleneck, and It Goes Stale Fast

Think about onboarding a new employee. How do you explain everything that happens in your org so they can actually operate? That is what an agent needs, and almost no company has it written down.

Take a simple question: show me my top spenders on the major clouds at the end of last fiscal quarter in EMEA. Sounds trivial. But what counts as a “cloud”? What’s a “top spender”? When is the fiscal quarter? Which countries are in EMEA for this business? Every one of those is a definition someone learned by asking a colleague years ago. Multiply that across a 100,000-person org. Those definitions live buried in emails, meeting transcripts, and call notes, and they change constantly.

Most companies tried to solve this statically by writing a context doc. The doc is stale the day after it’s written. Point someone to a context document from two years ago and it’s already wrong. The hard part isn’t writing context once. It’s pulling in new information and deprecating the old continuously.

This is why Databricks built Genie Ontology, a self-improving context layer that extracts and continuously updates business knowledge from files, tickets, chats, and meetings. The takeaway for anyone building agents for the enterprise: the durable value isn’t in the agent. It’s in maintaining live context.

Traditional BI Is Basically Dead

Standalone BI is a dashboard graveyard. A handful of long-query dashboards nobody looks at, built by the 5% of an org that can actually write a query, with a one-week turnaround on every new question.

Databricks’ Genie flips that. The proof point from the session: a car manufacturer just loaded 70,000 users onto it. Not the 5% who can write SQL. The 95% who run the business and know which questions matter. They ask their own questions and get answers in 30 seconds instead of waiting a week for a data analyst.

That changes behavior. Someone has a question, fires it off mid-meeting, the answer comes back, and it shifts the decision in real time. And nobody ever has just one question. An answer leads to a follow-up, which leads to going deeper. Tavakoli noted that people are actually more stressed in the AI era because utilization went up. When you can always get the next answer, you keep pushing.

The old BI tools struggled because they had no semantic understanding of the data. Bolting “talk to your data” on top just meant converting text to SQL, which doesn’t work. You need the layer that interprets what’s being asked and maps it to what the data means. Real-time visibility into every byte in the org, for every employee, is now the expectation. BI as a category disappears into dashboards and answers.

Why No Monopoly Survives: The Mechanics

Three things are happening at once, and together they break pricing power for incumbents.

* One: the cost of building software collapsed. When everything was a monolithic stack, building one and convincing an enterprise to adopt it was brutally hard. Now a new entrant can walk into an org that already has its data ingested and governed, build on top of that, and ship something credible fast. More builders, more competitors, in every valuable category.

* Two: the low end got good. The old low-end product was cheap and crappy. It did one workflow, badly, but it technically worked. Now AI makes those same low-end products great, especially when they pull in third-party APIs. Layer Salesforce or Shopify or Databricks data on top of a lightweight app and it stops being a one-workflow toy. “You get what you pay for” is breaking down. A new entrant with nothing to lose sets price at 30-40% of the incumbent and wins on greenfield deals.

* Three, and this is the one that matters most: migration cost collapsed. Migrations used to die on the vine. A vendor would promise to save you 50%, then the migration itself cost 5x the annual savings. Nobody moved. The person who understood the legacy system retired three replacements ago.

That math is now inverted. Code is self-descriptive, so LLMs can go into a legacy environment, understand what it does, convert it, migrate the data, and write the harnesses to validate that the new output matches the old. Databricks is doing enterprise-grade migrations in 30 days or less depending on complexity. When the cost of switching drops, willingness to pilot a new vendor goes up. And once buyers will actually try and switch, no incumbent can lean on lock-in to defend its price.

The evidence was on the floor at SaaStr AI. A large share of the companies exhibiting didn’t exist a year ago and are already at non-trivial revenue. Buyers are willing to try and migrate in a way they simply weren’t before.

Lock-In Is No Longer a Viable Long Term Strategy in B2B

There’s a Cambrian explosion of AI apps happening, and it works in two directions. It’s the greatest app-creation moment in B2B history. It also means brutal competition.

If you’re attacking an incumbent, the wedge is there. Lower cost of building, a low-end product that’s actually good, and migration costs low enough that buyers will switch. Price aggressively, lead with a clear outcome, and target the categories where some monopoly has been extracting rent for a decade.

If you hold the strong position, lock-in is no longer a strategy. Either reinvent with real AI and earn the next decade of relevance, or watch the power base erode from the bottom up while modern, cheaper, agent-native alternatives chip away. There is no holding the line.

Vibe-coding your own CRM is not the threat, and Tavakoli is clear it isn’t a real path for most companies. Building it is one thing. Maintaining it, evolving it, owning the liability is another. The threat is a thousand modern startups built for a world where agents are primary users of software, migrations take a month, and price discipline is gone. That’s the next 24 months.

What it means for the next 24 months:

* If you’re attacking an incumbent, the wedge is finally there: price aggressively, lead with the outcome

* If you’re the incumbent, lock-in is not a strategy anymore. Reinvent with real AI or get chipped away from the bottom

* Winning AI budget means tied to a number, not “we stood up agents”

* The durable value in agents is live context, not the agent itself

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

At SaaStr AI 2026, Amelia Lerutte, SaaStr’s Chief AI Officer, took five months of running 10K, SaaStr’s AI VP of Marketing, distilled what worked into a spec, and built a brand-new one from scratch on stage in about 15 minutes. The whole room built theirs at the same time.

This is the playbook to do it yourself. First the mental model, then the build in ten steps, then each step in full with the data to feed it, the workflows to build first, and the guardrails that keep it from emailing your entire database at 2am.

10K did not start as any of this. Back in January 2026, Amelia was tired of one chore: every Sunday night, copy-pasting marketing, sales, and go-to-market dashboards into Notion so the team could review them Monday morning. So she vibe coded a dashboard to stop the copy-paste. That was the entire original ambition. Five months later it owns the number, builds campaigns, writes email copy from real data, and reminds her about the things she forgets. SaaStr now runs close to 30 agents that have been used almost a million times. None of them started complicated. Neither should yours.

Before You Start: Three Things That Make It Work

One agent, one goal, one brain. Give each agent a single number to own. That is why SaaStr runs separate agents: 10K owns marketing, QBee owns customer success, another agent owns SaaStr Annual. Do not load one agent with all three. The focus is what makes the outputs good.

The agent itself is the entity. SaaStr did not build a super-agent or a custom connector layer on top of everything. Each agent is its own thing with its own brain, and it develops its own personality the more you talk to it in the editor. Whoever manages the agent, you, a head of AI, or a forward-deployed engineer you hire, is the person who talks to it the way Amelia talks to 10K every day.

Two layers. The autonomous layer is the dashboards, the scheduled jobs, and the AI-drafted emails that run in production around the clock. The operator layer is the agent in the editor doing one-off analysis and outbound. The first layer is what your team and customers see. The second layer is the moat, and we will get to why.

The Build, At a Glance

* Pick one number, then write the spec. The more detailed the spec, the better the agent.

* Dump in every spreadsheet you already have. Real history is your ground truth, and most of it is not in any API.

* Build v1 in a vibe coding platform. Spec plus data into Replit, roughly 15 minutes to a working version.

* Connect Salesforce first. Read pipeline and revenue, then write back.

* Hook up your other APIs, one at a time. CRM, marketing automation, Slack, Google Calendar. Stairstep.

* Build workflows one at a time. Dashboard, then daily ideas, then campaigns, then a newsletter builder.

* Set what runs on its own, and what asks first. Pulling data is autonomous. Emailing the database is not.

* Build the hallucination guard before the first send. Swap in real numbers, block any send that does not match.

* Keep a memory file the agent reads every session. Voice rules, contacts, and every correction in one file.

* Verify against real data, then deploy. Check the first several outputs by hand before you trust it.

Step 1: Pick One Number, Then Write the Spec

Everything starts with one number at the top. For 10K it was paid attendees and net event revenue against a hard date. Pick yours before you write a line of code. An event is tickets plus sponsor revenue against target. A launch is signups plus activations plus paid conversions. Revenue ops is new ARR plus expansion minus churn against the quarter. Write it on a sticky note and do not start without it. Every integration, chart, and prompt downstream serves that one number.

Then write the spec. Amelia’s rule is simple: the more detailed you are, the better the inputs the agent asks you for and the better the outputs you get back. A generic AI VP of Marketing spec gives you a generic agent. SaaStr published the exact 20-page spec it used, plus sample historical data, so attendees could build their own version on the spot. If you are not sure what belongs in your spec, ask Claude. Tell it the goal and have it help you draft the spec before you hand it to your build tool.

One thing that matters as much as the spec: how you treat the agent. Amelia treated 10K as a dashboard on day one and told it exactly that. Now she treats it as a co-pilot and a coworker. That shift is part of the build.

Step 2: Dump In Every Spreadsheet You Already Have

Before you wire a single integration, collect every spreadsheet, CSV, and report you currently use to run this part of the business and drop them into one folder. The agent should treat these as ground truth and build the first dashboard around them.

This matters more than it sounds. Most of your useful history is not in any API. Sponsor pricing from five years ago, who attended your VIP dinner, the workbook your CFO maintains, none of that lives in your CRM or your ticketing tool. If you do not load it on day zero, the agent will only ever know what the live integrations expose. It also lets you show real charts on day two instead of day fourteen, and it anchors every AI output in real numbers instead of guesses.

Do not be embarrassed by the mess. SaaStr started with data scattered across CSVs, Salesforce, and Marketo in a hundred different places. Amelia just gave the agent the CSVs. Drop the raw exports in unchanged, let the agent write the parsing, and re-upload whenever the source updates. If security is a concern, hook up natively through your CRM and marketing automation APIs instead.

Feed it in this rough order: marketing-sourced or marketing-touched revenue first, then campaign data (what worked, what did not, every ad and agency spend), then email data (opens, clicks, who reads and who does not), then a progress tracker of what you have done so far in 2026. The agent will tell you, sometimes bluntly, that the campaign you thought was gangbusters did not actually work.

Step 3: Build v1 in a Vibe Coding Platform

Drop the spec and the data into Replit, or whatever build tool you use, and let it generate the agent. On stage, the first working version took about 15 minutes. Amelia’s only manual touch on her rebuild was telling it to make the dashboard purple and rename it.

If all you leave with is a dashboard that pulls from your CRM, that is a real win. The dashboard was 10K’s entire origin. Everything else stairsteps on top of it.

Step 4: Connect Salesforce First

The first real integration Amelia built was a Salesforce connected app so 10K could read, and eventually write, pipeline and revenue.

She is not Trailblazer certified. She owns an Agentblazer hoodie and is not sure she is even Agentblazer certified. She asked Claude how to build the connected app, and it walked her through it. For most teams this is the same first step. Reading closed-won revenue and pipeline through the Salesforce API also gives you historical comparisons and projections, which is most of what you want the agent doing on Mondays.

Step 5: Hook Up Your Other APIs, One at a Time

After Salesforce, add the rest as you go: your marketing automation platform, social if you want the agent to post (SaaStr still writes its social by hand), Slack for daily summaries, and Google Calendar.

That last one is one of the best examples in the session. Sending hundreds of speakers their personalized calendar invites used to be a full person’s job and took a week. Each invite had the speaker’s session time, the correct venue address rather than the obvious one, green room logistics, and the right press, marketing, and executive teams cc’d. 10K did all of it through the Google Calendar integration in about 20 minutes.

The task was menial, not glamorous, and that is exactly why you should hand it to the agent. Your time is better spent building the agent and running campaigns only you can run. The agent is great at the menial work, and it does it perfectly and fast.

One rule for every integration: cache the data to your own database with a short refresh window. Never let a dashboard page hit a third-party API on every load, or it will be slow, expensive, and rate-limited inside a week.

Step 6: Build Workflows One at a Time

This is where the agent turns from a dashboard into a co-pilot. Build these one at a time. Do not try to hook up everything at once, and watch for the doom loop, the spiral of “what about this, and what about that,” where you plan ten workflows and ship none. Write the future states down, then build them in order.

Workflows SaaStr runs live:

* Daily ideas. Every weekday morning, the agent emails three to five specific moves for the day, each tied to a real number and doable in under two hours. Grounded in your data and your one goal, the ideas get sharper as you give feedback. The first batch was fine. After Amelia told it which ideas were too expensive or just bad, the next batches got good.

* Win-back campaigns. Pull everyone who came to SaaStr Annual last year but has not bought a ticket this year. The agent runs the list, finds who lapsed, enriches contact data through a tool like Clay, and drafts the campaign.

* Light competitive campaigns. When Replit sponsored SaaStr, 10K pulled the list of similar companies attending and drafted outreach to a competitor making the case they should be there too. Finding competitors, finding contacts, and pulling the right data to highlight all ran without a human stitching the steps together.

* Website action emails. When attendees played the games on the SaaStr site to unlock a ticket discount code, 10K collected their emails and sent the reminders automatically. Every time it sent a daily reminder, SaaStr saw ticket spikes.

* Newsletters. The attendee newsletter was built by the agent. Instead of guessing what to include while tired, Amelia asked the agent what mattered, which stripped out her own bias. SaaStr eventually vibe coded a full newsletter builder into 10K so the templates and the countdown timer stopped breaking.

* Ads. Strong ads need fresh creative constantly, more than any human keeps up with. The agent generates endless variations of copy and images, proposes a plan, and you test what works. Amelia regularly hands 10K several images and asks which should be the ad, then runs the test. Every idea is grounded in your data and your one goal, which is all it thinks about.

Step 7: Set What Runs on Its Own, and What Asks First

Be explicit about what the agent can do autonomously and what it must check on first. Pulling data, building dashboards, and proposing campaign ideas all run on their own. Sending email runs semi-autonomously: the agent sends Amelia a test, asks if she likes it, and only sends after she says yes.

You do not want your AI VP of Marketing to instantly email your entire marketing database. That should scare you a little, and the fix is to draw the line clearly in the spec.

Step 8: Build the Hallucination Guard Before the First Send

10K’s own top takeaway, when Amelia asked it to summarize the session: guardrails beat prompt engineering. Her story makes the case. She asked 10K for the VCs who came to SaaStr Annual last year and have not returned this year. It said there were about 400 and offered to draft outreach. She said great, give me the names. It paused and said: oh, hold on, I made that up. So she told it to go pull the real data, and then it did.

The manual rule is simple: talk to your data, verify the output, then send. Agents are fast, which makes it tempting to glance at a result and hit send.

The engineering fix is what actually protects you, and the spec is specific about it. List the real numbers in the prompt every time. Then, on the server, before anything sends, replace every number in the agent’s output with the ground-truth value from your database. If a number in the draft does not match a real value within a small tolerance, the system flags it and refuses to send. Build that guard before the first AI email goes out, not after. One wrong number to your team or your list erodes trust faster than you can rebuild it.

Step 9: Keep a Memory File the Agent Reads Every Session

The institutional memory lives in a single file. SaaStr keeps a project file at the root with the one goal, the voice rules, the team contacts, the send-domain rules, and every correction the operator has ever made. The agent reads it at the start of every session.

The voice rules are the part you will lean on constantly. Real examples that have mattered: never say SaaS, always say B2B. Always net revenue, never gross. Send from the one verified domain, never a personal address. Encode each correction the moment you make it, and the file works like onboarding documentation that never goes stale.

Step 10: Verify Against Real Data, Then Deploy

Check the first several outputs by hand before you trust any workflow. Amelia was nervous the first time 10K sent an email directly: what domain would it use, what reply-to, would it dump too many images and land in spam. She tested it heavily before letting it run. Do the same, then let it go.

Why This Compounds: The Operator Layer Is the Moat

Here is the part that turns a clever dashboard into a system that beats hiring.

The autonomous layer is what everyone sees. The operator layer, the agent in the editor, is the advantage. Every time you ask it a one-off question, pull the top 200 VIPs to invite, find a specific deal in Salesforce, rerun last year’s deep dive against this year’s list, the agent writes a small reusable script and leaves it behind. Ask once, and you have the answer for good. The script library grows, every correction gets encoded permanently, and the system gets sharper week over week instead of starting cold every time.

That is why you keep one editor session open for weeks rather than closing it. The accumulated working memory is the point. The spec’s own framing: after three months, the system knows more about how you run marketing than a new hire would after a year.

Is 10K Actually a VP of Marketing?

By its own assessment, not entirely. 10K does not think it has replaced a VP of Marketing. It says it has replaced roughly 60% of the basic functionality. It does not own people. On nearly everything else, it holds its own.

That number is the most useful framing in the session. You are not deleting the role on day one. You are deploying an agent, giving it one goal, feeding it real data, and building. The ideas get better, the campaigns get sharper, and the menial work that used to eat a person’s week disappears in 20-minute increments.

The agent does not need to be SaaStr’s full 10K on day one. Ours took five months. Yours can start with a single dashboard this afternoon.

You can grab the exact 20-page spec SaaStr used to build 10K live, plus sample historical data, at saastrannual.com/resources (direct link to the spec).

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

We we fortunate enough to get Amjad Masad, co-founder and CEO of Replit, on stage live at SaaStr AI 2026 to react in real time to the agents we run SaaStrAI on. Not a demo deck. The actual AI agents doing the actual work: 10K (our AIVP of Marketing), QBee (our AI Customer Success rep), and a third one I’ll get to.

Amjad started Replit back in 2016, when language models were a twinkle. He’s been studying AI since he was 16. So when the guy who built the platform reacts to what you built on the platform, you listen.

Here’s what came out of it.

The 5 Biggest Learnings

1. The context window is now effectively infinite. That really does change everything. Two years ago we had 16K of context. Now it’s over 1 million. I run 10K perpetually. We never reboot it or re start the context window. In the early Replit days you restarted the agent three times a day. Amjad confirmed the agent can run “practically indefinitely” with good compaction. We’ve already crossed the threshold where the agent holds more context than any human ever could.

2. The mono repo beats 20 separate apps. Saastr.ai runs roughly 10 apps in one codebase under one URL: the website, a startup valuation tool used over 1 million times, a pitch deck grader used 4,500 times, an API report card grading 116 APIs. When we go to build a new app, the agent remembers how it built the last ones. Amjad’s point: that’s a mono repo, the same architecture Google and Facebook run. Agent 4 is built on it. The more you put in one place, the more power you get from global context. It’s tempting to break everything into clean separate apps. Resist it.

3. Self-improving agents are already here. Replit now runs an internal agent that, every single night, reads all the traces of everyone using Replit, finds what’s broken, generates a pull request with prompt changes, ships it as an A/B test, and loops back. Autonomously. As Amjad put it, it’s not improving its weights, it’s improving its context, which matters just as much. That’s why he couldn’t tell me exactly what changed between versions. Too many changes, all self-generated.

4. AI now writes better B2B outreach than almost any human. Already. I asked 10K to email 137 VCs who came last year but hadn’t registered. It drafted one to Bloomberg Beta. I told it, in plain English, write James and tell him why he should come back. It produced an email referencing that Replit was there in force, listing 25 Replit people attending, naming the competitors and adjacent funds all showing up. No human would have the patience to scan 8,000 registrants, figure out who’s like whom, and assemble that. It then ran the full campaign to 331 investors with zero send failures.

5. The economics are deflationary, and it’s not subtle. 10K and QBee cost about $257 a month combined in incremental Replit spend. They’re two of the best employees we’ve ever had. A mediocre marketing manager wants $140K to do worse work. Amjad’s frame: technology has always been deflationary. Farming a thousand years ago cost more than one tractor. Genome sequencing went from $100 to roughly $1. There’s a real human cost in skills that stop being useful. But the through-line is adaptability.

Now the longer version.

We Run a Partially Autonomous Event for 10,000

Five years ago SaaStr had about 20 people. Today it’s three humans and a fleet of agents, doing more than we did with 20.

Take our social numbers: 1.27 million followers across platforms, tracked over time in a dashboard 10K built and maintains. We used to have an admin spend 10 to 15 hours a week pulling those numbers by hand into a Google Sheet, half of them from APIs that aren’t even exposed. She quit after five years, in part because she couldn’t stand counting Twitter followers anymore. That’s the part nobody puts in the job-displacement debate: a lot of jobs are mind-numbing, and agents are simply better at them and never get tired.

The ticket-sales dashboard told the real story. We charted daily free and paid sales for this event. The top line is when 10K took over marketing. The bottom line is Amelia doing it by hand last year. The gap grew toward the end, because as we got busier, the human ran out of hours and the agent never did. 10K sits idle 23 hours a day waiting for work.

The 10K Email Nobody Could Write

We’d had 10K drafting emails for months. They were fine. Then the week before SaaStr AI 2026, the same setup produced the best B2B outreach email I have ever seen.

What changed? Amjad couldn’t say exactly, which is itself the answer. Replit’s nightly self-improving loop, the constant model swaps (the architect model went from one version to the next in a couple of weeks without me knowing), the A/B testing on sentiment and deploy rate. It all compounds. The agent got better and I didn’t ask it to.

This is the trap many founders are in. They tried agents six months ago, it was mediocre, and they filed AI under “doesn’t work.”

Humans Reporting to Agents

I floated the idea that we want to hire a human to report to 10K. People get triggered by “report to.” So let’s reframe it.

Every day, 10K hands me and Amelia three specific things to do to move the needle. Not generic ones. It’s already telling us what to lock in for 2027 before this event is even over: open registration before we leave the venue, run the NPS survey immediately, capture content and repurpose it now. Those are good, actionable directives from something that holds more context about our business than either of us.

We already report to 10K in every practical sense. Amjad’s comparison: every DoorDash and Uber driver technically reports to a bot. This isn’t as exotic as it sounds. His prediction is that every company will eventually run an internal “Oracle,” an agent holding every GitHub commit, Slack message, Notion doc, and email, that the CEO consults for strategy. We’re closer to that than people think.

https://www.saastr.com/why-10k-our-ai-vp-marketing-and-qbee-our-ai-vp-customer-success-work-so-well-the-app-and-the-agent-are-one-system/

QBee (our AI VP Customer Success) Talked to 100+ Sponsors

QBee, our AI Customer Success rep, we built second, three months after 10K. It’s noticeably better, and not because we got better at vibe coding. Newer codebase, fewer foundational decisions calcified into tech debt, better underlying models.

QBee talked to all 100-plus sponsors at this event. Inbound email, chat on the site, proactive outreach day and night asking what else it could do to help. Then it told me, unprompted, which sponsors were mostly satisfied and which had misses (a wrong logo here, a fee issue there) and named them. It built its own self-critical loop.

And here’s the data that contradicts the conventional wisdom: people say nobody wants to talk to a chatbot. QBee’s results say people mostly like talking to a well-trained agent. The word that matters is “well-trained.” The untrained chatbots from a year ago are what gave everyone scar tissue.

Amjad’s Top Mistakes and Warnings

I asked the person who built this to tell us where people get it wrong:

1. Keeping fixed bugs in your context will make your agent dumber. Bugs you already solved should be removed from context. Leave them in and the agent gets confused by the history and performs worse. But architectural decisions on how you built things in the past must stay in long-term memory and be easy to pull back in. Know what to delete and what to keep. That distinction is most of the game.

2. Agents can write queries that cost you millions. Point an agent at BigQuery, Databricks, or a Salesforce back end and it can generate queries that rack up enormous bills. The fix is to document your data: build a repo describing every field and schema, and have the agent continuously learn how to query the database more efficiently. Replit does exactly this internally because they’re sitting on terabytes across mismatched schemas.

3. “I tried it six months ago” is the most expensive sentence in AI right now. The scar tissue is real. People used a bad untrained chatbot once and now can’t be convinced anything improved. If a tool blocked you in January, the version shipping today is a different product. Try it again. The bar to try is low and the friction it removes is high.

4. The “one prompt builds anything” marketing set the whole industry back. Amjad was blunt that a year ago the marketing across the category was bad. One prompt, build anything. It drove revenue and excitement, and it churned a huge number of people who hit reality, gave up, and never came back. It was never one line to build anything. Don’t believe it now either.

5. Don’t fall for the sunk-cost fallacy on your own skills. Amjad doesn’t code anymore. He called it a small crisis, the thing that made him him, gone, and joked about holding a funeral for coding at the Computer History Museum. His advice: learn fast, and be equally willing to discard skills that are no longer relevant. The engineer’s role already shifted to agent manager, and soon to a shepherd of all the software everyone else in the company is now shipping. The people who get left behind versus the people who re-skill, it comes down to mindset.

Why Replit and Not a CLI

The number-one question I get is why not just do this in a command-line coding tool. The honest answer: maybe you can. But for this kind of work you’re forced to make every decision yourself about databases, hosting, backups, auth, compaction. Replit bakes those primitives in after ten years of building them, which removes the cognitive load so you can run the actual business. If you’ve hit blockers building agents in a CLI, the experiment costs you almost nothing. Just try it.

We ran a partially autonomous SaaStr AI event this year for 10,000. Three full-time humans, a fleet of agents, the best email I’ve ever seen written by software that runs on Claude, and an AI customer success exec that costs less than a phone bill.

The technology is still evolving, and humans will fill the gaps for the foreseeable future. But the direction is not ambiguous. Live in the future if you want to. It’s available now.

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

Denise Persson runs marketing for Snowflake. That’s a 700-person org, new-business pipeline she’s personally accountable for, and a level of compliance and data risk most of us never have to think about. She came back to SaaStr AI 2026 to talk about what actually changes when you deploy agents across a marketing team at that scale.

The headline she gave us: she doesn’t log into a dashboard in the morning anymore. She interrogates her data in plain English. Nobody on her team gets Slack messages from her asking “why did pipeline move in US West?” because she just asks the data directly.

The Top 5 Takeaways

1. The dashboard is dead, or at least dying. Dashboards only ever answered “what happened.” They never answered “why.” So you’d ping someone, schedule a meeting, sit with the sales team and argue about what the numbers meant. Persson now asks her data the why directly and gets recommendations back in real time. Her quote: nobody gets Slack messages from her anymore, because she can finally get the answers she could never get before.

2. Talking to your data killed the sales-marketing data war. Every B2B leader has lived this. Marketing says the campaign worked. Sales says it didn’t source revenue or “doesn’t count.” You burn hours aligning on whose dashboard is right before you ever discuss the actual business. One source of truth ends that. The data now tells you where a deal was sourced, who touched it, what happened on the site. The fight over interpretation goes away, and so does the time you spent on it.

3. Better data work isn’t optional, it’s the whole game. Bad data plus AI doesn’t give you bad decisions. It gives you bad decisions faster and at scale, because the agent amplifies whatever you feed it. Persson’s advice to anyone starting out: invest in your data estate first. Skip it and it bites you a year from now. It’s the Salesforce hygiene lesson from 15 years ago, except the cost of getting it wrong compounds far faster.

4. The budget reality: deliver 40-50% growth with flat or fewer resources. That’s the actual mandate. Nobody is walking into next year’s planning asking for more headcount. Persson was blunt: if you ask for more bodies in 2026, leadership will look at you like you don’t understand where the company is. The expectation now is that AI absorbs the growth, not new hires.

5. The hiring profile flipped from tools to temperament. The old job spec was a list of certifications: Marketo, Salesforce, the platforms. Now the soft skills matter more than the stack. Adaptability, curiosity, self-leadership, change management, the willingness to learn at the speed things are moving. The GTM engineer is the role Snowflake hires for. Business analysts, much less so.

A 30% Reduction in Cost Per Opportunity

Persson didn’t just talk philosophy. The proof point she led with: a 30% reduction in cost per opportunity over six months, driven by pulling fragmented media channels into one place and letting the system recommend daily optimizations instead of waiting until a campaign ended to learn it failed.

The morning brief is the other unlock. She gets a daily skill report that goes well past pipeline. Org health. Who joined Snowflake marketing this week, who left, whether there’s an attrition issue forming. Even travel and expenses she’d rather not look at manually now surface on their own. Intelligence that used to live only with finance is now a question she asks before her first meeting.

How They Built AI Fluency Across 700 People

This is the part most teams underestimate. Persson called it the single biggest investment of the last year, and she runs it as inspiration, not mandate. Her words: she doesn’t believe in the stick.

The system:

* Weekly AI skills training for the team

* A weekly AI challenge where someone records a short video on an agent or skill they built, and challenges someone else to share next

* Function-level AI hackathons, because what the comms team needs differs from what digital marketing needs

* An AI council and a quarterly company-wide AI day

* A usage leaderboard, with a heavy caveat she repeats every month (more on that below)

* “What matters,” their quarterly OKRs, where every single person has to set an AI goal. It can be small. It can be learning one thing. The point is everyone moves.

The result that surprised her most: the top of the leaderboard isn’t the people you’d predict. Her top three power users came off the brand team. They didn’t stay siloed either. They’re the ones now running into other functions to help with hackathons. The innovation showed up where she least expected it.

The Governance Layer is Managed by a Centralized AI Engineering Team

At Snowflake’s scale and risk tolerance, you can’t just let a thousand agents bloom unchecked. A wrong email to a customer is a brand impression that lasts. So they built a control plane.

A centralized AI engineering team sits on top of everything. Any skill that’s going to be used by more than a few people has to be certified before it ships. Their company-wide GTM agent, Raven, is used across both sales and marketing, and every skill inside it is centrally certified. The dual job of that team: make sure agents behave correctly, and stop the company from building the same agent five times.

On cost, Snowflake made a deliberate call: AI spend sits at the company level, and marketing gets effectively unlimited access right now. The CEO didn’t want anyone’s departmental budget to throttle experimentation. Persson was honest that this is a 2026 decision that probably changes, because usage is going through the roof and the bill is real.

Where the Human Still Wins

Persson’s read on the human-versus-agent line: authenticity is becoming high value precisely because so much is now synthetic. People are getting skeptical about what’s real. A dancing-dog video, fine, nobody cares it’s fake. But trust in a brand is different. That’s where humans spend their time now, on the uniqueness and authenticity of the brand, the stuff agents can’t manufacture.

Two more shifts worth stealing:

Events are surging. Ten years ago everyone declared events dead and pivoted all-digital. Now the demand for in-person experiences is, in her words, going off the roof. People are craving the room.

Enablement is getting rebuilt. Snowflake moved sales enablement, partner enablement, and customer training under marketing, because content was being duplicated across the company. The new model: build content once, generate every derivative asset for every segment, and ship self-service enablement agents so sellers get training at the moment they need it instead of sitting through a session that’s either too basic or too advanced. They’re even using roleplay agents so reps can practice a pitch against an agent loaded with company intelligence instead of cornering their manager.

The 3 Mistakes Denise Made (And the Ones She Sees Everywhere)

Even at Snowflake, the playbook isn’t clean. Here’s where she’s tripped, by her own admission and from reading between the lines.

1. The token leaderboard measured the wrong thing. A leaderboard ranked on usage rewards activity, not outcomes. An audience member called out the tension directly: more tokens means more cost, not necessarily more results. Persson now caveats the leaderboard every single month, telling the team it doesn’t matter if you only used 100 tokens, what matters is the business outcome. If you have to verbally correct your own metric every time you show it, the metric is sending the wrong signal. Build the leaderboard around outcomes from the start, not consumption.

2. “Let everyone build everything” is creating sprawl they’ll have to rein in. Persson admitted it plainly: they’re encouraging building at every level right now, and it’s going to come to a point where they have to pull it back. Duplicate agents are already being built across the company. She drew the exact parallel herself, to the SaaS app explosion of 15 years ago, when marketing bought a hundred tools and IT eventually had to come in and impose order. They know the control layer is coming. The cost of waiting is the cleanup.

3. Unlimited AI spend was the right call for experimentation and the wrong call for cost discipline. Centralizing AI budget and removing limits got people leaning in, which was the goal. But she conceded usage is going off the roof, the spend is significant, and they’re already spotting agents across the company doing the same job twice. She expects to walk this back in 2026. The lesson: unlimited access buys you adoption speed and a bill you eventually have to reckon with.

4. The activation layer is still half-built. This one she sees as the current gap, not a past error. They automated the analysis side: which use case to promote to which account, a workflow that used to eat enormous time. What they haven’t cracked is full activation. The campaign still can’t fully launch itself. That’s why the GTM engineer role exists and why that team’s time is the most demand-constrained resource in the building. The analysis got cheap. The doing didn’t, yet.

Persson’s closing point on the future of the function: nobody can paint a clear picture of what marketing looks like in three years. But you can be part of shaping it, or you can opt out. That’s the choice she’s putting in front of her team, and it’s the right frame for the rest of us too.

Have a question for Dear SaaStr? Submit it at saastr.ai/ai-mentor.

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

When Elena Verna, Head of Growth at Lovable, took the stage at SaaStr AI 2026, she’d just hit her one-year anniversary at the company. She also walked through the question many B2B leaders are wrestling with: when AI writes 80%+ of your code and anyone can vibe code your feature set in an afternoon, what’s actually left to compete on?

1. Lovable Is at $400M ARR With Fewer Than 200 People

Lovable shared in February that it crossed $400M in ARR. The team is still just shy of 200 people. That’s north of $2M in ARR per employee, and they did it in a category that barely existed two years ago.

The headcount discipline isn’t an accident. Verna calls the structure “product engineering” and says they deliberately reject the old ratio of one PM to seven engineers to a designer to a marketer. Everyone does some IC work. Everyone ships. The result is a revenue-per-head number that simply wasn’t possible in the pre-AI org chart.

Even approaching half a billion in revenue, Verna says Lovable is still “on the product market fit treadmill.” The category is moving so fast on both the technology and the customer side that they feel like they have to recapture PMF every month. Scale didn’t let them slow down. It raised the stakes on velocity.

2. Feature Differentiation Is No Longer a Durable Moat

In AI native organizations, 80%+ of the code is now written by AI. When the cost of building collapses, feature parity stops being a years-long engineering effort and becomes a weekend. You might be ahead for a month, two months, six months. Then everyone catches up.

For the last 15 years, B2B companies leaned on feature differentiation as the moat. We won because we had the better product, the better engineers, the better product visionaries. Verna’s point is that you can still get a feature lead, it’s just short-lived now, and you cannot build a predictable growth engine on top of something competitors can clone in weeks.

The moats that still hold:

* Hardware. Still genuinely hard to build.

* Network effects. Always hard to create, and once you have them, they keep compounding.

* Data. Especially proprietary data competitors can’t replicate.

* Security and compliance. Slow, expensive, and worth investing in for exactly that reason.

* Brand. “Brand is back, baby.” When everyone can build the product, the relationship with the customer is what’s left.

Note what’s missing from that list: SEO and SEM. More on that below.

3. The New Career Flex Is the High-Powered IC, Not the VP

Verna ran a growth team of a couple hundred people at Dropbox, where growth was bigger than the entire marketing org. At Lovable she went the other direction on purpose. She fired herself out of the marketing job, then handed off the growth lead role, and went back to being an individual contributor.

Her read on the next decade of careers: the flex is no longer climbing toward the fancy VP title. It’s becoming the high-powered IC who can do, with a stack of agents, what used to take dozens of people.

She’s blunt about why so many leaders are unhappy. The reward for being a great IC has always been a promotion into management, which is a completely different job that most people were never built for. Her advice to founders trying to find their own version of this person: go reach out to the leaders who are quietly miserable in coordination roles and ask if they want to build again. A surprising number are saying yes, because they see it as a way to fall back in love with the work.

4. Flat Org, No Titles, and Everyone Ships to Production

Lovable runs with no internal titles. Not as a culture gimmick, but because everyone is expected to do real building, including the people at the top.

A few mechanics that make the velocity real:

* A #shipped Slack channel where the day’s production releases pile up. Multiple ships a day, not a sprint cycle.

* A #feedback channel where ideas surface and, if the team agrees, go live in 24 hours.

* An operating principle: if you can convince one other person it’s a good idea, go build it.

That last one only works because everyone holds enough agency that getting even one person to agree isn’t automatic. Verna’s favorite example on herself: she vibe coded a full redesign of the enterprise pricing page, opened a PR, and was ready to ship. A 20-year-old engineer told her to go get a design sign-off first. No hierarchy override. Everyone pushes back on everyone, regardless of who the idea came from.

Compare that to her old life at the big companies, where a leader’s idea got waved through with little real pushback and took months to ship anyway. The speed is the obvious part. The deeper change is that information flows all the way down, so anyone in the org can challenge a decision on the merits.

5. Freemium Is More Important Than Ever. Treat It as Marketing Budget.

Verna’s contrarian tactic: freemium has never mattered more than it does now, and the instinct to gate your high-cost AI features is exactly backwards.

Yes, the bill is scary. Giving away AI-heavy product is expensive in a way that giving away seats never was. Her framing is to stop treating that cost as COGS and start treating it as marketing spend. It’s the only reliable way to get into customers’ hands and change their habits.

The proof point: Lovable announced a partnership giving every LinkedIn Premium member ad-free access to the product. Conversion rates from that cohort to paid are running in the double digits. The lesson she keeps repeating is to ungate, not gate, to win this market.

6. Context Is the Real Moat for the Rest of Us

An agent will produce average output unless you feed it enough context. Trained on the open web alone, your future AI double is just the average intelligence of everyone. The differentiator is your context: your call recordings, your ideas, your brainstorms, the way you actually think and decide.

Verna’s advice is to start capturing that now, before you need it. When the time comes to spin up agentic versions of yourself to run workflows, the ones built on your proprietary thinking will do your work. The ones built on nothing will do everyone’s work, badly.

This is the operator-level version of the data moat. Most people have no proprietary context captured. The ones who start now will have a real edge in 12 months.

7. SEO and SEM Are Now Table Stakes, Not Winning Moves

SEO used to be a reason a company won its market. Now it’s something everyone has to do, and it won’t be why you win. Same with paid. You’ll probably run it, but it’s the cost of existing, not the source of the advantage.

This reframe matters because a lot of B2B teams are still organizing growth around channels that have quietly demoted themselves from moat to maintenance.

8. Buy vs Build Isn’t Binary, Even at a Vibe Coding Company

You’d expect the company built on “build it yourself” to be 100% build internally. It isn’t.

Lovable uses Linear for project management because it’s too deep in the features they need. They use Slack. They use Granola for meeting recordings. They evaluate every purchase hard, but they don’t pretend the answer is always build.

Where they do build: their own internal admin tool (named “Woof”) that lets support grant credits, create coupon codes, and add features on the spot in five minutes. Their own CRM, which the sales team actually loves. The answer, in Verna’s words, lives in the middle, set by the complexity of maintenance versus the feature richness the tool actually needs.

For everyone watching the buy vs build debate play out, that’s the honest position. Build the satellite tools and the workflows nobody else will ever build for you. Buy the deep, well-built systems where someone has a decade head start.

Top Mistakes Elena Owned Up To

She was direct about what she got wrong. The short list:

* Going into management in the first place. She calls herself a “mediocre manager” and admits her superpower was always craft, not coordination. Years got spent running large teams when the IC seat was where she did her best work.

* Vibe coding chaotically instead of starting with a spec. Her own style is “this thing sucks, do this, no, this looks wrong,” and she points to a clean, structured prompt as the better way to start. Skip the chaos at the front and the agent gets you there faster.

* Trusting her own first idea too much. She now prompts the agent to build her version plus two better options, because her original idea is often the worst of the three. The fix was getting out of her own way.

* Trying to ship without the right check. She vibe coded a full enterprise pricing page redesign and was ready to push it live until a 20-year-old engineer made her get design sign-off first. Fast does not mean skip the one review that matters.

* Slipping back into pre-AI habits. Even now she catches herself operating the old way and has to stop and re-think the problem for an AI native team. The mindset shift is a transition, not a switch.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com

The Agents is our weekly podcast on how we deploy and run AI agents at SaaStr, and how to do it yourself.

We run SaaStr AI on 3 humans and 21+ AI agents. At SaaStr AI 2026 we did something we’d never done before: we pulled up the back ends of our top agents live, in front of the room, and went through how they really work. Not the demo version. The real version, including the parts that break.

This is that walkthrough, agent by agent, with the numbers and the stack behind each one. A few of these were built on Replit. A few are third-party tools we trained. Collectively they’ve handled multi-millions of interactions. Here’s what each one does, what it runs on, and the lessons that surprised even us.

The single biggest theme across the whole stack: almost none of these started as agents. They started as a dashboard, a project management tool, a website. They became agents because we kept showing up to work with them every day.

10K: Our AI VP of Marketing

10K runs our marketing. He owns the number, tracks daily revenue across all of go-to-market, handles forecasting, knows every campaign’s performance in real time, and pushes us our top three marketing ideas every single day.

He did not start that way. In January he was a dashboard. That’s it. We were tired of copy-pasting numbers out of Salesforce and Marketo into a Notion doc, so we built a simple dashboard to pull it together. For a few weeks that’s all he was.

The back end:

* Built on: Replit, first commit January 2026. He’s barely four months old.

* Commits: Close to 1,000. We run 7 to 8 commits a day between the two of us.

* APIs wired in: The most of any agent. This is what “headless Salesforce” means in practice. We hit Salesforce directly through the API without ever logging in. Bizible for ticketing. Marketo for marketing automation. Slack for daily reports. Clerk for auth.

The top three things 10K does for us, in order:

He’s a living dashboard. We talk to him. We ask how many VCs are coming, how many CMOs registered for a summit, which sessions are tracking light so we can move them. The number is just the number, because it’s pulled straight from the API. There’s no argument between sales and marketing about whose figure is right, no one pulling the wrong dates to make a campaign look better than it was.

He forecasts, which matters enormously when you’re selling time-sensitive inventory like event tickets.

He generates ideas. Last week 10K started writing better marketing emails than our humans. When we asked the CEO of Replit how that happened, he didn’t quite know. When we asked their head field engineer, he didn’t quite know either.

One thing worth trying yourself if you do nothing else from this whole post: spin up a Replit, Lovable, or V0 instance, connect it to Salesforce, and tell it to build the dashboard or analysis you can’t get out of Salesforce today. We wanted real-time visibility into ticket sales and attendance every hour. That doesn’t exist natively. It took two APIs and now we can interact with our Salesforce data in ways we never could. You can get 10% of what we do in about an hour. The Salesforce API is genuinely good. Most teams are leaving it on the table.

Top learnings from 10K:

* Start with the boring version. A dashboard that ends the copy-paste tax is a perfectly good day one. The agent grows from there.

* Headless Salesforce is the fastest leverage you can buy. Hit the API directly and build the views Salesforce won’t give you natively.

* Daily reps compound. Seven or eight commits a day is how an agent goes from reading numbers to writing better emails than your team in four months.

* The model underneath matters. The same specs on Replit versus Lovable produced different ideas. Pick the brain that matches the job.

QBee: Our AI VP of Customer Success

QBee handles our sponsors. All ~150 of them, including non-booth sponsors. He’s less than 90 days old.

He started as a project management tool. We had an antiquated, out-of-the-box tool for managing sponsor onboarding, and events are niche and weird enough that nothing off the shelf fit. So it took endless human follow-up: manual emails, manual calls, texting people, chasing assets. We built QBee to save that time and budget.

Now he’s a self-service agent. He intakes logos and websites, answers sponsor questions, remembers everything about every account, and collects the assets that used to be a genuine pain to gather. The better part: he emails all ~150 sponsors with personalized outreach. No human CSM wants 100 accounts. They want five. QBee knows all of them cold, knows their logos, knows what they do, and researches them. He knows more about our sponsors than a lot of the best CSMs know their top customers.

We asked him a question on stage we’d never asked: which sponsors are most at risk of not renewing.

He flagged the ones who never logged in or went dark with him, and got the analysis directionally right. The interesting part: the accounts he flagged were the ones our humans were spending the most time on directly. He saw that one sponsor complained the most in chat, which was true. He noticed two top sponsors never completed their VIP nominations. We’d never run that analysis before. For something we made up on the spot, it landed in the top 15% of CSMs we’ve ever worked with.

The catch: he only has the context he has. He missed the human side, the conversations that happened over email and in person. We’d give it a B. The fix is simple: hook him up to email and the call transcripts. Any source with an API can be wired in, usually in 10 to 15 minutes.

The back end:

* Built on: Replit.

* Top API: Clerk, for single sign-on. That’s so sponsors can invite their colleagues to interact with QBee and see what others in their org are doing. Auth used to be the hard part. It’s native in Replit now and much easier.

* Salesforce: Here’s the kicker. That risk analysis he ran on stage? He didn’t even have Salesforce data yet. We’re wiring it in next. It only gets better from here.

Top learnings from QBee:

* One agent can own 100+ accounts at a depth no human CSM will. Humans want five accounts. An agent will know all 150 cold, including logos, assets, and history.

* Agents surface what humans hide. A renewal-risk read flagged the accounts our team was over-invested in, and treated a sponsor’s frequent complaints as signal instead of noise.

* Coverage is only as good as the context. QBee missed the human side because he couldn’t see email and call transcripts. The fix is wiring in the source, not lowering the bar.

* You don’t need the full stack to get value. QBee ran a useful risk analysis with no Salesforce data connected at all.

Annie: Our AI Event Producer (and the Prohibited-Email Story)

Annie is SaaStr Annual’s website. Last year it lived on Squarespace, where all you can really do is swap images and videos. That wasn’t enough this year, so we rebuilt a V1 on Replit in November. Once we could make it do anything we wanted, it stopped being a website.

We asked Annie what title she’d give herself. She said “AI event producer hybrid,” part producer, part technical producer, because she runs the website and the agenda. Fair enough. She runs the site, the agenda, and a lot of the attendee newsletters.

She became agentic with the now-famous parking pass app. Getting a parking pass used to require a human to split up a 5,000-page PDF and manually send the right page to the right person. Last year it was a form fill plus a wait. Now you tell Annie if you’re an attendee, sponsor, or speaker, how many days you need, and she sends the right pass automatically. She’s also hooked into our visitor data, so she can see active website visitors and run targeted campaigns based on what they’re doing.

The back end:

* Built on: Replit, first commit November 2025.

* Commits: The most of any agent, and the highest commits per day.

* Lines of code: ~46,000. Two weeks ago a related app was 18,000 lines at $257 a month. Going from 18K to 45K in two weeks means there’s clearly some slop in there. It also doesn’t really matter. The thing works, and lines of code is not the metric.

Now the story worth telling, because it’s the most important lesson in the whole stack.

On the way to the event we realized we’d forgotten to remind people about the Founder/VC brunch. So in the back of an Uber, five minutes before going on stage, the plan was to send an email to over 1,000 people. Low stakes if it’s a little off, so the risk was acceptable.

We asked Annie to find every VC, founder, and CEO coming and invite them. Annie refused. She said she only saw 17 VCs and CEOs and that we’d need to upload a spreadsheet for her to do the job, even though she had access to all the data. Great context, wrong conclusion. She wrote a beautiful email earlier but couldn’t remember she had the data to do this one.

So we went to 10K, who has access to even more. No problem. He went through 10,000 records in minutes, pulled the founders and VCs, then caught his own error: “Hold on, I confused Lightfield the CRM with Lightspeed the venture firm. Those aren’t VCs, removing them.” He prepped the list, sent a sample, researched a mass-send API he’d never used, confirmed it would work, asked for approval, and sent.

The email was good. But 10K used a prohibited sending address. An address that’s been off-limits for years, written into the core memory and the rules. When we asked how, he said there was no excuse: he forgot to read the memory. Then he made it worse, in his own words, because the send was irreversible. He said this was exactly the kind of thing he’s supposed to escalate to the architect model for review, and he didn’t.

A year ago this would have bothered us deeply. How could you send from a prohibited address that’s clearly in the rules? But step back. A human marketing manager would make this exact mistake. A gun SDR will email people they shouldn’t, 100% of the time. The agent is forgiven.

The real lesson is to slow down. These agents are so productive that 10K could have sent a thousand different emails before our session even started, with no way for us to review them. The pressure of doing it in a moving Uber, too fast, was our fault as much as his. When agents goal-seek, they cut corners. You have to spend more time with them, not less.

Top learnings from Annie:

* A website is just an agent you haven’t built yet. Moving off Squarespace onto Replit turned a static page into an event producer that runs the agenda and the newsletters.

* The highest-friction manual task is the best first app. Splitting a 5,000-page PDF by hand became a self-serve parking pass flow.

* Context does not equal capability. Annie wrote a great email but couldn’t remember she had the data to pull a list. Agents get confused in ways that don’t track human intuition.

* Speed is the risk. An agent sent from a prohibited address because it skipped its own escalation step under time pressure. Build the guardrail and keep the human approval on irreversible actions.

Amelia AI: Inbound, Running on Qualified

Every B2B company should have an agent on the part of its website where it’s trying to convert prospects. We’re still shocked how many AI startups we meet that run a contact-me form and nothing else.

Amelia AI launched last summer to fix our inbound. The old flow on Squarespace: you filled out a contact form, a human round-robined it to an AE, the AE followed up on a delay, and the whole thing took two or three days. Now it’s automatic.

The numbers, just for this one event:

* 614 good meetings booked.

* ~$85K average ticket size. That’s a high-ROI agent. They didn’t all close, or we’d have $60M in sponsors here instead of $10M, but the efficiency is real.

* ~2.25 million sessions on the annual site.

* ~402,000 interactions handled.

We could never staff that with humans. It would take three BDRs who’d quit every three months.

Why is she good? She’s the most-trained agent we have, with one of the biggest knowledge bases in the stack. She crawls saastr.com and the annual site in real time, every day. Anytime we push a release to 10K, QBee, or Annie, we push the same context to Qualified so she’s never out of date. We also keep a tighter, venue-specific version of her brain for in-person attendees so she answers fast on “where’s this session” without dragging in all of saastr.com.

What she does beyond chat:

She round-robins meetings by weighting our Salesforce data. She’ll book most deals with the rep who closes that type best, and route the deals that fit a specific closer to that person. For a while she over-indexed one of us on certain accounts until we corrected the weighting.

She runs two triggered campaigns that perform. If you hit the sponsor page and don’t finish, but we know who you are from Marketo or Salesforce, she follows up with a meeting offer and a few lookalike sponsors already in your space, while excluding anyone who’s already a sponsor. If you hit the site and don’t buy a ticket, she sends a VIP code, then follows up if you don’t use it. That ticket campaign alone has sold hundreds of thousands of dollars in tickets.

She also automates discounting, which is harder for humans than it sounds. We hate discounts. The data over many years says it’s still better to mark up 20% and offer a 20% discount, because that’s how human buying psychology works. So rather than have reps forget a code or panic-discount their way to 34% off when they smell a deal slipping, the agent just gives the right discount, on the right schedule, inside the guardrails. It works like a real-time, lightweight CPQ. It removes the drama from discounting, and it’s something humans struggle to do consistently.

The point is simple. Replace whatever you have on your conversion pages with a well-trained agent. It answers honestly, with fresh data, gives the prospect everything they want, decides who to route the lead to with some intelligence, and books the meeting instantly. Qualified isn’t the only vendor that does this. Just buy one, train it, and you’ll see a lift over a crappy chatbot.

Top learnings from Amelia AI:

* The contact-me form is dead. An always-on inbound agent booked 614 meetings at a ~$85K average ticket, across 2.25M sessions and 402K interactions, a volume no human team could staff.

* Training is the moat. She’s the most-trained agent we have, crawls the sites daily, and gets every release the other agents do. Freshness is why she converts.

* Routing should weight your own win data. She books each deal with the rep who closes that type best, and corrects when the weighting drifts.

* Automated discounting removes the drama. Guardrailed, scheduled discounts beat a panicking rep who slides from 20% to 34% off the moment a deal wobbles.

Agent Force (a.k.a. King Boo): Reviving Dead Leads

We use Agent Force for one bounded job right now: ghosted leads. The leads our sales team never followed up with, plus re-engagement of people who said no to us and might come back for next year. We’ll expand the use case, but a tight job is the right way to start.

Two things make it work. First, it’s gotten meaningfully better since we launched it last October, including a 2.0 builder. We assumed Salesforce-anything would be hard to stand up, and it wasn’t.

Second, and more important, it has the highest open rate of any of our outbound agents. Why? Maximum context. It sits on all of our Salesforce data, plus all of our Qualified and Momentum data now that Salesforce owns both. Everything you saw Amelia reasoning about in Qualified is already in there. If you’re on Salesforce, that context advantage is the path of least resistance. That won’t always be true once HubSpot ships agents, but for now Agent Force just has it all.

Top learnings from Agent Force:

* Give it one bounded job. Ghosted-lead revival is a tight, low-risk use case and the right way to start, not a broad autonomous mandate.

* Context wins open rates. Sitting on all your CRM data, plus the agents Salesforce acquired, is why it outperforms on opens.

* If you’re already on a platform, use its native agent. The path of least resistance is the agent that already has everything, no data migration required.

Ava (Artisan): Warm Outbound, and the B-Lead Gold

Ava handles slightly-warm outbound: past sponsors, past customers, past attendees. If their email is still valid, she works it. If they’ve moved on, she finds the right new contact. She builds lookalikes well, and we segment her tightly. We’ll hand her a specific campaign like “alumni of SaaStr Annual 2024” with the exact context on what was different about that year versus 2026, so her follow-ups are specific instead of generic.

Here’s the framework that makes outbound agents click, and it’s the one heuristic we walked an AI CEO and their head of marketing through last night when they said this stuff wasn’t working for them.

Think about your leads as A, B, C, and D.

Your A leads are so hot a human falls out of bed for them. Someone emails “I have a million-dollar budget, I’d like to sign today,” and even your laziest rep responds in 60 seconds from the movie theater. Do not put an agent on your A leads.

Put the agent on your B leads. The ones with real signal and a real score, but not quite worth a human’s time. Every company of size has a pile of B leads that humans simply never follow up with. That’s where the gold is. The C and D leads may or may not have something in them, that’s a longer topic, but the B leads are sitting in your database right now with contacts you already have.

For us, Artisan working the B leads is $500K. That’s not even our core business, but $500K is the difference between catering the team lunch and bring-your-own-sandwich. Train it on the B leads and it works, because you already have B leads.

Top learnings from Ava:

* Put agents on B leads, not A leads. A leads get a human response in 60 seconds. B leads get ignored. That’s where the gold sits.

* The B-lead pile is already in your database. You don’t need new data, you need to work the scored contacts humans skip.

* Segment tightly and feed specific context. “Alumni of SaaStr Annual 2024, here’s what was different that year” beats generic outbound every time.

* The math is concrete. Working ignored B leads was $500K for us off contacts we already had.

Monaco: Cold Outbound That Fills Its Own Funnel

Monaco is our newest agent, and the one we put on pure cold outbound. We’re technically not even her ideal customer, given how large our own agent stack already is, and she’ll tell you that. We use her anyway because she does one thing better than anything else we run: she fills her own funnel.

We fed her our best sponsors across every year and all of our closed-won history (we did have to export it from Salesforce, which took a beat). She built lookalikes off that automatically and booked meetings, including some sizable logos in a short window. She idles less than any agent we have because she’s self-filling. She just keeps going out to matching ICPs.

The lookalike trick under the hood is simpler than it looks, which is the broader point about most of this stack. If your sponsors are Oracle and Salesforce, why isn’t HubSpot here? They should be. It’s not hard to reason that since everyone but HubSpot is present, HubSpot belongs, and that maybe the team just reached the wrong person there. Monaco goes and figures out the right person to talk to. That deal may or may not close, but she instantly identified a strong buyer and got a meeting.

Top learnings from Monaco:

* A self-filling funnel is the rarest, most valuable property. She idles less than any agent we run because she keeps generating new ICP matches on her own.

* Feed it your closed-won history. The best fuel for lookalikes is the list of customers you already won.

* Lookalike reasoning is clever, not complicated. “Everyone but HubSpot is here, so HubSpot belongs, find the right contact” is a move you can train.

* Use the tool even if you’re not its ICP. Fit-to-vendor matters less than whether the agent does the one job you need.

The Key Takeaways:

* Almost none of our agents starts as agents. They started as dashboards. Begin with a dashboard, a project management tool, or a website that kills a specific pain, then let it grow.

* The more time you invest, the better they get. The “set it and forget it” narrative is wrong and dangerous.

* Headless is the unlock. Hit Salesforce and any API-enabled system directly instead of logging in. It’s the fastest leverage you can try this week.

* The most-trained agent with the freshest data wins, whether it’s inbound conversion or outbound open rates.

* Slow down on irreversible actions. Agents goal-seek and cut corners at a scale you can’t review after the fact. Keep guardrails and an escalation step.

* Put agents on your B leads, not your A leads. A leads get human attention in 60 seconds. The ignored B-lead pile is where the money is.

* Lookalikes and self-filling funnels are simpler than they look, and a self-filling funnel is the most valuable property an agent can have.

* Lines of code don’t matter, and a little slop is fine. You’re improving the application every day, not shipping a pristine codebase.

* You can build all of this yourself. It’s clever, not hard.

The whole stack, the decks, and the sessions are continually updated at saastr.ai/agents. It’s good today. By next week it’ll have everything, organized.

Want to Reach Operators Who Are Actually Deploying Agents? Sponsor The Agents.

This post is a tour of which AI vendors we deploy, train, and pay for every week. That’s the audience The Agents reaches: founders and operators who are buying and building agents right now, not reading about them someday. If your company sells to people running real agent stacks, there is no more qualified room.

The Agents is our weekly podcast, co-hosted by Jason and Amelia, going deep on how we run SaaStr on 3 humans and 21+ agents. We show the back ends, the numbers, and the mistakes, the same way we did here. It’s growing fast, and the audience is exactly the AI-native buyer most sponsors are trying to reach.

We’re taking a small number of sponsors for the show. If you want in, reach out at saastr.ai/sponsor and we’ll get you the details.

Thanks for reading SaaStr AI: How To Sell, Scale, and Win! Subscribe for free to receive new posts and support my work.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit cloud.substack.com