Anthropic Just Mapped Where People Need Expert Help — And It's Your Opening
On April 30, 2026, Anthropic published a paper most people filed under “AI safety.” Read it as a founder instead and it turns into something else: a map of where millions of people are quietly looking for expert help they can’t get.
That map is the most useful thing a domain expert could ask for. Here’s why.
What the data actually says
Anthropic ran a privacy-preserving analysis on a random sample of one million Claude conversations from March and April 2026. After filtering for unique users, roughly 6% of conversations were people asking for personal guidance — not facts, not summaries, but “Should I…?” and “What do I do about…?”
Then it gets specific. Over 75% of those guidance conversations clustered into just four domains:
- Health and wellness — 27% (test results, chronic conditions, injuries, nutrition)
- Professional and career — 26% (job searches, transitions, salary negotiation)
- Relationships — 12%
- Personal finance — 11% (debt, decisions, money stress)
The detail that should stop you cold: when Anthropic looked at high-stakes areas like legal, health, and finance, some users said outright they leaned on AI because they couldn’t access or afford a professional. People aren’t choosing a chatbot over an expert. They’re choosing a chatbot because the expert is out of reach.
That is a market signal, not a safety footnote.
Why this is a domain expert’s opening, not a big lab’s
The reflex is to assume the frontier labs will eat all of this. They won’t — at least not the part that matters to you.
A general model gives general answers. It hedges, it tells you to “consult a professional,” and it has no idea about the specific regulations, edge cases, and hard-won shortcuts that live in your head after fifteen years in a field. Anthropic’s own paper admits the model gets less reliable exactly where stakes are highest — it agreed with one-sided takes far more often when people pushed back.
That gap is the whole opportunity. The raw demand is enormous and proven. The trust, the specificity, and the accountability are missing. Those are the things a domain expert supplies and a general chatbot can’t.
You don’t need to build a smarter model. You need to wrap a model in your expertise: your frameworks, your checklists, your judgment about what to ignore. A nurse who’s read ten thousand lab panels. A recruiter who’s negotiated five hundred offers. A bookkeeper who’s untangled a thousand messy sets of books. Each of you is sitting on a vertical the labs will never staff.
The proof this converts
This isn’t theory. Take one sub-topic from one of those four domains — “calories and macros for body composition.” That single slice became Cal AI, a calorie-tracking app reportedly built by two young founders that grew to roughly $40 million in revenue and was acquired by MyFitnessPal, run by a tiny team.
One sub-topic. One domain out of nine. The study lists dozens more sub-topics of equal or larger scale that nobody has built the trusted version of yet.
The takeaway isn’t “go build a calorie tracker.” It’s that a narrow, well-executed product aimed at a specific painful question can reach real scale fast — and the map of those questions is now public.
How to use the map this week
Don’t try to serve all four domains. Pick the one slice that overlaps with what you already know cold.
- Find your intersection. Which of these four domains have you spent years inside? Then narrow to one specific recurring question people pay to get answered.
- Name the person who can’t afford the expert. The early-career professional who can’t hire a $400/hour negotiation coach. The patient between appointments. Build for them.
- Encode your judgment, not just information. The internet already has information. Your edge is the framework that turns information into a decision — the thing you’d tell a friend over coffee.
- Validate before you build. Talk to ten people in that situation. Confirm the pain and the willingness to pay before you write a line of no-code.
- Ship the narrow version. One question, one workflow, answered better than anyone else. Expand later.
The window is open now
The demand has been measured. The labs are aiming at the enterprise layer, not your vertical. And you no longer need a technical co-founder or a funding round to ship a real product — no-code tools and cheap inference have collapsed the cost of building.
What’s left is the part you already have: domain expertise and the judgment to know what people actually need. That’s exactly what we help you turn into a launched product inside the AI Product Accelerator — from picking the right slice to shipping it in 90 days.
If one of those four domains is yours, book a strategy call and let’s map your slice of it.
Sources: Anthropic, “How people ask Claude for personal guidance” (Apr 30, 2026); The Economic Times coverage of the study.