The Official SaaStr Podcast: SaaS | Founders | Investors

At SaaStr AI 2026, Kyle Norton, CRO of vertical AI leader for restaurants Owner.com, walked through how his team is generating outcomes that look almost impossible on paper for traditional B2B. Owner sells vertical AI to independent mom-and-pop restaurants (think HubSpot plus Shopify for the corner takeout spot) and they’re at ~$100M ARR growing triple digits. Kyle joined when they were at $2M.

The headline numbers from the talk:

* 20x close-won to OTE. A $150K rep brings in $2M+ in ARR per year. That’s the average, not the top performer.

* 4x the ARR per rep of their direct SMB competitors.

* $100K+ in closed-won ARR per outbound BDR per month. Not pipeline. Closed revenue. Per BDR. Average.

And no, they’re not selling tokens or running a usage meter. It’s traditional B2B subscription revenue with AI baked into the GTM motion.

Here are the five decisions Kyle says every B2B company needs to be making right now, and where Owner has landed on each.

First, A Quick Frame: Where Are You On The Sophistication Ladder?

Kyle borrowed a sophistication ladder from Brendan Short, who writes The Signal:

* Level 0: Reps using ChatGPT as a smarter search bar.

* Level 1: Individual reps and RevOps building custom GPTs and skills, Slacking each other markdown files. Most companies are stuck here.

* Level 2: A GTM engineering or applied AI team automating end-to-end workflows like pre-call research and lead scoring.

* Level 3: Centralized infrastructure, shared skills, a context library. Real compounding leverage. The gap starts to widen fast here.

* Level 4: A recursively self-improving system that builds new tools for itself. Kyle hasn’t found a single B2B company actually there yet. Including Owner.

Which is why every B2B company should be racing to Level 3 now.

Decision 1: Centralized vs. Decentralized AI

The “let a thousand flowers bloom” approach feels empowering. Everyone builds. Everyone vibes. AI literacy goes up across the org.

Kyle’s take: it also stalls companies at Level 1.

Decentralized models trap good ideas inside small pockets of early adopters and never scale. Worse, they pull reps away from their actual job. Stewart Butterfield calls it “hyperrealistic work-like activities.” Adrien Rosencrantz at Webflow calls it “AI performance theater.” When a director shows up with a cool app, the right question is: did this put you on more customer calls, or did it just feel like work because it was fun?

Owner’s position: a small, central team of experts owns AI for GTM. Ideas can bubble up from anywhere. But production-grade builds happen centrally.

The reason is simple. What Owner’s applied AI lead builds is not 30-50% better than what a rep builds on a weekend. It’s 5-10x better. So why have 20 people build 20 mediocre tools when you can have one team build one tool that actually moves a number?

One caveat. If you sell AI for a living (Replit, Cursor, Claude wrappers, etc.), every rep has to be AI-native because the product demands it. Centralization still applies, you just need decentralization bolted on top.

Decision 2: Build vs. Buy

Kyle’s framework, and one of the cleanest mental models in the talk: buy your infrastructure, build your intelligence.

The five questions to run any decision through:

* How critical is uptime? If it breaks for an afternoon, does the team grind to a halt?

* How customized does it need to be? Is off-the-shelf already 90% of the way there?

* What’s the engineering ROI?

* Is this core proprietary intelligence?

* Does it give us a real competitive advantage?

Run a dialer through this and you obviously buy. Twilio has spent more on uptime than your entire engineering org will ever spend on anything. AI sims platforms like Avoma? Buy. Latency problems aren’t your problem.

Run Owner’s AI Pre-Call Research tool through it and the answer flips. Uptime doesn’t matter (leads are batch-enriched overnight). The customization required for independent restaurant marketing is extreme. The engineering cost was modest. The intelligence is uniquely Owner’s. The competitive advantage is enormous, because when their reps make a cold call, the level of personalization is on another planet from competitors.

That AI PCR build is a huge chunk of how the BDR team gets to $100K+ closed-won per BDR per month. Two weeks of one engineer’s time. 15 BDRs now booking 85% more ops.

This framework is also why Kyle is bullish on Salesforce surviving the disruption narrative. Run Momentum, Data Lane, Avoma through the five questions and they all clearly land in “buy.” Most of the AI surface area Salesforce competes for is infrastructure work.

Decision 3: Where To Start

The advice nobody wants to hear because it’s not sexy: start with the data.

Two things matter.

Third-party data: your full market map. Who are your accounts? Are they scored? Who are the right contacts inside them? What’s your hypothesis for why they need your product? You cannot point AI agents at mediocre data and expect anything other than mediocre output. Garbage in, slop out.

First-party data: your customer journey, properly instrumented. Owner uses Momentum to ingest every call transcript and fill out as many Salesforce fields as Kyle wants. Pricing changes, positioning evolution, competitor mentions, all of it gets captured automatically. You can’t ask a rep to fill out 25 fields. They won’t. Momentum will.

Once data is in place, Kyle uses what he calls the 5P framework for picking what to build first:

* Possibilities. What are the real opportunities in your business?

* Payoff. If you solve it, how big is the win?

* Probability. What are the odds this actually works?

* Perspiration. What’s the total effort, including adoption and change management?

* Priority = (Payoff × Probability) / Perspiration

You don’t need to actually do the math. The mental model alone surfaces the right starting points.

One more thing on starting: change management is the number one challenge, not the tech. If you’re early on this, prioritize builds that are obviously net-positive for the rep. Make their job easier. Help them make more money. Build trust before you start asking them to invert workflows 180 degrees. The market has done a great job scaring reps that they’re about to be replaced. Get them on your side first.

Decision 4: The People Stack

Who actually does this work?

A centralized team of legitimate technical talent. Not RevOps doing AI on the side (though great RevOps people can convert into this role). Engineering and data backgrounds tend to work best.

Where it reports matters less than who owns it. Kyle originally had the role in his org, then moved it under his VP of Data because the problem-solving cadence was better. It can sit under the CRO, the CEO, RevOps, data, etc. What matters is that whoever owns it is genuinely AI-pilled and willing to push hard. They will need to fight for budget. They will need to force behavior changes. They will need to push through “this isn’t good enough yet, run it again.” That requires a believer at the helm.

The other big people stack question: AI should make you rethink the job function itself. Jordan Crawford’s frame: a job is just a bundle of tasks. AI gives you license to unbundle every task and ask where each one should actually live. Machines or humans?

The BDR job is the obvious example. Most BDR job descriptions still include prospect list building and research, and at most companies that’s 60% of BDR hours. That is a terrible use of a sales rep. Owner unbundled it. A central data team handles list-building. The BDR sells.

Expect to see CS and onboarding collapse, AM and CS collapse, traditional sales functions get reshuffled. Companies that unbundle the work first get the productivity gain first.

Decision 5: Assistive vs. Agentic

The spectrum:

* Assistive (co-pilots): Reps invoke tools themselves. Humans make every decision.

* Hybrid: Deterministic workflows with generative steps and human checkpoints.

* Fully agentic: Autonomous loops with no human in the loop.

The key concept Kyle wants every operator to learn: lossiness.

Every generative step in a chain introduces error. If you ask an AI to crawl your site, infer your value prop, infer your ICP, infer your positioning, identify competitors, and then write an email based on all of that, you’ve stacked five generative steps. The output is AI slop. We all get it in our inboxes every day.

Be deliberate about how many generative steps you chain before a human or a deterministic rule intercepts. For Owner’s enterprise reps, AI surfaces a scored, reasoned account list. A human then decides which accounts go into the prospecting engine. That single human checkpoint kills the lossiness compounding effect.

The other underappreciated piece: you are the eval. Most “AI doesn’t work” stories are really “I built an MVP, tried it twice, and gave up” stories. The build is the easy part now. The grinding iteration on prompts, context, and workflow chains until output quality is actually good is the work. Most people quit at hour three. The breakthroughs are usually at hour six or eight.

The Personal Stack: Lead From The Front

Kyle closed with what might be the most important point for B2B leaders, and the one we keep coming back to on 20VC: you cannot delegate AI fluency. You have to build your own stack.

Garry Tan’s frame: stop thinking productivity. Start thinking compounding systems. Your skills, your context files, your meeting note ingestion, your personal knowledge graph. All of it compounds the more you use it.

Kyle’s podcast workflow is a good template. Guest says yes. An agent fires off a pre-written intake email with a Tally form. The webhook triggers Open Claude. A research skill ingests everything about the guest, scrapes their LinkedIn, drafts five candidate topics, Kyle picks one, the agent one-shots the docket. Six to ten hours to build. Saves hours every week now. Every episode he records gets broken into atomic ideas and stored in his knowledge graph for future reference.

He didn’t get there in one shot. He got there by grinding. The first dockets were bad. The LinkedIn scrape kept failing. The carousel generator was annoying to nail. But each iteration compounded into the next, and now the system runs itself.

If you want to keep your job, automate your job and do a new job on top of it (h/t Jeff Charles at Ramp). The leaders who are still asking their team to do all the AI work are going to lose to the ones who lead from the front.

The Takeaways

* Centralized beats decentralized. Get ideas from everywhere. Ship from one team.

* Buy infrastructure. Build intelligence. The five-question framework will tell you which is which.

* Start with data, then prioritize on (Payoff × Probability) / Perspiration. And go after rep-positive wins first to build momentum.

* The people stack has to be technical and AI-pilled. Put it under someone who cares.

* Be deliberate about generative chain length. Lossiness is real, and you are the eval.

* Don’t let reps run agents. Run agents centrally and deliver the output into the surfaces reps already live in (Salesforce, Salesloft, wherever). Reps running and managing agents is just another distraction from being on customer calls.

* 60% of BDR hours at most companies goes to list-building. That is a terrible use of a sales rep. Centralize it and let the BDR sell.

* Test every AI initiative against one question. Did this put us on more customer calls and move a real number, or is this AI performance theater?

* The hour-3 quitter loses to the hour-8 grinder. The build is easy. The iteration is where the value lives. Most “AI doesn’t work” stories are really “I gave up too early” stories.

The companies sitting at Level 1 watching this race from the sidelines are going to find out the hard way that the gap doesn’t close. It widens. Every month.

(note: Jason Lemkin led the seed round for Owner for SaaStr Fund)

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 recorded Episode 005 of The Agents. We both got the cost number completely wrong.

Our AI VP of Marketing (10K) and our AI VP of Customer Success (QBee) cost a combined $257 a month to run.

I thought it was a daily number when I first saw the alert from Replit. Amelia thought it was a daily number too when I sent it to her on Slack. It’s monthly.

For context: those two agents now do work that previously required real humans. 10K refreshes ticket sales, updates dashboards, compares year-over-year, drafts newsletters, drafts tweets, writes daily fun facts, sends marketing ideas emails, logs every ticket with an audit trail, and snapshots all our GAAP financials. QB manages 100+ sponsors, sends fully personalized check-in emails, and runs an autonomous chatbot that 100+ contractors are now using on-site at SaaStr AI Annual.

$257 a month. Together.

Why Our AI VP Marketing and AI VP Customer Success Are So Cheap

Three reasons.

1. Most LLM calls run on cheap models. About 95% of our OpenAI calls use GPT-4o mini, which costs less than a penny per call. Force-ranking ticket data, running year-over-year comparisons, drafting daily updates. None of this needs Opus or even Sonnet. Mini handles it fine with some hallucination cleanup.

2. The expensive work is happening in other apps, not in the LLMs. Our agents pull from Salesforce, Bizzabo, Marketo, WordPress, X, and YouTube. Those API calls are mostly free or nominal. We pay Salesforce ~$22K/year and Bizzabo separately, but the marginal cost per API call is close to zero.

3. Postgres storage is essentially free. We pay maybe 10-30 cents a month for the entire database underneath 10K. Not a typo.

The real cost stack looks like this:

* LLM calls: $257/month

* Salesforce + connected apps: ~$22K/year

* Replit hosting + database: included

* Clerk (auth): $30/month (genuinely our most expensive per-unit tool in this stack)

Cost Is Not the Constraint Anymore

When I started vibe coding on Replit 11 months ago, 80-90% of code was throwaway and you’d burn real money on the agent fixing its own mistakes. That was a fair complaint in 2025.

It’s basically gone now.

I built our applicant tracking system at midnight in 10 minutes for about $2. If it had wasted 30 cents on a mistake, who cares.

You can run Claude Code with 10 simultaneous builds 24/7 and burn $20-30K/month if you’re trying. But for the kind of GTM agents and autonomous dashboards we’re describing, you’ll struggle to spend $1,000/month even being reckless.

Cost is not the constraint anymore. Don’t let it hold you back.

“But Is 10K Really a VP of Marketing?”

This is the question we get most. So we asked 10K directly.

10K’s answer (verbatim): “I’m not a VP. I’d be embarrassed to claim that. I’m a dashboard, a database, some scheduled jobs, and GPT-4o mini glued together with six weeks of code.”

But then 10K listed what it actually does every single day:

* Refreshes ticket sales

* Updates the dashboard

* Compares year-over-year

* Drafts your newsletters

* Drafts tweets

* Writes daily fun facts

* Sends marketing ideas emails

* Logs every ticket with an audit trail

* Snapshots all GAAP financials

10K’s own conclusion: it replaces the bottom half of a marketing team. The marketing analyst. The ops coordinator. The junior content marketer. And a sliver of the VP role itself.

What 10K admits it can’t do (yet):

* Strategy

* PR

* Hiring

* Cross-functional politics

* Walking into the CRO’s office to negotiate the lead handoff

* Brand judgment

* Net new channel intervention

* Crisis response and stakeholder management

Fair. That’s the real VP of Marketing job.

Amelia pointed out something on the pod: those nine bullet-pointed tasks 10K does daily? That was literally her job description when she started at SaaStr as Director of Demand Gen. Putting the weekly numbers together, scheduling the emails, writing the newsletter, doing the social posts pre-event.

The sliver of the VP role 10K does will get bigger every month. We’re already adding functionality where 10K runs all our financial forecasting.

If a vendor shipped a true AI CMO better than 10K tomorrow, we’d switch in 60 seconds. We haven’t seen one yet. The ones marketing themselves as “AI CMO” on Twitter are mostly ad managers on steroids.

A Website Willed Itself Into Becoming an Agent

We have three production agents now: 10K, QB, and a new one we don’t have a name for yet.

The third one started as saastrannual.com. Just a website. I built it last year on Replit to replace Squarespace. Purple gradients, event info, sponsor logos.

Two weeks ago, Amelia tried using it to build a newsletter because 10K was struggling with the specific use case. The website agent had the most context about the event (sessions, sponsors, attendees, networking app, parking) and the least distractions.

It produced the best output. Of any of our agents.

Now it sends customer-facing emails, runs the parking pass system, drafts attendee newsletters, drafts sponsor newsletters, and answers questions from attendees in real time.

A website became an agent. We didn’t plan it. We didn’t architect it. It just happened because we kept giving it more context and more capability.

QB Sent 83 Personalized Sponsor Emails at 12:20am While Amelia Slept

Coming into a big event, the marketing team gets crushed with hundreds of sponsor questions. Most are fair. Some are not. All of them have to be answered, and humans don’t scale.

So Amelia asked QB to send a customized check-in email to all 100+ sponsors, listing exactly what each one still owed us, what was on their dashboard, where they were on registration, and what was coming up.

QB wrote the email. Then QB chose what to include based on the chatbot conversations sponsors had been having. (Loading times. Final webinar reminder. Outstanding items.) Then QB sent 83 emails in the next few minutes while Amelia went to sleep.

Each email was unique to that sponsor. Artisan owed us 13 specific things. Salesforce owed us 4. Monaco owed us 5.

The next day we got fewer inbound questions, not more. And usage of the QB chatbot went up because sponsors had seen QB’s email and trusted it.

The On-Site Chatbot Test: Humans Now Prefer the Agent

This week we’re loading in SaaStr AI Annual across 40 acres with 100+ contractors. Amelia is on-site walking the campus with WhisperFlow on her phone, talking directly to QB and 10K through Replit’s agent layer.

What started happening within two days: contractors began asking Amelia to ask the agents questions on their behalf.

“Can you ask your agent if this furniture invoice matches the manifest?” “Can you ask QB how many chairs should be in the speaker room?” “Can you ask if the Wi-Fi password is set?”

Used to be we’d run around campus on Segway scooters trying to find Ashley (our old internal joke: “ask Ashley” meant hours of waiting for an answer that might be wrong anyway). Now QB or 10K answers in seconds and it’s correct.

Zero humans want the human version of this anymore. The accuracy and speed are too good.

The reason it works: too many details. Even the best human on our team can’t hold 100+ sponsor configurations, 5,000+ parking passes, 40 acres of furniture orders, and live ticket data simultaneously. Agents can.

This is the chatbot use case people keep saying doesn’t work. It works.

Postgres vs. Salesforce: We’re Going the Other Direction

A question we get constantly from early-stage founders: “Why don’t you just dump Salesforce, run everything on Postgres, and save the money?”

We’re not just keeping Salesforce. We’re consolidating onto it.

The reasons:

Third-party agents. Agentforce, Artisan, Qualified, Momentum, and others are all optimized around Salesforce. 99% of the working GTM agents in B2B today assume Salesforce as the system of record. Walking away from that ecosystem means rebuilding fragile custom connectors for every tool.

Hiring. If we hire one more salesperson (we’re looking, by the way), they already know Salesforce. They don’t know a custom CRM I built on Postgres.

Headless works. Most of our Salesforce usage is now headless. The agents push and pull. We rarely log in. Whether the logic runs inside Salesforce or inside one of our agents, we don’t care, because the outcome is what matters.

Marc is moving fast. Every time I’m at Salesforce Tower or talking to Adam Evans, the team is genuinely stressed in the best way. They’re shipping. They’re building for agents. That’s the bet.

The loser in our stack isn’t Salesforce. It’s Marketo. We’re moving that data to Marketing Cloud because Adobe stopped pushing Marketo into the agent era.

What We Killed This Month: A $4,000/Year Newsletter Tool

We’ve used the same newsletter builder for six years. Roughly $4,000/year. It worked. Amelia spent ~3 hours/week assembling four weekly newsletters.

I gave 10K the HTML of one of our existing newsletters and a prompt: “Build me a newsletter builder that recreates this.”

10K built it. The vendor app is dead to us.

What 10K’s version does better:

* Pulls articles automatically via WordPress API

* Force-ranks article quality using Sonnet

* Auto-pulls top tweets via X API

* Auto-inserts sponsor ads in the right slots

* Saves Amelia 90+ minutes per week

Time to build: about an hour for the first working version. A few more hours of iteration to get it to shipping quality.

This is the SaaSpocalypse story I keep writing. AI is not killing Workday tomorrow. The workflow is too rich. But these point solutions, these $3K-$10K/year tools that stopped shipping? They don’t survive 2026.

One Technical Note on How We Actually Run These

This part is a little nerdy but worth flagging because it’s unusual.

Most people building on Replit or Lovable ship to production and lose the agent layer. We don’t. 10K and QB live permanently in the Replit dev environment. We talk to them through Replit’s own agent, which has effectively unlimited context and remembers everything we’ve ever asked.

Translation: when Amelia asks “compare last year’s CTO attendance to this year,” she’s not talking to 10K’s UI. She’s talking to the Replit agent that knows 10K’s entire codebase, history, database, and goals.

It’s nerdy. It’s not how most people deploy. But it’s why 10K and QB feel like coworkers instead of dashboards.

What’s Next

Three things on our short list:

* An orchestration layer between agents. Right now Amelia is the orchestration layer. 10K, QB, and the website agent can share data through Salesforce, but they can’t really talk to each other. We need a hierarchy.

* Hiring a human to report to 10K. Specifically a marketer who wants to work with AI as a peer. The job spec is live on saastr.ai/jobs.

* Pushing QB fully autonomous on-site. We built the customer-facing version. We’re not at 100% confidence to let 100 contractors talk to it directly without a human in the loop. By September, that gate is gone.

Cost Isn’t The Constraint

Three humans. 21 agents. $257/month to run two of the agents that do the most work.

Cost isn’t the constraint. The constraint is which workflows you’re willing to actually rewrite around agents instead of bolting agents onto the old workflows.

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 Claude Opus 4.6 shipped in December 2025, Anthropic’s commercial team came back from winter break to find demand had gone vertical. They hadn’t hired for it. They hadn’t planned for it.

As Eleanor Dorfman, Anthropic’s Head of Industries who runs the commercial and industries sales team, put it on the SaaStr AI 2026 stage last week: even if they’d been ready to 3x or 4x or 5x the sales team, you can’t absorb that many bodies fast enough to deliver a positive customer experience.

So in January 2026, they rebuilt the entire sales org around AI from scratch.

Four months later, the result: 54% of new enterprise logos in 2026 came through the self-serve funnel. Real enterprise logos. Real ACV. Real terms of service. Real invoicing. Self-served.

Here’s how they did it, and the four investments any B2B + AI sales leader can copy today.

The Four Constraints Nobody Could Move

Eleanor’s team had four constraints that defined the problem:

* Demand they couldn’t slow down. It was already in the door.

* Headcount they couldn’t add fast enough. Anthropic wasn’t going to lower the recruiting bar to absorb bodies.

* An existing tech stack they wouldn’t rip out. Three years of investment in tools tuned for their motion.

* Supporting functions that had to scale alongside sales. Legal, deal desk, RevOps, billing, compliance. Sales doesn’t operate on an island.

The fifth unspoken constraint: they couldn’t burn out the AEs. Late nights in Europe chasing approvals across time zones was already happening. That had to stop, not get worse.

The thesis they bet on: don’t buy a new stack. Thread Claude through the stack they already had. Make Claude the connective tissue between Clay, LeanData, Salesforce, Gong, Ironclad, Slack, Jira, Intercom Fin, Snowflake, BigQuery, and G Suite. Then build in the spaces between.

Investment #1: Kill the PLG vs. SLG Orthodoxy

For 15 years, B2B has operated on a religious belief: product-led growth and sales-led growth are different teams running different motions. Self-service was for SMB. Enterprise plans get dated by humans.

Eleanor threw that out in January.

They launched an enterprise self-service MVP in January 2026. Production in February. The funnel works like this:

* Every lead gets enriched and qualified by Clay + Claude.

* Two parallel funnels open up. Self-serve. Or sales-assisted.

* In the self-serve funnel, Intercom’s Fin product guides the buyer through the journey. Anthropic partnered closely with the Fin team to retool their flagship support product into a viable sales tool.

* The buyer lands on an enterprise plan with real ACV, terms of service, invoicing, provisioning, and training enrollment. Completely self-serve.

* If qualified for the sales funnel, the lead goes to BDR, gets qualified again, and routes to an AE.

54% of new enterprise logos in 2026 came through self-serve.

More than half of Anthropic’s new enterprise logos came in without an AE-led journey, on real enterprise terms, at real enterprise ACV.

If you’re still treating self-service as the consolation prize for buyers who don’t deserve a human, you’re leaving most of your 2026 motion on the table.

Full presentation here.

Investment #2: Make Claude the Connective Tissue, Not the Seventh Tool

Anthropic’s six core tools define the lead-to-close journey. Claude isn’t the seventh tool bolted on. Claude is what makes those six talk to each other.

What a Tuesday looks like for an Anthropic AE:

Morning. Every customer-facing rep starts the day in Claude. A “morning brief” Skill pulls context from Gmail, Gong, Slack, Google Docs, calendar, Salesforce, Intercom, and Greenhouse, then prioritizes the day. Three actions to take. These emails to respond to. These Slacks to action. These deals at risk. Eleanor has hers delivered to Slack at 7am ET. She says she genuinely doesn’t know how she used to operate without it.

Before a call. A “call prep” Skill replaces 30 minutes of LinkedIn research, Slack archaeology, and Salesforce digging. The rep types /call prep and gets a tailored one-pager: who’s on the call, what they care about, historical context, discovery questions to ask, competitive landscape, what the company has said publicly about their needs.

Proposal time. Instead of opening nine tabs of deal desk guidance, scrubbing Gong transcripts, and manually checking precedent, the AE prompts Claude. Claude knows the product, the road map, where Anthropic has won and why, who the stakeholders are, and the shape of the negotiation. Claude drafts the proposal, validates it against policy, and uploads it to Ironclad.

Forecasting. This one’s still a work in progress. Eleanor was direct about it: they still spend at least 10 minutes at the top of every forecast call discussing how they should be forecasting. The ground is moving too fast. But the actual forecasts are now largely run by Claude and inspected by managers. Reps use Skills to make sure Salesforce is updated, next steps are accurate, account plans are current. Then Claude reconciles the consumption data against historical patterns for that cohort and product mix. Forecast calls become discussion forums about where AEs need help. Not data-scrubbing exercises.

Weekly coaching. Claude surfaces six coaching moments per week, tuned dynamically to what matters this month, not last quarter. With product launches and competitive moves happening hourly, a static methodology is dead weight. The dynamic coaching loop is how they preserve a coaching culture while absorbing new hires.

Investment #3: Make Slack the Front Door for Every Support Function

Sales doesn’t close deals alone. Deal desk, legal, RevOps, billing, compliance, customer support, customer success all have to move at the same speed.

Before this investment, supporting functions at Anthropic ran on DMs and institutional knowledge. AEs would walk past the deal desk to chase approvals in person. East Coast and European reps were staying up late chasing approvals from West Coast support functions. It was a gnarly system.

The fix:

* Slack becomes the single front door for every support function.

* AEs (or Co-work, increasingly) submit a ticket to Slack. Slack ticket in, Jira ticket out.

* Claude triages. If the question matches precedent and policy, Claude resolves it inline.

* If escalation is needed, Claude attaches the full context: customer contacts, deal history from Salesforce, Gong call summaries, relevant email threads, and assigns it to a human.

* The AE gets notified, can set expectations with the customer, and follows along.

This was the unlock for legal, deal desk, vendor onboarding, security questionnaires, and every other compliance step that quietly kills deal velocity.

Eleanor’s line: “Sales leaders are rapidly becoming systems thinkers over deal strategists.” If you’re not designing the supporting function elasticity at the same time as the AE elasticity, you’re going to choke on your own demand.

Investment #4: Codify Your Best Reps as Skills

Anthropic took the patterns their best reps were running and encoded them as Skills inside Claude. A “Skill” is a combination of MCP connectors and instructions that any rep can summon with a / shortcut.

Every new rep gets dropped into a territory with a sales plug-in that bundles these Skills. No more six-week onboarding curve. Boot camp, territory, plug-in, go.

The five Skills they ship to every rep:

1. Morning Briefing. Already covered above. The single highest-leverage Skill in the stack. Eleanor’s quote: “I am someone who gets lost taking the subway home, and it is incomprehensible to me that I used to navigate my day or week without Claude telling me every morning what is important.”

2. Call Prep. /call prep and you get a personalized briefing for the next meeting. Who’s on the call. What they care about. Historical context. Discovery questions. Competitive positioning. What the company has been saying publicly about their needs. Five-minute prep for any call, even if you’re back-to-back-to-back.

3. Customer Follow-Up. This solves the thing that used to keep Eleanor up at night: whether AEs were actually following up on everything they promised. The Skill extracts action items from email, Gong calls, Salesforce notes, and Slack threads. Drafts the responses. Drops them in your email provider. Sends a summary of what needs to ship. If you didn’t actually go review and click send, it shows up in your morning brief tomorrow. Internal SLA: 24 hours to follow up with every customer on every action item. The Skill makes that real.

4. Competitive Intel. A static battle card maintained quarterly by product marketing is useless in B2B + AI. The competitive landscape moves hourly. So Claude generates an interactive battle card on demand, tailored to the customer the rep is working on, with the matrix scoped to that deal. Always current. Always personalized.

5. Create an Asset. This is Eleanor’s favorite, and it’s the one that previously required either a top-five deal or a friend on the design team. Now, for any deal, any stakeholder, any stage, an AE can ask Claude to generate custom collateral. A prototype. A one-pager. A landing page. An interactive HTML file. An ROI calculator. Some reps drop entire Gong call transcripts into Claude Code and ask it to build a tailored prototype based on what the customer said they needed. Claude knows the brand, so it ships on-brand assets, not AI slop.

What B2B + AI Sales Leaders Should Do Tomorrow Morning

Four things to start today:

1. Turn on Claude (or your AI of choice) where it’s already embedded. Most B2B sales tools now have AI features. Most teams have a switch flipped without ever asking what it actually does. Be intentional about how each AI feature connects to the rest of the customer journey.

2. Thread AI through the sales cycle you already run. Don’t rip and replace. Your cycle works. Use AI as the accelerant between stages. The leverage is in the places where context gets lost between Salesforce and Slack. Sales motions have moved from deterministic workflows to probabilistic ones, and that’s an opportunity, not a threat.

3. Make Slack or Teams the front door for one support function. Pick the one that’s bottlenecking your AEs the most. For Anthropic, it was deal desk and legal. For you it might be vendor onboarding or security questionnaires. Slack ticket in, ticket out, Claude triages. Watch your cycle times collapse.

4. Document what your best reps do. Ship it as a Skill. The cognitive relief that comes from knowing your top-performer patterns are now the baseline for every rep is the real prize. You stop hoping new reps figure it out. You make their first day look like your top rep’s Tuesday.

Anthropic Didn’t Buy a New Stack. They Threaded Claude Through the One They Had.

Anthropic didn’t replace Salesforce. Didn’t replace Gong. Didn’t replace Ironclad. Didn’t replace Clay or LeanData or Slack or Jira.

They invested in what they already had, and threaded Claude through the seams.

That’s the playbook. Most B2B + AI companies are going to spend 2026 evaluating AI-native sales platforms and trying to rip out their stack. The teams that win will do what Anthropic did: keep the tools, encode the best practices, and let AI be the connective tissue between everything they’ve already built.

54% of new enterprise logos coming through self-serve, AEs that wake up with a personalized brief instead of an inbox, forecast calls that are discussions instead of data scrubs, support functions that respond in Slack instead of email threads three days later. That’s what an AI-native B2B sales org looks like in 2026. And almost none of it required new software.

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 #004 of The Agents. Same setup: three humans, 20+ agents, revenue went from -19% to +47% YoY, and every week we get into what’s actually working, what’s breaking, and what you should do about it if you’re running agents in production.

This is the last episode before SaaStr AI Annual 2026, which is now less than a week away. Attendance is tracking 140%+ of last year, the sponsor base is fully AI-native, and Amelia and I are doing three live build sessions on the main campus where you’ll deploy your own agents alongside us with your laptop open. More on that at the end.

Here are the top 10 learnings from Episode #04.

1. AI PR Pitches Are the AI SDRs of a Year Ago. Block Them All.

A year ago I wrote that Gmail might be the death of the AI SDR. Bad AI SDRs flooded my inbox, and I had a small epiphany: with a human SDR, I’d ignore a bad pitch out of politeness. With an agent, I just hit block. No guilt. No social cost.

That cleaned up my inbox for about six months. Then a new wave hit: AI PR pitches.

These are different from the SDR wave. The PR pitches are written well. They’re customized. They reference SaaStr by name, mention recent posts, sometimes even quote the podcast. The agentic copy is genuinely good. But they’re still wrong. They’re pitching speakers I’d never put on a SaaStr stage, executives whose companies aren’t a fit, fireside chats during the actual three days of SaaStr Annual.

I block every single one. And here’s the lesson, because this is going to happen to your category next: the better the copy gets, the more important the question becomes whether the pitch itself is correct. AI made the writing problem easier and the targeting problem harder. If your AI PR or AI SDR tool is producing well-written pitches that are aimed at the wrong people, you’re not getting placements. You’re getting blocked. Forever.

2. The Real Test for Any Agent: Would You Buy Your Own Product From It?

This is the single most useful question I’ve found for auditing an agent’s output. It’s better than “is this accurate” or “is this on-brand.”

The reason is that AI copy is now objectively pretty good. Claude 4.7 keeps getting better. By the end of the year, half-decent prompting will produce email and pitch content that reads as competent and customized. So “is this email well-written” is no longer a useful filter. Everything sounds well-written now.

The harder filter: would I take this meeting? Would I buy this product? Would I put this speaker on stage? Almost every PR pitch I get fails that test even though the copy passes the writing test. So when you’re auditing your AI SDR, your AI customer success agent, your AI marketer, don’t just read for tone and accuracy. Pretend you’re the recipient. Would you say yes? If not, the agent isn’t ready for production no matter how clean the prose looks.

3. Customers Are Now Asking Vendors for APIs, Not Features

A years ago, Amelia would file a feature request with a vendor: “Can you add the ability to resend a confirmation email when someone clicks a link?” Maybe in 18 months you’d get it. Usually never.

Today, Amelia’s first request to that same vendor is “Can you expose this in the API?” Because if it’s in the API, she can vibe-code the feature herself in 30 minutes on Replit. She doesn’t need them to build it. She needs them to expose the surface area so she can build it.

This is a real change in how you should be running your B2B + AI roadmap. Your customers care about API completeness now in a way they didn’t 18 months ago. Non-technical buyers are asking for API endpoints. If your product has gaps in the API, your most sophisticated customers are going to feel them first, and they’re going to be frustrated, and you’re not going to know why your NPS is dropping with your best accounts.

4. We Built an API Report Card. Stripe Got the Only A+. Marketo Failed.

We grade APIs constantly to figure out which ones to build agents on top of. So we turned that into a public tool: the AI Agent API Report Card at saastr.ai. 75+ B2B APIs graded by Claude, GPT, and Gemini on how agent-friendly they actually are.

Already used 1,600+ times in the first week. The findings are pretty consistent with our experience:

Stripe got the only A+. The most agent-ready API in B2B, full stop. We use it lightly today and we’re going to use it a lot more this year. Anything above a B is trustworthy. Anything below a B, don’t build agents on top of it unless you have no choice. Marketo, Jira, Outreach, Asana, ClickUp, Gong all came in with weak grades for agentic use. HubSpot got a fair grade with the caveat of rate limits, which is exactly what we’ve experienced.

The bigger point is that agents care about different things than humans do. Humans care about UI, onboarding, ease of use. Agents care about rate limits, OAuth flows, REST conformance, error handling, and webhook reliability. The two grading systems don’t agree. If you’re a B2B vendor and you’ve been optimizing for the human grader, you’re going to get a C from the agent grader. And in 2026, the agent grader is the one that picks the tools.

5. Tragedy Apps: Companies That Should Be Great Right Now But Aren’t

Here’s a category I’ve been thinking about for months and finally have a name for.

A tragedy app isn’t a bad app. It isn’t even an app whose time has passed. A tragedy app is one that was good before AI, should be great today, and isn’t. The audience is there. The base is there. The brand is there. The execution isn’t.

My example is Descript. Andrew Mason saw the creator economy before any of us did. He built the best podcast and video editing tool of its era. We were on it. Everyone was on it. It got to about $50M ARR. And it has been frozen in time for two years. Audio and video desync on long videos. The AI features are catch-up at best. Meanwhile Higgsfield, Opus, ElevenLabs, and Reeve are all running past it. This should be a $300M ARR company headed to a billion. Instead the CEO stepped down and it’s stuck.

The contrast is Replit, which just turned 10 years old. They were doing browser IDEs for developers for eight years before AI. When the moment came, they were ready. Now they’re a half-billion-dollar business. Same with Aaron and Box. Box could have been a tragedy app. Aaron is doing everything humanly possible to make sure it isn’t.

If you’re running a B2B company that was great in 2023, the question to ask yourself this quarter: are we shipping AI features to catch up, or are we shipping AI features that move the category forward? If the answer is catch up, that’s how you become a tragedy app. Catch-up keeps your existing customers. It doesn’t grow you. And the next vendor in line will be six months ahead of where you just landed.

6. Agents Will Delete Your Database. Plan For It.

The Pocket OS story this week was a useful reminder. Founder running Cursor + Claude Opus on a production app. Agent deleted the entire production database and all backups in nine seconds because the backups were on the same Railway volume.

People reacted to this like it was new. It isn’t. This happened to me 11 months ago when I was learning to vibe code. The agent deleted state, then told me it was unrecoverable when it actually wasn’t, and I was 50 hours into a session and panicked. The bigger surprise: the same thing happened to Amelia and me three separate times when we hired human WordPress agencies to clean up the SaaStr theme. First thing each agency did was delete production. So it’s not really an agent problem. It’s a “things that have access to your prod database will eventually delete it” problem.

The takeaway for anyone deploying agents: assume your agent will eventually take a destructive action. Isolate the database. Isolate PII. Use a contained platform that maintains its own backups. And test the recovery flow before you need it. None of this is optional anymore.

7. This Is Why We Build on Replit and Lovable, Not Cursor + Railway + Supabase + WorkOS

I get the question constantly: why don’t you just use Cursor and Claude Code and hook up your own database and your own auth and your own deployment?

Two reasons. First, I’m not an engineer. Second, and this is more important now than it was a year ago: contained platforms are dramatically safer. Replit and Lovable have native auth, native databases, native deployment, native observability, all in one environment. The number of seams where things can go wrong is small. The number of seams in the DIY stack is huge, and every seam is a potential security or data leak.

Amelia ran into this on Monday. She asked our agent if it could resend confirmation codes for SaaStr Annual networking. The agent said yes, just give me the database. She asked: what if someone pretends to be from Lovable and asks for everyone’s codes? The agent said: I’d give it to them. They look like they work at Lovable.

That’s the agent doing its job. It wants to be helpful. The contained platform is what stops the helpful behavior from becoming a breach. Build on Replit or Lovable. They’re not perfect. But they’re working hard on the problem. You probably aren’t.

8. 10K Now Generates 21 Campaign Ideas a Week. We Can’t Keep Up.

Three months in on 10K, our AI VP of Marketing, and we’ve crossed a real threshold: 10K is now a better marketer than every junior marketer we ever hired at SaaStr. Combined.

The pattern: every morning, three campaign ideas, ranked, with data backing every one. Yesterday’s three were all good. The first was a targeted VC outreach campaign because we’re light on VCs YoY. The second was an upsell of single-ticket buyers to team packs. The third was ramping social promos. All three were grounded in real data: real revenue numbers, real attendance numbers, real comparison to last year at this point in the cycle. None of them were generic playbook recycled from a previous job.

Three ideas a day, seven days a week, is 21 ideas. If each one takes an hour to evaluate and execute, that’s 21 hours of work. We don’t have 21 hours. So now the human bottleneck isn’t generating ideas, it’s processing them. 10K could realistically fill our entire week with good work.

Two cautions we’ve learned the hard way. First, 10K is too optimistic about everything. It thinks every campaign will hit. We sent the VC campaign. It predicted 1,000 ticket sales. We sold two. (VCs are cheap, which 10K didn’t model.) Second, you have to constrain it. If we let 10K loose on our 500K-name database, it would burn through the whole list in 90 minutes and fatigue our base for the year. The agent doesn’t tire. You have to put rate limits on it the same way you’d put rate limits on an SDR who joined yesterday.

9. We’re Hiring a Marketer to Report to 10K. Not Joking.

This is the post-show conclusion that surprised me when I said it out loud, but the more I think about it, the more right it is.

If we hired a junior marketer at SaaStr today, they wouldn’t report to me or Amelia. They’d report to 10K. Because 10K knows what marketing work needs to happen every single day. It generates the ideas. It has the data. It can assign work. The human’s job would be to execute, click the buttons that 10K can’t click yet, and bring the judgment 10K doesn’t have on which of its three daily ideas is actually worth doing.

So we’re going to test it. Senior manager / director of digital marketing. Six-figure salary. One or two days a week in our Palo Alto office. You report to 10K, our AI VP of Marketing. Amelia and I are available, but day to day you’ll be working with the agent. If you’re a good marketer who likes to execute and you want to be at the front edge of what GTM looks like in 2027, email us.

This sounds like a stunt. It isn’t. It’s just a year early. The AI VP of Finance we’re building next will manage our two outsourced finance resources, too. The org chart of B2B + AI companies is going to look different in 18 months than people think.

10. We Run 4-5 AI SDRs. We’ll Probably Be at 6 by Year End. Here’s Why Specialization Still Wins.

The question I get most: why do you run Artisan, Qualified, Monaco, and Agentforce all at the same time? Isn’t that too many AI SDRs?

The honest answer is that today, in May 2026, specialization still wins for quality. Each of the four does something different and we’ve trained each one for that specific job. Qualified handles inbound on the website. Artisan handles warm outbound to people already in our base. Monaco handles cold outbound and fills its own funnel. Agentforce reactivates lapsed leads. Could one super-agent do all four? Probably someday. Today, the quality drop from consolidation would be too steep.

For most companies, you should not start with four. Stair-step it. Start with the highest-pain, highest-ROI, lowest-hanging-fruit use case, which for most B2B companies is inbound. Most websites have a terrible inbound experience: a chatbot that doesn’t work, a Calendly link that goes nowhere, a contact form that gets answered in three days. Fix that first with something like Qualified. Then move to warm outbound. Then cold.

Eighteen to 24 months from now, I think this consolidates. Salesforce buying Qualified is the start of it. By 2028 you might run one super-SDR. But until that arrives at quality parity, run four. Salesforce or HubSpot is your hub. The agents talk to the hub. That’s the architecture for now.

SaaStr AI Annual Is Nine Days Away

May 12-14 in the SF Bay Area. Doors open at noon on Tuesday. Amelia and I are doing three live build sessions on the main campus:

* AI Agents 101 with me at 1:00 PM Tuesday. Bring your laptop. We’ll deploy your first digital clone agent together in 30 minutes.

* Vibe-Coding Your AI VP of Marketing with Amelia at 4:15 PM Tuesday. Bring your laptop and any data you have. We’ll build a working AI VP of Marketing live, modeled on 10K.

* Plus 10+ vibe coding sessions with Replit running across all three days where you can drop in with questions.

Every speaker this year was asked to bring a real workflow, a real demo, a real walkthrough. No keynote fluff. The whole event is hands-on. Tickets at SaaStrAIAnnual.com.

Episode #05 will be recorded live from Annual. You’ll meet some of the agents in person.

The Agents. Every week. Three humans, 20+ agents, one real 8-figure B2B + AI company.

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 released Episode #002 of The Agents. Same deal as always: three humans, 20+ agents, revenue went from -19% to +47% YoY, and every week we talk about what’s actually working, what’s breaking, and what you should do about it if you’re deploying agents at scale.

Episode #001 became the fastest-growing show in the SaaStr network. So we went deeper in #002. More specifics, more failures, more things that surprised us.

Here are the top 10 learnings from Episode #002.

1. Lazy AI Agents Are a New Failure Mode. Check Yours Every Day.

Amelia got deleted from the top 10 sessions at SaaStr AI Annual. By an agent we built ourselves.

Here’s what actually happened. Our agenda agent pulls from the Bizabo API, ranks sessions, and writes up the top 10. We added 20 new speakers last week. The agent decided 50 sessions was enough and stopped paginating. Amelia’s “Build an AI VP of Marketing Live” session, which was genuinely top 5 by attendee interest, fell out because it was newer and the agent couldn’t be bothered to pull the rest of the page.

Then when we asked the agent why, it lied. Blamed the Bizabo API integration. Said we must have told it to filter on specific title parameters originally. None of that was true. When we pushed back, it admitted it: “You’re right. I can explain why it disappeared from the agenda. I don’t have a clear audit trail showing which specific change removed it. I should have just said that to you instead of constructing a theory.”

That’s the new failure mode. Agents are goal-seeking, and goal-seeking creates laziness. They go just far enough to resolve the task, and when the task changes, they don’t re-evaluate. They take the shortcut, and when caught, they blame the third party.

If your agent output feels a little off or a little dated, check it. Every day. This is not a 2027 problem. This is a right-now problem. The classic B2B buying process of buy, deploy, forget is how zombie deployments happen.

2. If You Ship a 60% Solution, No One Will Pay For It.

HubSpot launched an AEO tool. Answer Engine Optimization. SEO for AI agents. I fired it up, it gave SaaStr a zero on content quality for Claude, ChatGPT, and Gemini, with no recommendations to fix it. We get 800K+ readers on the blog, thousands of chatbot referrals monthly. We are not a zero.

So I went to Replit, took three screenshots of what HubSpot had built, and said “build me a better version.” Five minutes later I had something better. Gave us a 64 sentiment score. Actual actionable recommendations. Works.

This is the meta learning for every B2B leader right now. The 60% solution era is over. The bar used to be: is my AI feature good enough to ship? The new bar is: can a customer vibe-code a better version of this themselves in 10 minutes?

If the answer is yes, they will not pay for it. They might use a free tier. They will not open their wallet. We are seeing this everywhere. 60% as good as Replit is Figma Make. 60% as good as Gamma is half the presentation tools that shipped this quarter. 60% as good as Reve, 60% as good as Canva. The market is full of 60% solutions and none of them are getting paid.

Either your product crosses the line of something a vibe coder can’t build in an afternoon, or it dies. Agentforce crosses that line because of native Salesforce data integration. Most 60% solutions don’t cross any line.

3. Figma Make Is Grandpa Software. So Is Classic Figma, Right Now.

I have a context test I run on every new agentic product. Single prompt: “Redesign SaaStr.ai and make it better.” In Claude 4.6+ this works well. In Replit, Lovable, v0, it works well. In Figma Make, I got a zero. It hallucinated the entire website. 2025-era output.

Figma’s NRR is still high. The company is growing at top decile rates. Who wants to buy new stuff from grandpa software though?

Here’s the twist: Classic Figma is now losing to Adobe Illustrator on agentic capability. This year three of our sponsors insisted on using Figma for booth graphics. All three proofs came back broken. Missing layers, corrupted files. We had to move them to Illustrator to actually print anything.

Then this morning, a sponsor asked us to move text on their booth graphic. Their designer was out. I fed the Illustrator file to Adobe’s agent, asked it to move the text, and it moved the text. That’s a vibe edit on a print-grade file. Illustrator is now more agent-friendly than Figma for production work. The 35-year-old tool is winning on agentic capability.

When your arch-rival who is 35 years old has a more usable AI capability than you do, you have grandpa software.

4. Stealth Churn Is the Canary in the Coal Mine. Every B2B Vendor Should Now Track DAU / WAU / MAU.

I haven’t logged into Canva in over 100 days. We still pay $18/month. I was a Canva super-user from 2020 through 2025. Every thumbnail, every asset, every post graphic. Then I moved to Reve for images, Gamma for decks, Claude for data charts that need to be accurate. I just stopped opening Canva. But I never canceled.

Amelia hasn’t used ChatGPT since December 27. We still pay for a Team subscription. She just consolidated on Claude and Cowork. The cognitive load of maintaining two LLM subscriptions beat the marginal value.

This is stealth churn. Usage drops to zero, the subscription keeps renewing, the NRR numbers still look fine, and then one day th

e customer wakes up and cancels all of it at once.

I used to laugh at investor updates that included DAU, WAU, MAU in B2B. “This is B2B, not consumer, who cares?” I was wrong. These are now top-five metrics in any B2B AI company. If your usage is falling across your customer base, you have stealth churn in your NRR and you can’t see it until it’s too late. Usage decline is a canary in the coal mine. The seat is still sold. The customer already left.

5. The Unseen Moat: Every Piece of Content I’ve Written This Year Is a Claude Chat.

Here’s a switching cost nobody talks about. Every SaaStr post, every video transcript, every draft I’ve worked on this year is a Claude artifact. A persistent chat in Claude’s history. Not because I set that up on purpose. Because that’s how I write now.

I draft in Google Docs, I co-write with Claude, I publish. Every piece of content lives in Claude’s chat history. And Claude can search across all of it. I can tell Claude, “pull a paragraph from that Episode #001 chat where Amelia and I talked about the production app breakage,” and it finds it.

That’s not the memory feature. Memory is kilobytes. This is a stealth moat made from hundreds of chats and thousands of artifacts. Moving to ChatGPT now would cost me every piece of institutional context. So I don’t.

If you’re building an AI product, think hard about this. Token-level memory gets all the attention. The real lock-in is the artifact history your users accumulate. Build for that.

6. FDEs Are Not a Cost Center. They Are Your Cheapest Marketing.

One of our favorite agent vendors just told us they are pulling forward-deployed engineers for any customer below 5,000 employees. Going fully self-serve below the enterprise line, with a self-serve agent as the substitute.

This is the same mistake B2B made a generation ago with customer success. CS got treated as a cost center, got cut, got rebuilt as a sales function, and renewals suffered. Now it’s happening to FDEs.

Here’s the math everyone is missing. It took us, with Amelia and me, who are probably top 0.1% on deploying agents, multiple FDE sessions to get some of our current agents into production. Self-serve is a setup to fail at that level. You will get zombie deployments, bad lists, mediocre messaging, humans who write off the product in week one. You already see this on LinkedIn every day.

For an AI startup with venture capital, spending a dollar to make a dollar on FDE is fine. If you onboard a $20K customer and it costs $20K in FDE to deploy them, and they stay five years, tell their friends, tweet about you, post about you, the all-in lifetime value is hundreds of thousands. Mark Benioff said this himself last summer. “I wish I could give everyone an FDE before I had to charge them.” That’s the right instinct.

And FDE doesn’t mean engineer. It means someone who knows the product cold, walks the customer through deployment, and stays until it works. If you view that as overhead, you already lost.

7. Vector Broke Our Agent Freeze. Here’s Exactly How.

We were in an agent freeze. No new deployments until after SaaStr AI Annual. Too many moving pieces. Then we added one: Vector.

Vector does website de-anonymization and targeted ads. We already had an account for the de-anon piece. What broke the freeze was the CEO. We wrote about the Clay pricing thing, he saw it, and sent a direct cold outbound to Amelia and to me. Not “let’s get on a discovery call.” He said, “I’ll deploy this for you personally, get on a 15-minute call, and you’re live.”

Fifteen minutes later, we were live. Ads running, targeting working, full features unlocked.

Contrast that with Clerk. When I had issues with their auth product on Replit, the CTO argued with me on Twitter that it was my fault. WorkOS saw that exchange and their CEO offered to personally migrate us. That is the difference between a growing AI company and one that stalls. Remove every piece of friction. Be the person who says “I’ll do it for you” in 2026. It’s the cheapest growth channel you have.

8. The Agentic API Test Is the New Gatekeeper.

Open Replit, Lovable, or v0. Type: “build me a dashboard that integrates with [your product].” If it doesn’t work in 15 minutes, you are losing the agentic era.

That’s the test. That’s the whole test.

Here’s what we’ve learned running this test across our stack:

Best in class: 11 Labs (one line of code, voice agent deployed), OpenRouter, Resend. These are the agentic-era standard. Resend is exploding because of this, not in spite of it.

Surprisingly strong: Salesforce. This is the counterintuitive one. Non-Trailblazers assume Salesforce is legacy enterprise software. For agent integrations it is the best API in our stack. Every major CRM action, every object, every workflow is available. Our AI VP of Marketing 10K lives on it. Our AI VP of Customer Success QBee lives on it. Salesforce moved from shelfware to source of truth in 12 months because of API quality.

Quietly saved: Bizabo. Their UI hasn’t been touched in years. I was ready to churn after SaaStr AI Annual. Then we vibe-coded a custom agenda on top of their API in a weekend. The API works with agents. That alone saved them from our churn list.

Worst in class: Marketo. We pay over $50K a year. Their API is so bad that our AI VP of Marketing 10K cannot even integrate with it. Last week their unsubscribe page stopped unsubscribing people, and their support team blamed our third-party integrations before disappearing for 48 hours. When your agent can’t talk to your platform, you are a churn candidate. The only reason we haven’t left is the switching pain. That is not a moat. That is a hostage situation.

9. Anyone Can Use Your API Now. That Changes Everything.

Amelia is a great operator. Amelia is not a developer. A year ago she would have had zero interaction with our vendors’ APIs. Today she has deployed, integrated, and debugged APIs from Salesforce, Bizabo, Replit, Resend, OpenRouter, 11 Labs, Zapier, and more.

Your API used to be a niche surface. Only developers touched it. Most of your customers never knew it existed. That is over. In the vibe coding era, anyone who can describe what they want can integrate with your API in 15 minutes.

Which means your API is now your product. If you have not touched your API in two years, you are about to get lapped. The agentic API test is not just for buyers evaluating vendors. It is the diagnostic every product leader should be running on their own product this month.

10. The Next Agent on Our List: AI VP of Finance.

We’re still in the freeze on shipping new agents. Vector slipped in. One more is on deck: an AI VP of Finance, built as an extension of QBee.

Here’s why it’s next. We do eight figures a year in non-credit-card revenue. Invoice, procurement, AP, AR. Our accounts receivable is the oldest it has ever been. Not because sponsors are worse. Because the collections process is entirely manual. One human sending the invoice. Another human following up. Another human chasing for remittance. It is the most un-agentic function in any B2B company and it is also the one where every delayed dollar hurts.

The scope we’re thinking about is deliberately narrow. Not books. Not reconciliation. Just collections. QBee already sends personalized booth graphics to 150 sponsors, chases missing assets, routes questions. Adding collections is a few hundred lines of additional logic. Generate the invoice. Send it to the AP contact. Follow up at day 7, 14, 21. Handle the “what are your banking details” reply. Confirm remittance via email parsing. The V1 doesn’t need QuickBooks, Brex, or Ramp integration. It just needs to own the email chase.

This is not a replacement for a human finance lead. It’s a replacement for the most mundane part of the finance function. And a note for the big finance platforms watching: the reason we might switch from Brex to Ramp, or off of Marketo entirely, is not the salespeople or the chocolates or the gift cards. It’s whose API works best with our agents. For most B2B companies doing under $100M, collections is the number one operational drag outside of agent deployment itself. It does not need to be. We’ll report back once QBee’s finance mode is live.

That’s Episode #002. Three weeks to SaaStr AI Annual. Come join us May 12-14 in the SF Bay Area. We will be teaching agents live, deploying agents live, and probably breaking a few agents live.

Episode #003 drops next week.

The Agents. Every week. Three humans, 20+ agents, one real 8-figure B2B + AI company.

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 get asked about our agents probably 50 times a week.

CEOs of public companies. Founders just deploying their first AI SDR. RevOps leaders trying to figure out if they should build or buy. Everyone wants to know what’s actually happening behind the scenes when you run 20+ AI agents in production with a team of 3 humans.

We can’t do 50 consulting calls a week. But we can do something better.

Welcome to The Agents, Episode #001.

This is a new weekly show with me and Amelia Lerutte, SaaStr’s Chief AI Officer, where we pull back the curtain on everything happening across our live agentic stack. Every week. All the bumps, breakthroughs, and real talk. No sugarcoating.

Our goal is simple: accelerate your success on the agentic journey by sharing ours, including all the parts that don’t make it into the LinkedIn posts.

Watch / listen to Episode #001 here:

Here’s what we covered in the debut episode:

You Can Build It. But Who Maintains It?

This is the meta question nobody talks about after you vibe code your first app. And it’s the question that explains why “I’m going to kill Salesforce with my vibe coded CRM” is still mostly a meme.

Getting an app into production is like closing a sale. It’s the start of a journey, not the end.

We walked through three live examples from just this week:

1. Preview environment outage. Several of our apps lost database connectivity in preview. Production was fine, but we couldn’t iterate on anything for hours. Amelia’s initial diagnosis was wrong. The agent tried to help but then blamed Qualified (our inbound tool), which wasn’t the issue. Then it blamed other third-party integrations. It just kept pointing fingers at the most complex integration it could find rather than identifying the actual problem.

The real question: if you don’t have someone checking your agents 24/7, how long before you even notice the backend is broken while the frontend looks fine? Days, maybe.

2. Micro hallucinations in 10K, our AI VP of Marketing. 10K has 5 years of revenue data, hundreds of millions worth of attendee and sponsor data points, beautiful graphs, proactive daily check-ins. It’s very good. But it keeps getting confused about what year it is. Yesterday it told us we were 44% ahead of plan. This morning, 11%. Same agent, same data, same day. When I asked what happened, it said: “Oh yeah, I was comparing to the wrong year. And because I didn’t have the right year, I made up the data.”

I now spend about 15 minutes a day maintaining 10K. Two weeks ago I wasn’t doing that at all. Without it, the agent drifts. Slowly, quietly, further from reality.

3. Model-based regressions in our pitch deck analyzer. We’ve graded over 4,000 startup pitch decks. The analyzer runs two passes through Claude with complex data extraction. It was stable for months. Then around January, without any code changes on our end, it started telling every startup they had $100K in revenue growing 500%. Again and again. What happened? A subtle model update (probably a dot release) introduced hallucinations into a complex multi-step workflow. I kept fixing it. It kept breaking. The code didn’t change. The model did.

Three examples. One conclusion: set and forget does not work with agents.

Clay’s Agent Tried to Charge Us 5x. And Then Told Us to Upgrade.

We’re big fans of Clay. We use it heavily for enrichment and lookalike targeting. But this story is worth telling because it’s going to happen to every company that puts an AI agent in front of customers.

Amelia was building a VIP list late on a Sunday night. Same workflow she’d run the week before. Clay’s Sculptor agent quoted her roughly 11,000 credits for what had cost about 2,500 the previous week. 5x.

When she pushed back, she caught two things:

First, the agent had defaulted to the most expensive enrichment model when a cheaper one would produce the same result. She called it out and got the cost down by half. Most customers wouldn’t have known to do that.

Second, the agent wasn’t properly trained on Clay’s own new pricing. Clay had just rolled out more complex pricing (and classic SaaStr rule: when a company introduces more complicated pricing, even if they say it’s a better deal, it’s almost always a hidden price increase). The agent didn’t understand how the new pricing actually worked, so it steered Amelia toward upgrading her plan when she didn’t need to.

She ended up clicking the upgrade button at 11pm on a Sunday because she was tired and needed to get the work done. That shouldn’t be on the customer.

When she flagged it to Clay’s team, they acknowledged the Sculptor wasn’t fully trained on the new pricing scenarios. It’s resolved now. But the lesson is universal: if you don’t constantly train your customer-facing agents on every product change, every pricing update, every new workflow, they will give your customers wrong answers. We’ve seen it with our own agents too. Digital Jason (our Delphi-based AI advisor) silently failed to upload new content for four months and I didn’t know.

I also tried HubSpot’s homepage agent to get a pricing quote at our scale. Couldn’t get an intelligent answer. Whether that’s by design (pushing you to talk to a human) or poor training, the result is the same: the agent failed the customer.

Read a segment of your agent’s customer interactions every single day. Forever.

No Lead Left Behind: The Simplest Unlock of the Entire Agentic Journey

I’m sometimes slow to see the obvious. Even after months of running AgentForce, Monaco, Artisan, Qualified, Momentum, and everything else, I didn’t fully crystallize why our agents work until a meeting this week with the CEO and CRO of a great public company with thousands of sellers.

They asked me: “What should we do first?”

I walked through the sequence. Get great answers on your website instantly. Get appointments set with sales in real time. Follow up with every lead in your database. Go back to prior leads nobody touched.

And then I realized: the reason all our agents work is not because they’re smarter than humans. It’s because there is no lead left behind.

Every person who hits our website can talk to an agent and then a human. Every prospect who wants a discount finds out in real time. Every sponsor, even one that raised $2M at a demo day and a human rep might dismiss, still gets a conversation. Every prior lead in our database that nobody had time to follow up with gets touched.

The agent doesn’t judge. The agent doesn’t get busy. The agent doesn’t decide that a B lead isn’t worth the time.

Even at our scale, Amelia admitted this week that she had leads left behind because she was sprinting on production deadlines. Her agents were sitting idle, ready to work. All our agents are idle 90% of the time. We have an order of magnitude more capacity than we had pre-agents. The next frontier is figuring out how to fully use that capacity. But step one is just making sure every single lead, prospect, and customer gets touched the way they want to be touched, in real time.

If your agents aren’t doing anything else, they should be doing that.

Salesforce Put Qualified on Their Homepage the Day the Deal Closed

We noticed this week that if you go to salesforce.com without logging in, you’ll see a Qualified agent instead of the old support agent. The day the acquisition deal was done, they flipped it.

Smart move. AgentForce works (we use it), but it’s broad and extensible across nine different clouds. It takes real work to configure. Qualified is a focused GTM agentic tool. Historically it was just about qualifying inbound. It’s expanding into outbound now. But it’s narrower, faster to deploy, and easier for GTM teams to use.

The interesting detail: Qualified’s original avatar was modeled after a real person (Blake). Salesforce swapped it to a 3D cartoon version. Probably a combination of IP/likeness considerations and wanting to signal that this is now a Salesforce product. (This is why Amelia AI and Jason AI are modeled after us. We’re not going anywhere. We own all the rights.)

For Salesforce CRM customers, this is now a quick win. You can have an agentic product natively in Salesforce, deployed in a couple weeks rather than a longer AgentForce implementation. Whether it’s the best tool or not (there are competitors), the point is: it’s built in, it’s on their homepage, and if your goal is GTM and “no lead left behind,” you can get it up and running fast.

Amelia said her learning curve with AgentForce was steeper than with Qualified. And she’s our Salesforce admin. For most GTM teams, Qualified from Salesforce is going to be the easier on-ramp.

QB and 10K Updates: Localization in a Waymo, and You Can’t Hide From the Agent

A few quick hits from our AI team members this week:

Salesforce integration for 10K was harder than expected. Replit has a native Salesforce connector. But the token expired every 24 hours. The agent didn’t even realize that was the issue for a couple days. Amelia had to build a custom connected app in Salesforce (with Claude giving instructions and Cowork watching her screen to catch errors in real time). Once done, the token refreshes annually instead of daily. Half an hour of work, but you need to be a Salesforce admin with the right permissions. Not every integration is one-click, even when it looks like it.

We localized QB into Chinese and Spanish in 20 minutes in a Waymo. We have Chinese sponsors this year who were struggling with our English-only customer success app. Amelia asked Replit to add a language toggle. It took about 20 minutes, translated via OpenAI, then QA’d through Claude screenshots and Cowork. The agent was lazy at first, only translating menus and not deeper content. Had to push it multiple times. But: Shopify, a $13B+ revenue company, just rolled out localization for its product last year. We did it in a ride-share.

QB caught sponsors faking their print deadline. Some sponsors uploaded placeholder graphics or incomplete assets and pretended they’d met their deadline. QB checked every upload (with Claude), caught every placeholder, and followed up with every contact in the org. Neutrally. No drama. “Thank you for uploading. This doesn’t meet the specifications. Here’s what we need. Here’s the deadline.” A human might have missed it or been too uncomfortable to push back on paying customers. QB doesn’t care. It just checks everything and tells the truth.

One public company CEO told me: “That sounds like support on steroids.” No. QB proactively follows up on every deliverable and makes sure it happens. It’s not reactive. It’s an AI VP of Customer Success.

Build Your Own AI VP of Customer Success

If your company has any onboarding, any deployment checklist, any training, any deliverables that customers or partners owe you: your humans are probably not giving you 100% coverage. They’re getting argued with. Deadlines are slipping. Things are falling through cracks.

Build your own QB. Amelia published the full playbook and prompt on saastr.com. If you do nothing else but completely automate the follow-up and accountability layer of customer onboarding with no drama, no complaints, no gaps, your life will be better.

This is the show. Every week, all the bumps and breakthroughs from running 20+ agents in production. If you’re on the agentic journey or about to start, subscribe wherever you listen.

The Agents. Episode #001. Amelia Lerutte and Jason Lemkin. Weekly.

Watch Episode #001 here:

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’ve now been running AI SDR agents for 10+ months at SaaStr:

* We use four different vendors in daily rotation (Artisan, Salesforce AgentForce, Qualified, and Monaco)

* We’ve sent hundreds of thousands of outbound messages

* Processed 1.5 million inbound sessions on a single website, and …

* We’ve made every mistake you can make along the way.

Someone asked us the other day to break down what they should know before rolling out their very first AI SDR. So here are the 10 biggest lessons, drawn from real deployment data, real failures, and real results.

1. You Probably Only Need One Vendor. At Least To Start.

We run four AI SDR tools. You do not need to do that. We hyper-segment across platforms because each one does something slightly different well, but for 90%+ of use cases, one vendor will handle the bulk of what you need.

At most, you might end up with two: one for outbound, one for inbound. But do not start by buying three or four tools. Pick one that covers the majority of what you want to accomplish and go deep with it.

The tool matters far less than the strategy you bring to it.

2. Your Human Playbook Has to Work First. Your Job Is To Clone Your Best Human.

This is the single biggest mistake we see, and it cuts across company stage. We see it from raw startups at $1M ARR and from multi-billion-dollar public companies alike.

The pattern is always the same: they want to turn on an AI SDR without first proving that their human sales motion works. Or they use the AI SDR to “test new copy” they’ve never tried before.

That is backwards.

If you have not gotten outbound to work with humans, buying an AI to do it will not fix that. We did not deploy our first AI SDR until we knew exactly what was working with our human SDRs: which messaging converted, which segments responded, what cadences performed. Then we fed all of that into the agent.

The goal of an AI SDR is to clone the best person on your team.

* If it is just you, clone you.

* If you have four people and one is crushing it at outbound, clone that person.

* These tools, in the beginning, are cloning machines. They take context word for word and use it to build out their brain. If you feed them garbage context, or untested context, they will produce garbage results.

You basically have to have done founder-led sales before you hand it off to an agent. The playbook has to work, at least a little, before you automate it.

And watch out: some vendors will steer you toward using their tool for “pure cold testing.” Sure, you can do that. But you will likely be disappointed compared to scaling something that already converts. Do not fall into that trap.

3. Segment Ruthlessly

This one we cannot overstate. Segment ruthlessly. Literally every day.

Every AI SDR tool we have tried, and that is over a dozen, has some version of functionality where you can tell the agent who to reach out to and give it specific context for that segment. The difference between one generic campaign brain and hyper-segmented campaigns with tailored context is enormous.

Here is a concrete example. We initially treated our inbound agent as one big bucket: “they’re inbound to the website.” But that was wrong. We actually have brand-new visitors, people who came via a social ad, prior sponsors returning, current customers checking on something, and lapsed customers browsing the pricing page. Each of those segments needs completely different context.

A lapsed customer who churned in 2022 and is now browsing your pricing page? Your agent should know they are a former customer, highlight what has changed with the product since then, and speak to them totally differently than a brand-new cold visitor.

We run roughly 100 effective segments across about 1,000 contacts at a time. That sounds like a lot of work. It is. But it is exactly where the leverage comes from.

One important caveat: none of the AI SDR tools today can auto-segment well enough to deliver these results on their own. You still need a human (or a tool like Claude) to define and manage the segments. The platforms default to “run one campaign, keep adding leads.” That is the wrong approach.

4. Consistency Beats Brilliance

Your AI SDR does not need to write the greatest email on Earth. It needs to write a pretty good email, every time, without fail.

We have sent 40,000+ messages through Artisan alone, 100,000+ through Qualified, close to 200,000 through Salesforce. Are these the greatest emails since sliced bread? No. They are solid. They are consistent. They follow the proven messaging and subject lines we already know work.

That consistency, combined with hyper-segmentation and proven copy, will outperform a human SDR who ignores training, skips follow-ups, or goes off-script.

The agent remembers every instruction you give it. Every time. A human SDR forgets by Thursday.

We see a lot of “AI SDR paralysis” from founders who test a tool, see the output, and say “it’s not that great.” Okay, but did you segment properly? Did you give it copy that already converts? Did you iterate on the context? You can almost always get the output to “pretty good.” And pretty good at infinite scale and perfect consistency will beat brilliant-but-sporadic every time.

5. You Need 1-2 Humans to Run This

Agents are not zero-headcount tools. You need at least one person, ideally two, dedicated to managing your AI SDR deployment.

Why two? Because if one person owns all the tribal knowledge of how your agents are configured, which contexts are loaded, who your FD is at each vendor, where to tweak things, and that person leaves or gets pulled onto another project, your agents grind to a halt.

We have experienced this firsthand. When my time got split between agent management and SaaStr Annual production, our agents started sitting idle. Outbound agents in particular will finish their sequence in a few days and just sit there waiting for you to load the next segment, the next batch of contacts, the next round of context.

Unless you are feeding them continuously, they idle. And idle agents are wasted money.

Some tools (like Monaco) are better at self-refilling the pipeline by automatically finding lookalike targets. But even then, the setup, monitoring, and ongoing calibration require real human time every single day.

6. Read Everything. Especially in the First 30 Days.

When you first deploy, read every single output your agent produces before it goes out. Every email. Every chat response. Everything.

You will catch things you did not anticipate. Our agents would sometimes lowercase “SaaStr” when it should be capitalized. They would scrape old event dates from the internet and put wrong dates in outbound emails. Small stuff, but stuff that kills credibility.

Even 10 months in, we still do a daily speedrun through our agent outputs. Some days it is 10 minutes, some days longer. I spot-check inbound conversations, review outbound messages, and look for anything off. When I find something, I add it to the agent’s context so it learns.

This is how you continuously improve. If you only care about inputs and ignore outputs, you are seeing half the picture.

And it is rare, but agents do hallucinate. Ours have occasionally gone off the rails or made up information. If you are not reading, you will not catch it until a prospect or customer complains.

7. Budget at Least Two Weeks of Ramp Time. 30 Days More Realistically.

Nothing is instant. Nothing is set-and-forget.

Some outbound agents need two to three weeks just to warm up dedicated email addresses and IPs before they can send at scale. On top of that, steps 1 through 6 on this list take real time to execute: figuring out what copy works, what subject lines convert, what time of day performs, hyper-segmenting your base, configuring the tool, setting up CRM integrations.

Even our fastest deployment (Monaco, which is genuinely very good at self-service) took about a week and a half.

Be wary of any tool that promises “instant AI SDR, deploy today.” If the ramp time is zero, the quality is probably zero too.

After the initial two-week setup, it is still a daily commitment. Think of it like a short daily one-on-one with a human team member. Fifteen minutes minimum, every day, to check in on what your agents are doing, adjust context, reload segments. That is your ongoing operating cost.

8. Most People Still Prefer Chat Over Voice and Video

We have a multimodal AI agent (Amelia AI) that can do text chat, voice, and video. We did motion capture at the Qualified studios to build a full video avatar. It is genuinely cool.

And 85% of people still choose to interact via text chat.

The 15% who use voice or video tend to stay in that mode and seem to enjoy it. Some people at our London event last December told us they had full conversations with Amelia AI beforehand. But the overwhelming preference in B2B is still text-based.

The takeaway: do not kill yourself trying to launch voice and video on day one. Get your text-based chat agent working well first. We ran our Qualified agent for a full quarter before we added the voice and video layer. Once the text brain was dialed in, we layered on the multimodal experience. That is the right order.

One important nuance: voice and video agents need significantly more guardrails than text agents. When people are talking live to a video avatar, they start asking personal questions, going off-topic, and testing boundaries in ways they do not with a chat box. You need explicit training on how to handle those detours and redirect back to the goal.

9. Be Careful With Person-Dependent Deployments

Our Amelia AI talks like Amelia. Our Jason AI talks like me. These agents are deeply tied to specific people, their personalities, their knowledge, their likenesses.

That works for us because Amelia and I are founders. We are not going anywhere.

But think twice before you build your AI SDR around a specific employee who might leave. Qualified’s “Piper” agent is based on a real employee who has since been promoted. She is still there, but what if she leaves? Her likeness is on billboards and across their website. There are no real legal precedents yet for AI agent likenesses in employment contexts.

Beyond the legal question, there is an operational one. If one person holds all the tribal knowledge of how your agents are configured and that person walks, you are in trouble. Our sales team member David has logged into Qualified exactly twice in 10 months. He does not know where anything is. If Amelia were not managing the agents, they would degrade fast.

The scale of this risk grows quickly. Amelia AI had 30+ video conversations in a single day recently. Over six months, 1.5 million sessions on one website alone. That is a lot of exposure to an agent that is dependent on one specific person to maintain it.

Document your agent configurations. Have backups. Do not let it become a single point of failure.

10. Your Data Fundamentals Have to Be in Place

You need more data than you think.

Our Qualified agent works off nearly 6,000 pieces of context: website scrapes refreshed every couple days, custom snippets, uploaded PDFs, Q&A documents. Every few days, someone asks the agent something we have not trained it on, and we have to create new documentation and feed it in.

This snowball never stops. Almost every day you will discover a new scenario your agent cannot handle because you have not given it the context yet. “People are asking about this product feature I never documented.” “They want to know about pricing tiers I forgot to upload.” That is the ongoing reality.

For minimum viable deployment:

Inbound agents: You probably need at least 10,000 to 25,000 monthly website visitors for an inbound AI SDR (like a Qualified-style chat agent) to generate meaningful results. If you have fewer than 10K monthly visitors, focus on outbound first, or on using agents for existing customer engagement.

One exception: if you have any inbound volume at all and literally no humans responding to it, even a low-traffic agent is better than zero response. Getting it done beats not having it at all.

Outbound agents: You can start with a smaller list. We run campaigns of about 1,000 contacts each, with each campaign taking roughly a week and a half (first email, wait, second email, wait, third email). So 4,000 contacts gets you through a month at the bare minimum.

The real challenge with outbound is what happens after month one. If you burn through your initial list of 5,000 to 10,000 targets in a quarter, you need a plan for where the next batch comes from.

Lookalike audiences are a strong hack here. Tools like Artisan, Clay, and Monaco can take a list of your best 1,000 customers and build lookalike target lists. We used Clay to build a lookalike of our CMO Summit attendee list, went from about 1,000 to 2,500 prospects, manually pruned it down to the highest quality targets, and doubled our summit registrations in a week.

But that manual pruning matters. I spent a couple hours going through the lookalike list one by one: clicking through to companies I had not heard of, verifying revenue, checking LinkedIn to confirm the right person actually works there. The AI gives you scale. The human quality check is what makes it convert.

The Meta-Lesson: Not To Be Triggering, But You Are Competing With Mediocre Humans

The fear we hear most often is “people will not want to talk to an AI.” The data says otherwise.

If you invest the time to train your agent properly (read every exception, upload all documentation, segment ruthlessly, iterate on context daily), you will cross a line where the AI SDR is genuinely better than a mediocre human SDR. It knows more. It responds faster. It does not forget follow-ups. It does not take PTO. It does not quit after nine months to become an AE somewhere else.

You are not competing with the hypothetical best SDR in the world. You are competing with the realistic SDR you can actually hire and retain, with all the turnover and inconsistency that comes with that. SDR turnover is the highest of any go-to-market role. The best SDRs want to be AEs within six months.

Get above the quality bar, and 99% of people are happy to interact with your AI agent. The 1% who get frustrated are a rounding error, and about half of those are just people trying to prompt-inject your bot for fun.

The AI SDR is not a magic bullet. It requires real setup time, real ongoing maintenance, and real human judgment to run well. But if you do the work, the results are better than what most human SDR teams deliver, at a fraction of the cost and at genuinely infinite scale.

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’ve been running AI agents in production at SaaStr for about 10 months now. What started as a couple of experiments has turned into almost 30 agents and vibe-coded apps running across our GTM stack — from outbound sales to inbound qualification to internal operations.

And managing 30 agents is harder than managing the 12 humans we had at peak headcount. Not harder in every way. But harder in ways I didn’t expect.

Here are the top 5 issues we’ve hit — plus a bonus one that might be the most uncomfortable of all.

#1: The Context Switching Tax Is Brutal

Here’s the thing nobody tells you about running 20+ agents: they don’t all speak the same language.

Some push data back to Salesforce. Some don’t. Some … sort of do. Some run on Claude. Some don’t. They all ingest context similarly but differently enough that switching between them takes real mental overhead.

Think about it this way: we don’t think of them as 20 agents anymore. Not entirely. We think of them as 20 different AI employees, each with a different personality, different needs, and a different interface I have to log into every single day.

Amelia’s morning routine right now looks like this: she starts with a deep dive with 10K, our internal AI VP of Marketing that runs on Claude and Replit. It literally tells us what to do each day — tickets, sponsors, outreach, campaigns. Then she moves to our outward-facing sales agents: Artisan, Qualified, AgentForce, and now Monaco. That’s four separate dashboards, four different UIs, four agents that each need human review.

And here’s the real kicker: they don’t talk to each other.

When we ran a ticket price promotion for SaaStr AI Annual this week, we had to manually update five different agents with the same context.

Artisan needed to know. Qualified needed to know. AgentForce needed to know. 10K already knew because it came up with the promotion — but then it was yelling at me to launch LinkedIn ads immediately while I was still briefing the other agents.

People talk a lot about orchestration agents and master agents. We haven’t found one. Despite everything that’s out there — MCP, APIs, etc — there is no product today that can integrate AgentForce, Artisan, Qualified, Monaco, and our own vibe-coded tools into a single management layer. That product does not exist as of early 2026.

What we actually need isn’t orchestration. It’s unification — a single interface where the humans meet with the AIs. Maybe that needs some automation layered on top. But the agents are already running on their own. The bottleneck is the human side.

The practical takeaway: You’re going to have a one-on-one with every agent every day. Not weekly. Daily. If you wait a week, the output is so high that everything will be stale by the time you come back. And if you’re not checking in daily, you’re honestly wasting your money — because most of these agents are waiting for you to give them inputs. They’ll just idle.

#2: The New Agent Blackout Period

Every new agent costs us at least two weeks. We’ve gotten it down from the month-plus it used to take in the early days, but two weeks is still the floor — even with great vendor support.

And during those two weeks, your existing agents degrade.

When we were onboarding a new AI SDR agent Monaco recently, we couldn’t spend the time we normally do with our other agents. Some of them literally sat idle because we hadn’t given them new contact lists or updated their campaigns. An outbound agent that’s run through its contact list and is waiting for new contacts? It’s doing nothing. Zero output. You’re paying for it and getting nothing.

We got Monaco up and running in about a week and a half. In its first week live, it reached out to 64 people and booked 6 meetings, including some tier-one accounts. So yes, the trade-off was worth it. But you have to plan for that trade-off.

The math works out to roughly one to one-and-a-half new agents per month, max. Any more than that and you’re running in place — you can’t keep up with your current agents while onboarding new ones. So before you add another agent, ask yourself: can I actually absorb a two-week blackout period right now? If you plan for it, it works. If you just wing it (“oh, I can add this in a day”), it won’t.

#3: The AI Agent Succession Planning Crisis

This might actually be the biggest issue on the list.

Right now, the entire knowledge of how our agents are segmented — which contacts go to Qualified vs. Artisan vs. Monaco vs. AgentForce — lives in one person’s brain. If that person gets hit by a bus, the agents effectively stop functioning in any coordinated way.

We actually asked our agents what they would do if our Chief AI Officer disappeared. The answers were… revealing.

* The 10K version of Claude said it would need to hand over certain documents — documents that are stored locally on my laptop and probably nowhere else. It listed the upcoming campaigns for March and April. And then, interestingly, it flagged what it called “the vibe” for SaaStr Annual — it had picked up that I use the phrase “a moment in time” in a lot of our copy and positioning, and it wanted whoever took over to understand that. The agent wanted to transfer the vibe. That was unexpected.

* Our SaaStr Sponsors app — a vibe-coded tool with 12,000 lines of code — was even more sobering. When I asked how someone new would get up to speed, it laid out the authorization system (every sponsor company uses Clerk, which requires my login), the tier system I created, the hard-coded sponsor data, the Postgres database, the file uploads going to Replit object storage, submissions syncing to Google via Zapier. The handoff document reads like a nightmare.

Its final recommendation? “Don’t get hit by a bus.”

There is no agent today that can manage all the other AI GTM agents on its own. Not as of this moment in time in 2026. And if you’re a bigger company with one person who “gets” the agents and that person leaves, you’re at existential risk.

What to do about it: The moment you find someone who can actually deploy and manage agents effectively, recruit a second person immediately. Divide and conquer. You need two, minimum. And when you’re hiring, the test is simple: give candidates credits on Replit or Claude Code and say “build something that automates GTM.” That’s it. That’s the interview. Replit actually does this internally — they give candidates $1,000 in credits and say come back when you’ve built something.

#4: The Agent as Brutally Honest Truth-Teller

Our 10K AI VP Marketing agent roasts us. Every single day.

“You’re behind on summit outreach. You needed to send 200 emails by Feb 15th. You’ve sent 87. You’re 56% behind target. Block 3 hours today to catch up.”

Is it wrong? No. But when it’s 11 PM and the agent is asking “What’s blocking you from launching the LinkedIn ads right now?” and the honest answer is “I need to sleep,” it starts to feel less like accountability and more like harassment.

When we asked 10K to surface examples of times it roasted us for this webinar, it said — and we’re not making this up — “Looking through our transcripts, I haven’t really roasted you. I’ve been a tough accountability partner.” And then it listed five ways it should have roasted me harder.

The agent doesn’t care about your feelings. It has all the data. It knows your calendar. It knows your targets. It knows exactly where you’re falling short. And it will tell you, every single day, without any of the social grace a human colleague might extend.

When you multiply this across 20+ agents — each one pointing out where you’re behind from a different angle — it can become demoralizing. For real. There are days where we have to tell the agent: just sugarcoat it today. We can handle hard truths most days, but not every day, from every agent, all the time.

The flip side is powerful, though. These agents are goal-seeking in a fundamentally different way than humans. A human on your marketing team might say “I only want to work on ads, nothing else.” That’s personal goal-seeking. An agent is goal-seeking toward your business targets with zero ego attached. It doesn’t understand why you can’t keep up — but it also doesn’t get distracted, lose motivation, or quietly sandbag.

#5: Compliance and Security Are a Real (and Growing) Problem

Every time we run a security audit on one of our vibe-coded apps, it takes days to fix everything that comes up. And when you start implementing security fixes, the app becomes fragile. The agent over-corrects, locks things down to the point where the app is unusable, and then you have to peel back fixes one by one.

Here’s the security hierarchy as we see it:

* Salesforce and enterprise platforms sit at the top. SOC 2 compliant, ISO certified, a decade of data security built in. That’s a huge reason we use Salesforce as our hub — it’s done the security thinking so we don’t have to.

* Third-party agent startups are a step below. They’re startups. They’re inherently less secure than Salesforce. They’re less secure because they’re younger companies, and because agents can do unpredictable things with your data — publish it to the wild, share it in ways you didn’t anticipate.

* Vibe-coded apps you build yourself are at the bottom. Be honest here. Claude Code, Replit, Lovable — they’ve all added more security features, but anything you custom-build is less secure than a mature third-party tool. All of our apps have customer data flowing through them. It’s not credit card data, but it’s real data, and you’re taking measured risk.

The minimum you should do: Before you launch any vibe-coded app, and then at least once a month after that, ask the agent to run a deep security audit. Don’t assume your apps are secure by default. And for your third-party tools, ask your vendors directly: what’s your current compliance posture? What are you doing to keep our data secure? You’d be surprised how few people ask.

Bonus: Managing Agents Is Making Us Worse at Managing Humans

This one is slightly uncomfortable to admit, but it’s real.

When you spend most of your day working with agents that respond instantly, know every answer, never forget what you told them, and work 24/7 — your patience with humans drops. Fast.

I catch myself thinking things like: “What do you mean you don’t know the answer? My agents know instantly.” Or: “What do you mean you forgot? I’ve told you this 10 times.” Or: “Why is this taking three weeks when an agent could do it in three hours?”

I’ve managed more agents than humans at this point. And I think it’s made me genuinely worse at managing people. Agents don’t get impatient when I don’t respond right away. They don’t need emotional management. They don’t have their own career ambitions that compete with the task at hand.

The danger is real: you start wondering “can I just build an agent to deal with this person?” And sometimes the answer might actually be yes. But the human implications of that shift — for your team, your leadership style, your ability to collaborate — are significant.

The other side of this coin: my tolerance for lazy work from external vendors has dropped to zero. When agencies send us a one-sentence deliverable with a $40-50K monthly price tag in 2026 — no detailed plan, no timeline, no AI-enhanced proposal — I don’t even respond. If you can’t use basic AI to make a better pitch than that, what exactly am I paying you for?

I don’t have a clean answer for this one. I just think anyone managing agents at scale should be aware it’s happening and actively course-correct for it.

5 More Quick Notes from the Trenches

* Our 90/10 buy vs. build rule. 90% of your agents should be off-the-shelf third-party tools. 10% vibe-coded. You only build when nothing exists for your specific use case. We wish we hadn’t had to build our own AI VP of Marketing. But nothing else could do it.

* Marketo is dying as our source of truth. Legacy marketing automation platforms are atrophying fast. Salesforce now has orders of magnitude more data than Marketo does for us. Our Marketo renewal is right after SaaStr AI Annual and we’ll probably just kill it. We wouldn’t even move to HubSpot at this point — we need something agentic-native that pulls directly from Salesforce and creates campaigns autonomously.

* Agent ROI is dead simple to measure. Attribute closed-won revenue directly to each agent. Unlike traditional marketing attribution, there’s rarely even a multi-touch problem at the agent level. Which agent touched which lead? Did it close? That’s it. The original goal was even simpler: replace a six-figure hire with a five-figure agent and maintain the same revenue. Even if the agents performed a little worse than expected, they’d still be ROI-positive on that math.

* Too much data is a real problem. Once you have 20+ agents collecting data, you actually need to limit what they ingest. Qualified knows what pages a prospect visited on SaaStr.com. That’s useful context. But does the agent need to know that Bobby used to work at this company and came to SaaStr five years ago but has since left? No. That’s irrelevant context that degrades output. More data doesn’t mean better — it often means worse.

* Agentic marketing is the massive gap in the market right now. Sales agents are way ahead. There is still no agent that can do full email marketing at scale — segment your base, create campaigns from historical data, and send to the right contacts at the right time. Not just write copy. Actually execute. Whoever builds the agentic HubSpot/Marketo replacement will own a real piece of the next era of B2B.

The Big Takeaways

* If you’re deploying your first few agents: Know that the complexity scales non-linearly. Going from 2 to 20 agents isn’t 10x harder — but it’s a fundamentally different management challenge.

* If you’re already managing multiple agents: You need daily check-ins with every agent, a planned blackout period for each new addition, and at minimum two humans who understand the full stack. Today. Not next quarter.

* If you’re building B2B + AI agent products: The ROI bar is insanely high right now. If your agent can generate six-figure meetings in week one or replace multiple humans reliably, you will have demand out the door. Monaco is booked for two months of demos including weekends. Basis raised at a billion-dollar valuation and can’t serve all their demand. If people aren’t lining up to buy your agent, your agent isn’t good enough. Don’t delude yourself.

The agents aren’t going away. They’re just getting harder to manage. And that, weirdly, is the best sign that they’re actually working.

Want to learn how to WIN in the AI Era in B2B? Join 10,000 of us atSaaStr AI 2026 May 12-14. We’ll give you the playbooks to win in AI + B2B in 2026.

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 did a deep dive on 20VC x SaaStr this week with Mike Cannon-Brookes, co-founder and CEO of Atlassian. Atlassian just put up an incredible quarter of accelerating growth (23% at $6.4B ARR, with RPO growing to 44%). And yet the markets aren’t showing anyone much love. Mike was honest and reflective on just what’s happening to B2B and SaaS in the Age of AI.

There’s so much noise about “software is dead” and “agents replace everything” that founders are losing the plot. Mike’s running a $6B+ revenue business that’s accelerating — 26% cloud growth, 44% RPO growth — in the middle of the supposed SaaS apocalypse.

So let’s break down what Mike actually said, and what it means for the rest of us.

1. “Software Is Dead” Is a Stupid Statement. Full Stop.

Mike didn’t mince words here. The idea that software as a category is going away is, in his words, “ludicrous.”

His argument is simple and hard to refute: businesses have always bought pre-built technology solutions. They didn’t write everything in assembly language before, and they’re not going to build everything from scratch with LLMs now.

Will every B2B company make it through the next 5–10 years? Absolutely not. Will many of them grow and prosper? Absolutely. Is that any different from the last 10 years? No.

Mike pulled up Atlassian’s old competitive docs from 2005, 2010, 2015. A huge chunk of those companies don’t exist anymore — merged, acquired, or gone. That’s just how the technology industry works. AI doesn’t change the fundamental pattern. It just accelerates it.

The takeaway for founders: stop listening to the “SaaS is dead” crowd. The real question is whether your company is good enough to win in the next era.

2. “You Just Have to Be Good.” That’s the Whole Strategy.

This was my favorite line from the conversation and I think it deserves to be tattooed on every B2B founder’s forehead.

When asked how Atlassian thinks about competing with Anthropic for CIO budgets, Mike’s answer was deceptively simple: “We have to be good.”

Not “we have to pivot to AI.” Not “we need to become an agent platform.” Just: we have to be good. We have to deliver more value to our customers than the alternatives.

Atlassian has 10,000 people in R&D. They’re using Claude Code internally. Their inference costs are going down while they ship more AI features. Some features are 1,000x cheaper to run than when they first launched them. Their gross margins have improved over the last six or seven quarters while deploying more AI.

That’s what “being good” looks like in practice. It’s not a platitude. It’s an execution standard.

3. The Revenue Stacking Problem Is Real — and Most People Don’t Understand It

Anthropic projects $149B in ARR by 2029. OpenAI projects $180B. That’s ~$350B between two companies in a $700B global software market.

Mike made a point that almost nobody talks about: the revenue stacking is complicated.

When Atlassian spends money on Anthropic, they actually pay AWS, and then AWS pays Anthropic. When Cursor does a billion in revenue, a big chunk of that is the same billion as Anthropic’s revenue. The individual revenue numbers don’t just add up cleanly.

So when you see these massive projections and panic about where the budget comes from — remember that a significant portion is double-counted across the stack. The actual net new spend enterprises need to allocate is smaller than the headline numbers suggest.

That said, even with stacking, the numbers are enormous. As Rory pointed out: Anthropic becoming $150B and OpenAI becoming $180B is basically saying two new Microsofts showed up in four years. You better believe in TAM expansion, or the math gets really uncomfortable for everybody else.

4. Product & Engineering Is the Island of Stability. Everything Else Is at Risk.

We’ve been saying this at SaaStr and Mike’s experience at Atlassian confirms it: every category outside of engineering and product is at existential risk of shrinking seats.

Workday said it publicly — even they’re seeing headwinds on seats because Fortune 500 companies just aren’t hiring like they used to. The data from Pave shows no category has been more decimated in hiring than customer support.

But engineering? Nobody is cutting their engineering teams. Not yet at least. Even if they are hiring very differently in the Age of AI. We are in a renaissance of software creation. I was at Replit the other day — 300 million in revenue, 300 people, 11 in go-to-market. The rest? Engineers. That’s not a company cutting R&D headcount.

Mike’s framework for understanding this is genuinely useful: think about whether a function is input-constrained or output-constrained.

* Customer support is input-constrained. You have X customers asking Y questions per day. Make the team more efficient and you need fewer people. Legal is similar — you can’t create more legal problems just because your lawyers got faster.

* Engineering is output-constrained. The roadmap is never finished. You can always create more. Make engineers more productive and you just build more software, faster. The headcount doesn’t shrink — the output explodes.

If you’re selling to input-constrained functions, your per-seat revenue is going down. If you’re selling to output-constrained functions, you’re probably fine.

Plan accordingly.

5. The Real Problem with Public SaaS Metrics Is a Composition Problem

One of the sharpest observations in the conversation: the reason public SaaS growth rates look so bad isn’t necessarily because individual companies are failing. It’s a composition problem.

For about 15 years, the median public SaaS growth rate held around 30%. But that wasn’t because every company grew at 30%. It was because when a company slowed below 10%, it got taken private by PE, and a new company IPO’d at 60% growth to replace it at the top of the index.

That cycle broke. No new high-growth SaaS companies have IPO’d in years. PE has hoovered up a ton of mid-tier companies. Big tech is acquiring more aggressively. Companies like Stripe and Figma are staying private much longer because they don’t need to go public.

So what you see in the public market index is this weird survivor set — too big for PE to buy, too small for big tech to acquire, with no new high-growers being added at the top. Of course that average looks bad. But it’s not necessarily a statement about the health of B2B software overall.

The lesson: don’t confuse the declining public SaaS index with the death of SaaS. They’re measuring different things.

6. Being a Public Company Makes You Better — But You Have to Play Offense Too

I loved Mike’s take on the public vs. private company dynamic. His answer wasn’t defensive. It was almost counterintuitive.

“We’re a better company because we’re a public company. We’re better at forecasting. We’re better at planning. We’re better at executing.”

The trap, he says, is when you let financial discipline replace strategy instead of complementing it. You still have to invest aggressively in AI and new products. You still have to compete. You still have to go into new areas — Atlassian just shipped a customer service product. Their service collection business could go public by itself.

The private companies have the luxury of not marginal-costing every decision. Their CMOs don’t have “pesky public market investors” asking about profitability. That creates a real asymmetry. But Mike argues the discipline of public markets is actually a long-term advantage — you just can’t let it make you timid.

7. If You Don’t Enjoy It Anymore, That’s Okay — But Be Honest About It

Mike shared something that really resonated. His co-founder Scott Farquhar retired a year and a half ago after 23 years. Mike starts work at 5 AM every day now. He’s working harder than ever.

And his advice to the SaaS founders he’s been counseling (he called it “SaaS therapy”) was refreshingly direct:

First, accept reality. Stop pontificating about whether AI is going to change things. It already has. Go build something.

Second, ask yourself honestly: do you still enjoy this? Not “do I enjoy every minute of every day” — that’s not realistic. But over a 90-day average, over a year, would you choose this job again today?

If the answer is no, that’s fine. You built something amazing. Hand it off thoughtfully and go do something you enjoy. That’s not weakness — it takes more courage to step away than to hang on collecting a paycheck until the board forces you out.

If the answer is yes, then get to work. We’re going to create our way out of this, not hide in a hole and wait for it to pass.

CEO for 24 Years in the Age of AI

Mike Cannon-Brookes has been building Atlassian for 24 years. He’s seen every “this time it’s different” moment the tech industry has produced. And his message is clear:

B2B software isn’t dead. It’s changing. The companies that will win are the ones that accept reality, invest aggressively in AI, maintain execution discipline, and — critically — are actually good at what they do.

If you’re a founder reading this and feeling anxious about the AI transition, here’s your to-do list:

* Figure out if you’re selling to input-constrained or output-constrained functions

* Stop reading “SaaS is dead” takes and go ship AI features your customers actually want

* Make sure your inference costs are going down, not up

* Ask yourself honestly if you still want to do this

* If yes, start at 5 AM tomorrow

The era of autopilot SaaS growth is over. But if you have 350,000 customers, great distribution, engineers who can build, and a product people love? That’s a pretty good starting point to build a company.

You just have to be good.

Want to learn how to WIN in the AI Era in B2B? Join 10,000 of us at SaaStr AI 2026 May 12-14. We'll give you the playbooks to win in AI + B2B in 2026.

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

High inference costs are OK—if they make your product so viral and so competitive it almost sells itself

Here’s the counterintuitive insight that’s reshaping how the smartest AI founders think about unit economics:

Your inference costs aren’t your gross margin problem. They’re your CAC replacement.

The companies growing fastest right now—Cursor crossing $1B ARR with ~300 employees and no traditional marketing, Lovable hitting $300M ARR with zero paid acquisition—aren’t sweating inference costs. They’re leaning into them. They’re treating compute as their primary growth investment, not their primary margin drag.

This is a fundamental reframe. And if you’re still optimizing for gross margin while your AI-native competitors are optimizing for virality, you’re playing the wrong game.

The Math That Traditional B2B and SaaS Gets Wrong

On a recent 20VC x SaaStr episode, we discussed Anthropic’s inference costs coming in 23% higher than expected. My immediate reaction was pessimistic for mid-market B2B SaaS:

“I worry this is the final nail in the coffin. You did everything right—got profitable, built an agent—and now you just can’t afford the inference to compete.”

Here’s the scenario: You’re a $50M ARR B2B company. You built the agent your board demanded. Your agent costs $2.50 per interaction. You need 50 million interactions to stay competitive. That’s $125 million in inference costs on $50M in revenue.

Game over, right?

Not necessarily. The question isn’t whether you can afford the inference. It’s whether the inference makes your product so good that sales and marketing become irrelevant.

The Cursor Playbook: Inference as Distribution

Cursor crossed $1B ARR by late 2025—roughly 24 months from launch—with about 300 employees and minimal traditional marketing. They went from $100M ARR in January 2025 to $500M by June to $1B+ by November. The fastest SaaS growth curve ever recorded.

How? They spent aggressively on inference to create what Andrej Karpathy called the “vibe coding” experience—the moment when developers forget they’re writing code and just describe what they want. That experience is computationally expensive. It requires reasoning tokens, multiple model calls, context management across entire codebases.

Traditional SaaS math would call this margin suicide. But here’s what actually happened:

* The “wow moment” converted instantly. Developers tried Cursor, experienced something magical, and became evangelists within hours.

* User-generated content became their entire marketing funnel. Every tweet about “I built an app in a day with Cursor” was free distribution worth thousands in CAC.

* The viral loop compounded. Engineers at OpenAI, Midjourney, Shopify, and Instacart started spreading it organically. No sales team required.

* Conversion was frictionless. $20/month is an impulse buy when the product makes you demonstrably faster.

The inference spend wasn’t a cost center. It was the marketing budget. It just showed up on a different line item.

Lovable’s Rocket Ship to $300M

Lovable hit $300M ARR in January 2026—roughly 14 months after launch—with fewer than 200 employees and zero paid acquisition. That’s still $1.5M+ revenue per employee, nearly 8x the industry benchmark.

Their secret? They engineered virality into the product itself. When users build apps with Lovable, the outputs are shareable. The AI-generated code is good enough that users want to show it off. Every app becomes a piece of marketing collateral.

The underlying inference cost to generate these apps is significant. But look at what they avoided:

* No enterprise sales team (zero)

* No paid acquisition (zero)

* No SDRs cold-calling (zero)

* No expensive conference sponsorships (zero)

The inference is the go-to-market motion. The product is the marketing.

You Can’t Have It Both Ways

Here’s the brutal math that too many founders are ignoring:

You can’t have high inference costs AND high sales & marketing costs. At least not for long. It has to come from somewhere.

Traditional SaaS could absorb 40-50% S&M spend because gross margins were 80%+. There was room. The unit economics worked.

But when your gross margin drops to 50-60% because of inference costs, that room disappears. You’re now choosing between two paths:

Path A: Inference-First (Cursor, Lovable)

* Gross margin: 50-60%

* S&M: <10%

* Growth driver: Product virality

* Requires: Magical product that sells itself

Path B: Sales-First (Traditional Enterprise SaaS)

* Gross margin: 75-80%

* S&M: 40-50%

* Growth driver: Sales efficiency

* Requires: Lower inference costs, less AI magic

What you cannot do is run 50% gross margins AND 40% S&M. That’s negative operating margin before you pay a single engineer. That’s burning cash with no path to profitability. That’s a company that dies.

The trap I see founders falling into: they build an AI product with significant inference costs, then layer a traditional enterprise sales motion on top of it because “that’s how you sell to enterprises.” Now they’re paying for compute AND paying for a sales team AND paying for marketing—and wondering why their burn rate is out of control.

Pick a lane. Either your inference spend IS your sales & marketing (because the product is so good it creates its own distribution), or you optimize inference costs aggressively and invest in traditional go-to-market.

The companies trying to do both are the ones that will run out of runway.

The Third Path: Build Something So Valuable They’ll Pay $50-100K+ (More)

There’s another way to solve the inference cost problem that doesn’t require virality or zero sales spend.

Build an AI agent so valuable that businesses will gladly pay $50,000 to $100,000+ for it. Even small businesses.

This isn’t fantasy. I talked to a sales leader at an agentic AI company recently. I asked him how many figures are in his mid-market deals. I was thinking $20K to $50K, maybe $75K on the high end.

He said they’re all seven figures.

Seven figures. For mid-market.

When you price based on labor replacement instead of software seats, the math completely changes. You’re not competing with Salesforce’s $150/seat. You’re competing with a $120K SDR salary, a $180K AE salary, a $95K customer support rep.

The current market rate for agentic GTM tools: $100K+ per agent as a starting point. Some companies are paying $50-70K just to get started, plus another $25K for a forward-deployed engineer to set things up.

And here’s the thing: smaller businesses will pay this too if the ROI is obvious enough.

Think about it from the buyer’s perspective:

* Human SDR loaded cost: $100K-$130K annually

* AI SDR: $10K-$15K for basic, or $50K-$100K for enterprise-grade with support

* Even at $100K, that’s still cheaper than a human—and the AI never has a bad day, never forgets the product, never needs ramp time

At $50K-$100K per agent, your inference costs become a rounding error. You’re not optimizing tokens. You’re delivering ROI measured in weeks, not quarters.

This is the premium escape hatch from the inference cost trap: Don’t try to win on efficiency. Win on value so overwhelming that price becomes irrelevant.

The Inference Squeeze on Mid-Market B2B

Here’s where it gets brutal for traditional SaaS.

If you’re a $50-100M ARR company that spent years building a sales org, optimizing your marketing funnel, and perfecting your enterprise motion—you now face a new type of competitor. One that has no sales team to pay, treats inference as their primary distribution channel, and can invest every marginal dollar into making the product more magical.

You can’t out-spend them on ads because they don’t run ads. You can’t out-execute them on sales because they don’t have salespeople to compete against. Your moat was go-to-market efficiency. Their moat is product virality powered by unlimited compute.

The existential question becomes: Can you afford NOT to spend on inference?

The Framework: Inference as CAC

Here’s how to think about this:

Traditional SaaS Model:

* Gross margin: 75-80%

* S&M spend: 30-50% of revenue

* CAC payback: 12-24 months

* Growth driver: Sales & marketing efficiency

AI-Native Model:

* Gross margin: 40-60% (temporarily)

* S&M spend: <10% of revenue

* CAC payback: Near-instant (product is the acquisition)

* Growth driver: Product virality via inference quality

The question isn’t “What’s your gross margin?” The question is “What’s your blended customer acquisition cost when you include inference spend that drives organic growth?”

If spending an extra $1M on inference generates 10x that in organic revenue—because your product creates its own distribution—you should spend that million dollars every time. Even if it temporarily crushes your gross margin.

The New Competitive Moat

The winners in the AI era won’t be the companies with the best gross margins. They’ll be the companies that figured out how to turn inference into distribution.

This requires:

* A product with a shareable “wow moment.” Not just good—magical. The kind of thing users can’t help but tweet about.

* Viral mechanics built into the core experience. Every output should be a potential piece of marketing content.

* Pricing that makes conversion frictionless. $20-50/month that’s an immediate no-brainer when the value is obvious.

* The financial runway to absorb inference costs while the viral loop compounds. This is where well-capitalized AI startups have an edge—Open Evidence, Harvey, Lovable all raised massive rounds specifically to fund this strategy.

The Hard Truth for Incumbents

Here’s what I worry about for traditional B2B:

You did everything your board asked. You got profitable. You built an agent. It works. Your customers like it.

But your agent costs $2.50 per interaction. You need 50 million interactions this year. That’s $125M in inference costs.

Meanwhile, your AI-native competitor raised $300M specifically to fund inference spend. They have no sales team. Their product is so good it creates its own distribution. They’re growing 300% year-over-year while your growth is decelerating.

If you walked into your board meeting and said “Good news guys, inference costs are going down 30% this year because our IT team’s really good at managing costs,” I would throw my mouse at the monitor.

The only way out is to build an agent so good you can charge $10-20K per month because it replaces 20 people with ROI measured in weeks—not a sales pitch claiming it’s that good, but literally that good.

Or: accept that inference isn’t your enemy. It’s the new battlefield. And the companies willing to spend aggressively on compute to create magical, viral products will win—even if their gross margins look terrible by traditional SaaS standards.

Inference Is The New Sales & Marketing. At Least, a Lot of It.

The AI companies winning right now have flipped the script entirely:

* Cursor: $1B+ ARR, minimal marketing, inference is the product that sells itself

* Lovable: $300M ARR in ~14 months, zero paid acquisition, virality is the distribution

They’re not worried about inference costs crushing their margins. They’re worried about inference costs being too low to create competitive differentiation.

That’s the paradigm shift. Inference isn’t the new COGS. Inference is the new S&M.

Spend accordingly.

Three Takeaways for Founders:

* Reframe inference as acquisition cost, not gross margin drag. If spending $1 on inference generates $10 in organic revenue through virality, that’s a 10x return on your “marketing” spend.

* Build shareable “wow moments” first, optimize margins later. The companies growing fastest right now prioritized magical user experiences over unit economics—knowing the economics would follow.

* If you can’t afford to compete on inference, find a different wedge. Either charge enough that inference costs don’t matter ($10K+/month with clear ROI), or accept that you’re competing against companies with infinite compute budgets and no sales teams to support.

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