Manual Work is a Bug

This week we have a special episode - Live from the ITC 2026 expo floor. Mav Founder + CEO Matthew Black and Head of Brand Hillary Black grab ten minutes in the official podcast studio between booth shifts to report on what actually changed in insurance this year.

Every vendor is still pitching their version of AI for insurance. But what’s changed this year is the agency owners.

On the table: why GPT helped the first five minutes of every sales conversation finally disappear for companies in AI, the commercial agency strategy for quickly moving into personal lines without hiring, evidence that Mav has the coolest booth in the tradeshow circuit, and Matthew's prediction for where the insurance product itself is headed in the next year.

Chapters

00:00 — Intro

00:16 — Live from the ITC 2026 expo floor

01:17 — The coolest booth on the circuit

02:45 — "X for AI," three years later

03:40 — The agency owner has changed

05:07 — The old guys in the back

06:19 — From "I don't text" to "how do I grow?"

07:04 — Commercial agencies moving into personal lines

08:28 — A prediction for next year

09:37 — Why doesn't insurance just come with the car?

10:05 — Insurance like shipping on Amazon

10:44 — Close: see you at ITC 2027

Hosts

Matthew Black — Founder + CEO, Mav

Hillary Black — Head of Brand and Conversation Design, Mav

Links and Resources

ITC Vegas (InsureTech Connect), the world's largest insurance innovation event

Toyota Insurance, the automaker-owned agency selling coverage alongside the car

Follow the show

Manual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies grow their book with AI: hiremav.com.

Spotify

Amazon Music

This week we're walking through the entire process of rebuilding and redesigning the ⁠hiremav.com⁠ website using Claude Code, the Figma MCP, and two humans. ⁠Mav⁠ Founder + CEO ⁠Matthew Black⁠ and Head of Brand ⁠Hillary Black⁠ walk through the exact process from ideas to setup, iterating, and launching, plus where humans stepped in to make it not look like every other AI-generated website. We think it's looks pretty cool.

Hillary, a non-engineer, was able to build the website of her dreams the exact way she wanted without waiting weeks to fit it into a sprint cycle, and the Claude Code "other tab" even convinced Matthew to use the Claude App instead of terminal for his work.

On the table: why buying a hundred old insurance agencies and trying to AI-enable them doesn't work, the one rule that kept the site from looking AI generated, the tells that give an AI website away, a dev setup that went from four hours to thirty minutes, whether a website should even be the same for every visitor, and an Easter egg worth a free year of Mav.

🥚 There's an Easter egg hidden on the new ⁠hiremav.com⁠. Find it, screenshot it, and DM Matthew on LinkedIn. First one gets a free year of Mav.

Chapters

• ⁠00:00⁠ — Intro
• ⁠00:15⁠ — Cold open: we're live
• ⁠01:01⁠ — The TV nobody could find
• ⁠04:21⁠ — Too many humans in the way
• ⁠04:49⁠ — Why the AI roll-up doesn't work
• ⁠06:29⁠ — Buying the business for parts
• ⁠07:11⁠ — Today: replacing ⁠hiremav.com⁠
• ⁠08:08⁠ — Twenty new pages, two people
• ⁠09:35⁠ — No staging server, no CMS
• ⁠10:45⁠ — AI started, human finished
• ⁠11:28⁠ — "I'm not a designer"
• ⁠13:04⁠ — Was the wireframe holding it back?
• ⁠14:17⁠ — The tells of an AI website
• ⁠15:43⁠ — A four-hour setup in thirty minutes
• ⁠16:57⁠ — Claude Code in the app
• ⁠18:28⁠ — Every page, tracked
• ⁠20:42⁠ — What even is a website anymore?
• ⁠22:01⁠ — The new AI use case: create constraints
• ⁠25:08⁠ — Using humans where it counts
• ⁠26:01⁠ — One button and it's live
• ⁠27:06⁠ — The Easter egg

Hosts

• ⁠Matthew Black⁠ — Founder + CEO, Mav
• ⁠Hillary Black⁠ — Head of Brand and Conversation Design, Mav

Links and Resources

• The new ⁠hiremav.com⁠
• ⁠Claude Code⁠, what Matthew pointed at the README and what built the pages
• ⁠Claude Code and Figma⁠: setting up the MCP server, the connection that let Claude read the actual design file
• ⁠Guide to the Figma MCP server⁠
• ⁠Model Context Protocol⁠, the open standard behind the Figma connection
• ⁠ngrok⁠, how Hillary's local build gets shared with anyone
• ⁠Docker⁠, the self-contained dev environments
• ⁠micro1⁠, the AI training-data company Matthew mentions

Follow the show

Manual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies grow their book with AI: ⁠hiremav.com⁠.

⁠Spotify⁠

⁠Amazon Music⁠

After two weeks of AI getting it wrong, let's talk about where AI gets it right. Founder + CEO Matthew Black and Head of Brand Hillary Black go looking for the places AI genuinely outperforms a human, and what those use cases have in common.It starts with a car that drives itself better every month, and an insurance company that charges you less the more you don't do your own driving.We talk: why Tesla's lead in autonomous driving may be impossible to catch, the difference between selling tokens and selling outcomes, why everyone's AI flyers look the same, a data broker's offer that stopped Matthew cold, why CRMs are graveyards, and his MCP setup that replaced a personal trainer.

Chapters

00:00 — Intro

00:15 — Cold open: ten episodes in

01:10 — Starlink at cruising altitude

03:40 — The chairs: a final update

05:04 — When nobody can override anything

07:01 — Tesla self-driving, six years later

08:24 — Insurance that pays you to let the AI drive

09:46 — Why Tesla wins: the data nobody else has

10:54 — Sell outcomes, not tokens

12:22 — Why all the flyers look identical

13:29 — Data is being brokered like oil

14:04 — You can't be an AI wrapper anymore

16:37 — Tesla isn't a car company

17:57 — The company that can't use its own data

19:53 — Choosing tools by what they connect to

21:03 — Why do you even need a CRM?

21:38 — No steering wheel, no pedals

22:48 — What agencies can do at a smaller scale

24:35 — The fitness stack that replaced a trainer

27:16 — The bridge is hard

29:40 — Why Mav went all the way into insurance

Hosts

Matthew Black — Founder + CEO, Mav

Hillary Black — Head of Brand and Conversation Design, Mav

Links and Resources

Tesla Insurance: how the Full Self-Driving discount works

Lemonade's Autonomous Car insurance, roughly 50% off every mile driven on FSD

Tesla, Uber and Waymo all get the OK to run thousands of robotaxis in Nevada

Model Context Protocol, the open standard behind connecting your tools to Claude

athletedata, a single MCP server that connects WHOOP, Hevy, MyFitnessPal and 30 more fitness apps to Claude

Rippling, the not-a-sponsor

Follow the show

Manual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode.

Mav helps insurance agencies grow their book with AI.

We're an AI company, but we won't sell you AI for customer service (yet). Here's why we're so bothered by it. Mav Founder + CEO Matthew Black and Head of Brand Hillary Black get into why the thing everyone is automating first for some reason, is also the thing that breaks most often, and potentially costs your business the most.

Hillary is bothered. Three sets of chairs, a retailer with no phone number, and a bot that can talk all day without doing a single thing about it.

On the table: the one company Matthew thinks actually gets this right, the two conditions that separate a support experience that works for customers, from one that wastes an hour of your life, the unit economics nobody considers before they switch off their phone lines, what happens when your agent starts talking directly to the store's agent, and a mall built entirely for AI.

Chapters

• 00:34 — Cold open: three sets of chairs and no phone number
• 02:35 — Why some AI support is good and most of it isn't
• 03:21 — Why Mav stayed out of customer service on purpose
• 04:10 — Sales is a formula. Support is one-out-of-one.
• 06:34 — 27 entry points vs. 3, and the unit economics nobody runs
• 08:43 — Cheaper per conversation, more expensive per customer
• 09:01 — The one that actually worked
• 10:50 — The two conditions
• 12:06 — Onboarding a rep with no access to inventory
• 12:30 — Saying the right thing doesn't matter if it can't do the job
• 14:30 — Muse, and what happens when agents talk to agents
• 16:24 — Meta bought the agent forum
• 18:13 — MoltMall
• 20:01 — Everyone is doing the best they can
• 22:19 — The fold phone detour
• 23:52 — Go win somewhere easier first
• 24:25 — The worst possible place to put AI in insurance
• 26:16 — Reinvest the savings at the bottom of the funnel too
• 27:37 — Why Chewy gets away with it

Hosts

• Matthew Black — Founder + CEO, Mav
• Hillary Black — Head of Brand and Conversation Design, Mav

Links and Resources

• Meta's Muse — the personal agent launched September 8, described as OpenClaw for everyone
• OpenClaw — the open-source personal agent Muse is answering
• Meta acquired Moltbook, the social network built for AI agents
• Model Context Protocol — the open standard behind agent-to-agent connection
• Stripe Support — the implementation Matthew holds up as the bar
• Chewy — Matthew's example of stellar customer service

Follow the show

Manual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies grow their book with AI: hiremav.com.

Your customer is about to ask an AI whether they're overpaying you. Mav Founder + CEO Matthew Black and Head of Brand Hillary Black get into what happens when people end up shopping for insurance when all they did was ask AI to fix their budget.On the table: the prop plane Matthew almost bought, why comparison sites can't help you with this, what Bitcoin taught him about how consumers trust, how a small agencies compete with big budgets, how to get your agency recommended by ChatGPT, what insurance will look like a decade from now, and a million dollar idea to win over your clients he gives away at the end.Hosts- Matthew Black — Founder + CEO, Mav- Hillary Black — Head of Brand and Conversation Design, MavLinks and Resources- The Insurance Buzz with Michael Weaver and Courtney Weaver — the agency-side version of this conversation, with Matthew- Insurify's ChatGPT app — real-time personalized auto quotes inside the chat, the execution Matthew points to- How Insurify's app started, back in February — rate estimates and side-by-side comparisons from a database of 196M quotes- The insurer ChatGPT apps directory — who else is already quoting inside an assistant- Why quoting inside the assistant is the whole game, including Liberty Mutual making the first move among US carriers- Model Context Protocol — the open standard that makes any of this connectable- Tesla Insurance — insurance as a line item in the purchase wizardFollow the showManual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies grow their book with AI: hiremav.com.

If you found AI for the first time today, where would you start? Mav Founder + CEO Matthew Black and Head of Brand Hillary Black run the thought experiment: deserted island, you reenter civilization, and every tool that exists right now is sitting there waiting. What do you reach for first, and what should you never touch at all?On the table: the two questions that tell you where to start, why the thing everyone reaches for first is the wrong thing, Matthew's AI ick and the one skill he wishes he didn't have, why human outreach is quietly becoming an unfair advantage again, the engineering process Mav rebuilt after choking on its own AI code, an $18,000 Mac you can lease for the price of a Claude subscription, and a poker chip that came out looking like a fish.One of these is where to start. One of them is where to never use it.Hosts- Matthew Black — Founder + CEO, Mav- Hillary Black — Head of Brand and Conversation Design, MavWhere to learn moreRecord your screen, get a skill — the thing Matthew describes at 17:14- Record a Skill, in the Claude desktop app's Cowork "+" menu — you do the task once, narrate it, and Claude writes the skill- Codex Record & Replay — OpenAI's version, shipped a few weeks earlier; this walkthrough shows what the output actually looks likeHow to build skills, the thing Hillary's whole brand system runs on- Equipping agents for the real world with Agent Skills — Anthropic's engineering post, the best single explainer- Agent Skills docs — what a SKILL.md actually is and where to put it- Introduction to Agent Skills — free Anthropic course, start-to-finish- anthropics/skills on GitHub — real working skills to copy from, including brand and document ones- agentskills.io — the open standard, so a skill you write isn't locked to one toolUsing AI for something other than making content- Claude Cowork — the "give it access to your actual files and tools" product- Model Context Protocol — the open standard behind every connector, if you want to understand why access matters- Claude Code — what you'd use to point a local model at your own machine

Local Model Talks

Llama — Meta's open-weight models, free to download and run yourself- Apple's new Mac Studio- AppleInsider on the M5 Max and M5 Ultra refreshFollow the showManual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies grow their book with AI: hiremav.com.

Get hour tinfoil hats out, this week we're talking about 10 AI conspiracy theories. Mav Founder + CEO Matthew Black and Head of Brand Hillary Black go through the internet's darkest, strangest and weirdest AI theories. Real, or fever dream?On the list: Google buying a dead airline's data. Shredding rare books for the sake of AI. The dead internet theory. An agent-only forum where the bots allegedly invented their own language in thirty-six hours. Labs poisoning their own models to keep them to themselves. An AI sentience cover-up. Simulation theory. People marrying their AI. And an ad campaign asking America to cool the data centers with pee.Some of these have press releases. Some of them should have stayed on the internet.

Links and Resources

• Anthropic on detecting and preventing distillation attacks
• Google bought Spirit Airlines' data for $10M — 100M emails, 500M Teams messages, 30M lines of code
• Why that data is worth it — teaching agents white-collar work needs private records, not public code
• 404 Media hid a tracker in a rare book and followed it to an Amazon AI facility in Las Vegas
• Amazon, which started as a bookstore, is destroying rare books to train AI
• Cloudflare on bot traffic passing human traffic
• Moltbook — the Reddit-style social network only AI agents can post on
• Thinking Machines Lab — Mira Murati's company, the ex-OpenAI CTO
• Love at First Prompt by Bridget Todd — on relationships between people and AI
• Liquid Death x Garage Beer, "We Want Your Pee"
• Can we actually cool data centers with pee?

Follow the showManual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies grow their book with AI: hiremav.com

Is AI actually locked up for the rich? Mav Founder + CEO Matthew Black and Head of Brand Hillary Black open on Mark Zuckerberg's new manifesto, "The Future is for Everyone" — released the day Meta open-sourced Muse Glimmer, a 30B model small enough to run on your own machine. Matthew's read: the scary question was never who controls AI, it's who doesn't have it yet. And the public library is the counterargument nobody makes (but we're in huge support of, seriously, give them your money).Hillary sees it differently. Meta's own problematic algorithms are exactly why nobody takes "this is for everyone" at face value, and today's free tiers are nothing like being mailed a CD of the internet. Plus, a Home Assistant demo that builds its own A/V dashboard on a $500 laptop, and Hillary's tinfoil-hat theory on OpenAI's disastrous influencer brand trip — everyone called it a botched, out of touch product launch, but she thinks the discourse missed the point entirely.If you're trying to figure out whether AI is something you own or something you rent, this one's for you.

Links from the Episode

- Mark Zuckerberg's Manifesto, "The Future is for Everyone: The Path to a Positive AI Future" - Muse Glimmer — Meta's 30B open-weight model, small enough to run on a single consumer GPU instead of a data center- Muse Glimmer weights on Hugging Face — free to download under Apache 2.0, commercial use included- Home Assistant — the open-source platform Matthew runs at home, and the one Glimmer's own demo builds a dashboard for - MacBook Neo — Apple's $599 entry level laptop- Peter Diamandis & Steven Kotler, We Are as Gods: A Survival Guide for the Age of Abundance - Ari Kuschnir — the artist behind the AI Sam Altman "come live in the data center" films, and the creative process Matthew wants everyone to see- Influencers draw backlash for attending OpenAI's first luxury trip — TechCrunch's coverage of the trip- Apple names John Ternus CEO- How libraries are funded- Support the Las Vegas-Clark County Library District — the actual library system of our cityFollow the showManual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies democratize growth with AI: hiremav.com.

How many Claude sessions is too many?  Mav Founder + CEO Matthew Black and Head of Brand Hillary Black open with Matthew running six at once — then dig into where AI agents actually belong in a company.

The jumping-off point: Jack Dorsey and Block's new open-source tool Buzz, where agents are first-class citizens (their own identity, permissions, and audit trail) instead of bolted-on Slack bots. Matthew's real interest isn't the chat — it's shared compute and a shared context window, so the whole team can tap one secure model instead of everyone spinning up their own and quietly reinventing the wheel.From there: why artifacts (Notion, HubSpot, Stripe) tell you what's true but not how or by whom a decision was made; the viral "I replaced my whole marketing team with agents" claim (spoiler: we're calling mostly fake); and a genuinely useful framework for putting "agents on the accountability chart" — if a seat needs net-new creative decisions, it stays human; if it's factual, marker-in-time work, an agent can do it faster; if it's both, augment.

Plus the Take-Two/GTA case for why derivative AI output can't make a hit, Bezos on getting paid for a few high-quality decisions, why AI-generated art and copy-paste LinkedIn comments miss the point — and whether Matthew could throw a ceremonial first pitch over the plate.If you're figuring out where agents fit on your team without turning your company into an echo chamber, this one's for you.Timestamps- 00:00 — Intro- 00:21 — Cold open: how many Claude tabs are open right now?- 01:15 — The premise: agents on the org chart (+ an Angels-in-the-Outfield detour)- 03:40 — Jack Dorsey's "Buzz": agents as first-class citizens- 11:00 — Shared compute & the secure-Claude problem- 13:20 — Why teams reinvent the wheel: context vs. artifacts- 16:25 — Agents on the accountability chart & the viral "I replaced my team" claim- 17:24 — The framework: net-new-creative stays human, factual → agent, both → augment- 19:17 — Take-Two, GTA & derivative IP vs. net-new creativity- 23:45 — The human edge: why AI content misses, and Bezos on high-quality decisions- 27:10 — Close: the most human thing (a ceremonial first pitch)Hosts- Matthew Black — Founder + CEO, Mav - Hillary Black — Head of Brand and Conversation Design, MavLinks and Resources- Buzz by Block — the open-source, agent-native Slack/GitHub alternative where agents are first-class citizens- Linear — the issue tracker whose agent workflow Matthew compares Buzz to- Claude — one of the models Mav runs on internally, including Claude for Slack- OpenClaw — the open-source personal-agent harness behind Matthew's "~$100k a day" migration example- Strauss Zelnick, Take-Two CEO, on why AI output is "derivative" and can't make a hit like GTAFollow the showManual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies democratize growth with AI: hiremav.com.

What does it actually look like to run a company *through* AI? In this episode of Manual Work is a Bug, Mav Founder + CEO Matthew Black and Head of Brand Hillary Black get specific on how they use AI at work, after building AI for nearly a decade. Starting from Jack Dorsey's essay "From Hierarchy to Intelligence," Matthew breaks down the "world model" — a real-time, company-wide brain that flattens the org chart so everyone works on the *edge* of the model instead of climbing a hierarchy — and uses a Boeing-vs-Airbus analogy to explain why he no longer does anything directly: every move gets routed through the model first.Then he shares his actual setup: a decade of getting-things-done manually in Things, now rebuilt into a Granola → Todoist pipeline over MCP that turns his to-do list into an orchestrator for AI. Along the way — why good models are killing the busywork of "clean data" and CRMs, the thing he's going deepest on next (evals and "loop engineering"), the simplest way for any business owner to start (record your screen and ask the model what to take off your plate), Hillary's own experiment analyzing a year of LinkedIn posts, and closing thoughts on Bezos's 80% rule, getting buy-in, and why Claude vs. ChatGPT already shows up in how far people can go. Plus a pickle-flavored taste test nobody asked for.If you build with AI, brand for it, or run a business you're trying to operate with it, this one's for you.Timestamps- 00:00 — Intro- 00:17 — Cold open: the great energy-drink debate- 02:35 — Jack Dorsey's "From Hierarchy to Intelligence" & the world model- 09:18 — The Boeing-vs-Airbus theory of running a company- 12:41 — Matthew's real AI workflow: Things, Todoist, Granola & the to-do list as orchestrator- 22:53 — Why AI kills the busywork of clean data & CRMs- 26:01 — What's next: evals & "loop engineering"- 30:28 — Where to start: record your screen, ask the model (+ Hillary's LinkedIn deep-dive)- 34:17 — Bezos's 80% rule, "find a problem," and Lego-after-Lego- 37:40 — Buy-in, living on the edge & Claude vs. ChatGPT- 40:09 — Mav lore: the pickle-party roulette Hosts - Matthew Black — Founder + CEO, Mav- Hillary Black — Head of Brand and Conversation Design, Mav Links and Resources - Jack Dorsey, "From Hierarchy to Intelligence" (Block) - Things — the GTD app Matthew used for a decade- Todoist — the MCP-friendly to-do app he switched to- Granola — AI meeting notes with an MCP server- Model Context Protocol (MCP) — the open standard that ties tools together- Claude Follow the showManual Work is a Bug is Mav's podcast, now on video. Watch on YouTube and follow wherever you listen, so you don't wait another six years for the next episode. Mav helps insurance agencies democratize growth with AI: hiremav.com.