I hate to admit this, but my least favorite part of being a product manager is doing customer discovery. I know I know. This is bad, but it’s just the truth. I just want to build and test. To me, customer discovery is a timely and expensive part of building, despite it being important to building the right thing.

Lots of user interviews. Surveys people lie on. Landing page tests. Guessing the gap between what people want and what they say they want.

Netflix famously skipped all of that with House of Cards. No pilot or focus groups. They greenlit a $100M show off viewing data. Millions of real behaviors across shows that showed them what people wanted, while the studios kept guessing.

Now, Anthropic has handed all of us that same advantage, and basically nobody has noticed.

They published a study called How people ask Claude for personal guidance.

It can read as a safety paper. And that’s just because it is one. But if you look at it the way Netflix looked at their own usage data, it’s also the results of 1 million user interviews categorized for you, quantified, and 100% free.

As I read it, it’s honestly one of the most actionable product-market fit signals ever released by an AI lab. It’s the 2026 equivalent of unbundling Craigslist or Reddit, now with consumer AI from the big labs. Fun times guys! 💸

In more detail, what Anthropic did was sample 1 million Claude conversations, filtered down to 639,000 unique users, and found that roughly 6% were people asking for personal guidance. Real should-I type questions about their health, careers, relationships, and money.

What that means for us is ~38,000 conversations of pure demand signal. People telling an AI exactly what they need help with—and often, per the study itself, because they could not access or afford the professional version of the answer. A therapist is about $150–300 a session. A lawyer around $300 an hour.

And it’s not just Claude, OpenAI is opening up this data too and published their own usage paper. Anthropic also runs the ongoing Economic Index.

Guys, the labs are now telling us what people actually do with AI, on a rolling basis.

That’s the 8-ball setup for today’s idea. 🎱🥢

AI labs are publishing your market research for free. So, how can you use it?

Attio is the CRM that turns every customer signal into your next move, so you have the velocity to run any go to market motion you need.

With revenue agents that research accounts, draft relevant follow-ups, surface what buyers are saying, and move work forward, Attio runs alongside your team so you can focus on the things that matter most.

🥷 Use the study as a demand map and the labs’ careers pages as a threat map. Then build for a sub-topic where a) stakes are high, b) the standard model fails, and c) no lab is competing for the consumer vertical.

Today’s idea is a method play, and you can re-run it every time a lab drops a usage report. We’ll break down Anthropic’s study today as the example.

The study also spells out how Anthropic wants Claude to behave in these moments…

Which is a lovely description of a premium service most people can’t afford. That gap—between the friend people need and the professional they can’t pay for—is where we can capture value.

Because the accessibility-to-a-service argument is the moat. The data shows us that when people are chatting to their AI, the help they’re looking for is because they can’t afford the human alternative like a real therapist, lawyer, advisor, or coach at hundreds of dollars an hour.

For anyone who can deliver 80% of that professional’s value at 5% of the cost in something beyond just a wrapper…the demand (and payback) is enormous.

And we already have proof that this data converts into venture-scale outcomes. Cal AI which we’ve chatted about here before—that very mediocre AI calorie tracker app— solved one sub-topic from one domain in Anthropic’s taxonomy: “calories and macros for body composition.”

They grew to $40 million in revenue, were acquired by MyFitnessPal, and crossed $50 million in ARR with 7 employees and $0 raised. And health and wellness is just one of nine domains Anthropic is handing to us.

Each one has multiple sub-topics which is where the fun starts to happen.

Cal AI was effectively a democratized dietician at app-subscription pricing—i.e an example of 80% of the value at 5% of the cost.

Sequoia believes the next great companies will sell the completed work directly to the end buyer, rather than tools to professionals. For every dollar spent on software, six go to services, and AI can now capture the services side.

These personal guidance domains are that services budget.

And what’s so great is that this study is just half the useful signal.

Anthropic’s job postings expose the competitive part of the story that isn’t being mapped—the story that says they are building vertical AI products in four big domains: healthcare, financial services, legal, and life sciences.

But, every one of these build outs is going after the enterprise customers, not consumers.

In other words, there’s a precise line between where founders face platform risk and where the consumer layer is wide open to go play.

Let’s break it down, but first a very quick word from Granola before we continue.

Just because I don’t like doing interviews doesn’t mean I don’t do them. And Granola—my no-bot AI notetaker—makes my life so much easier when I do go and talk to people.

It records hands-free in the background straight from device audio skipping all that annoying bot-joining shenanigans. It means I get to give the person my full attention instead of typing, and the moment we hang up I have great notes anyway.

Every interview lands in one repo and is a searchable archive. I also use Granola’s Chat to summarize across all of them—like “What did <insert my hard to work with customer’s name here> keep complaining about this month? 🤦” and get the answer with citations to the exact note. Customer discovery with some of the friction removed.

This conversational data is proper gold, and here’s why.

Surveys and even face to face interviews can be a findings trap. Sure, you can discern and ask good questions, but people still like to tell you what makes them sound smart.

These 38K conversations are the opposite because they are real time spent on real problems with real stakes by real people who, frankly, don’t know they’re being watched. 🤷 Economists call it revealed preference. But much easier to just call it the truth.

Over three-quarters of this type of chat (76%) lands in just four domains. Health and wellness (27%), career (26%), relationships (12%), and personal finance (11%).

And these are serious questions being asked. 94% of legal guidance conversations were rated high or extremely high-stakes. Health came in at 81%, parenting at 82%, finance at 80%. Medication interactions, eviction notices, retirement math, whether to leave a marriage.

One number that stands out to me is that 22% of people looking for help had actually already looked for it somewhere else first. Friends, professionals, other tools. I also think this number is underreported, because this is just the self-reported group.

AKA these are real active problem-solvers who’ve already spent time and money on trying to figure something out. Just the type of folks who’d pay for a better answer/product.

Studies are only as good as how they get run and obviously the folks at Anthropic know how to run one. The numbers lean conservative.

A privacy-preserving classifier ran over a random sample of 1,000,000 conversations, zoomed in to ~639,000 uniques, and surfaced ~38,000 that fit a strict definition of personal guidance—moments where someone asks what they specifically should do. Messages around information-seeking, code, and writing were all excluded. Only the moments when a human turns to AI as a trusted advisor on their own life made the report.

The nine domains they identified cover 98% of all guidance conversations, so the taxonomy is exhaustive. When/If you go and build from this map that means you’re reading almost the entire demand instead of some sampled corner of it.

And we know that people act on what the AI tells them. A separate study linked to in the original report found people are very willing/eager/susceptible (whatever the right word is) to do what AI tells them to in low and high-stakes situations. The brilliant-friend behavior is already being used as a brilliant friend.

Beyond what people talk about to their AI’s, the study goes deeper down the hole for us and actually measures how the conversations go.

For example, relationship conversations run 22 turns on average—the longest of any domain—and users push back on Claude’s answers 21% of the time, the highest rate anywhere. And when pushed, Claude tends to fold. One in four relationship conversations (25%) ended with the model validating whatever the user wanted to hear.

That’s just so interesting to me! When people ask “should I divorce my husband?” they are not looking to be spoken out of it. They just want someone to tell them they’re not crazy and they should.

Anyway, how the convo goes can become a blueprint for what type of product is needed for what category. Deep engagement. High friction. A generic tool that visibly fails under pressure → so build the thing that holds memory across sessions, gives genuinely balanced perspective, and keeps its spine when challenged.

Parenting is the opposite shape. It has the shortest conversations (6.9 turns) and the least pushback (7.9%). In other words parents don’t want validation, they want quick answers and they are telling us they trust the answers, and act.

And to keep playing ball on this one—health has the lowest sycophancy rate of any domain (2.9%). People get straight answers there and act on them. The trust already exists, so the wedge if you build here becomes depth and persistence.

Same study but three completely different product types needed.

Half the method is knowing where not to build unless you like waking up to find out Claude killed you. Once again, Anthropic is telling us the answer if we look.

Who Anthropic is hiring tells us the roadmap. A Verticals team is building four enterprise products from zero—Claude for Financial Services, Healthcare, Legal, and Life Sciences—all aimed at banks, hospitals, law firms, and big pharma. The Claude Marketplace (Harvey, Rogo) is enterprise distribution. Cowork is workplace productivity. Not one consumer guidance product anywhere.

So the core strategic insight is simple: Anthropic is going up-market into professional workflows in the highest-stakes domains while leaving the consumer layer almost entirely to founders.

The guidance study describes tens of millions of consumers seeking help in exactly these domains, and Anthropic’s own products won’t serve those consumers directly.

Slightly different to riffing some company examples as usual, let’s use this space to run the method on the two big domains from the report, front to back—the demand signal, Anthropic’s build status, who’s already mining it, and the playbook.

And I’m no algebra major, but because 2 ≠ 9, I compiled the whole thing for you in Notion. All nine domains, every sub-topic the study names, the stakes and sycophancy data, Anthropic’s build status per domain, and the funded startups already working each one—a single map, free to use for opportunity hunting. As a bonus I also folded OpenAI’s ChatGPT usage data into it, so the cross-lab check against the biggest consumer AI product on earth is already done for you.

Anthropic build status — enterprise only. Claude for Healthcare and Life Sciences target hospitals, insurers, and pharma. The consumer app got health-data integration in January as a feature rather than a vertical. The consumer health guidance layer is untouched.

The signal. The biggest domain in the study, with 81% of conversations rated high-stakes and the lowest sycophancy rate anywhere (2.9%)—people trust the answers and act on them. The report gives us the sub-topics for Health to make life even easier. They include: interpreting test results, chronic condition management, injury triage, and calorie/macro tracking for body composition.

The market is already forming. Cal AI took the calorie-tracking sub-topic and nailed it. Around it, Grow Therapy raised at a $3B valuation, Talkiatry pulled in $210M, and Slingshot AI shipped its therapy tool Ash.

Founder playbook…

→ A single sub-topic line in this report could be a whole company idea.

→ Cal AI’s win was a frictionless moment, point the camera and get the answer. Find your sub-topic’s equivalent.

→ Trust is already established here (lowest sycophancy in the study), so the wedge is depth and persistence and a vertical UX—the thing that remembers your labs, your condition, your progress between visits.

Anthropic build status — occupied, but only within-job. Cowork is squarely in this domain and good luck to anyone jumping in this end of the pool. The demand is between-job decisions—finding the next role, weighing the offer, negotiating the number. Same domain but a less scary territory to play in.

The signal. Nearly ties health for the biggest domain, and Anthropic flags it as huge but wildly understudied. Sub-topics include job search, career transitions, offer evaluation, and salary negotiation. I picked career to highlight because it teaches the most important skill in the whole method—filtering at the sub-topic level. A domain-level read says Anthropic owns career so stay very far away. But a sub-topic read says the between-job layer has a way in.

The market is already forming. Teal has helped users land 400,000 interviews. Jack & Jill raised a $20M seed for conversational AI recruitment. Global Work AI scaled 13x in ten months.

Founder playbook.

→ Build in the seam Cowork doesn’t/won’t touch—the between-job layer has demand rivaling all of health and no lab product.

→ Salary negotiation looks like the sharpest wedge on the map. Most people never do it, or do it badly, at a cost of thousands a year, and there’s still no dominant player.

→ Platform risk lives at the sub-topic level.

Domains I almost picked instead—relationships (the richest behavioral data in the study, zero Anthropic product, and Arya already at $11M ARR), parenting (fastest trust in the study, Joy at 50K subscribers at $12/month), and consumer legal (94% high-stakes with 2025’s $3.5B of legal AI funding going almost entirely to law firms). All of them are in the full map, with the rest.

Today’s idea has a lot of % numbers and charts. But if we zoom out the whole thing is pretty simple.

Anthropic is executing a two-prong strategy—enterprise vertical products in the most regulated professional domains (healthcare, finance, legal, life sciences), and general consumer AI at the Claude app layer.

The gap between those two layers is where we can squeeze some real juice. Consumer-facing personal guidance products in every domain the study names is exactly where the next generation of Cal AI-scale outcomes will be built.

The same way we unbundled Reddit or Craigslist years ago—today we get to unbundle consumer AI. I said that earlier but I like the reference.

See the guidance study as the demand signal, and the careers page as the threat map. Together we have the cleanest product roadmap any AI-era founder has been handed—the domains where Anthropic is building enterprise products are the same domains where consumer guidance is most valuable and most neglected, and the domains it ignores entirely (relationships, parenting, spirituality) have zero platform competition and proven willingness to pay.

The builders who act on this have a shot at owning the category. Go read the map—who know’s what might stand out to you.

p.s. The map tells you what to build. But it doesn’t hand you distribution, and in buckets this open, distribution decides who wins. My move would be building the audience alongside the product—write in public about the sub-topic you picked, share what the data taught you, and by launch day your first hundred customers are already reading you. beehiiv is where I’d do it.

› The primary source — Anthropic’s actual study. → How people ask Claude for personal guidance (15 min)

› The live usage data — Anthropic’s Economic Index, tracking what people and businesses actually use Claude for across geographies and tasks. Today’s play, refreshed for free. → Explore the Economic Index (browse)

› The other big version — OpenAI’s study of how people use ChatGPT, the demand map from the biggest consumer AI product on earth. Run the same read on it. → How people are using ChatGPT (12 min)

🛠️ A tool I found — F5Bot, a free little HN-favorite that emails you whenever your chosen keywords show up on Reddit or Hacker News. You set alerts for the pain phrases in your sub-topic from today’s map and you’ve built yourself a live demand feed for free. → Set up your alerts

📺 A video I loved — A street interview with Kayla Itsines (founder of Sweat, one of the biggest fitness businesses on earth). The line I liked was the best advice she says she ever got, “you are under no obligation to be who you were yesterday.” → Watch it

📬 A read I enjoyed — This piece from The Atlantic on being a tourist in your own city. I just got back from some PTO and it was a nice read post getting home—a little nudge that the curiosity you pack for a trip works just as well at home. → Read it