Founder Playbook · Consumer AI · 2026

Building Trusted AI Guidance Products

What Anthropic's personal guidance report means for founders building narrow, trusted, high-retention AI products.

I do not read the Anthropic report as a signal to build AI doctors, AI lawyers, or AI therapists. I read it as a signal that people are already bringing high-stakes personal decisions to AI, and the opportunity is to build safer, narrower, more trusted workflows around those decisions.

Read the Playbook View the Framework

What the report says

The signal is behaviour.

Anthropic's report is useful because it shows people already asking AI for decision support in real personal contexts.

Anthropic analysed a privacy-preserving sample of Claude conversations.
A measurable share were personal guidance conversations where people were asking what to do next.
Those conversations covered health and wellness, career, relationships, personal finance, legal, parenting, personal development, ethics, spirituality, and adjacent consumer decisions.
The important signal is not: build generic advice bots.
The important signal is that users are already using AI to reason through real personal decisions.

Official source: Anthropic, "How people ask Claude for personal guidance"

Figure 1 recreated

Guidance conversations by domain

10 clusters. 37,657 guidance-seeking conversations. Percentages are recreated from Anthropic's published Figure 1 values.

Health / WellnessTest results, chronic conditions, injuries, respiratory symptoms, nutrition, and body-composition goals.
Professional / CareerJob search, opportunity identification, career transitions, offer evaluation, and salary negotiation.
RelationshipsCurrent relationship communication, romantic intent, leaving difficult relationships, and conflict repair.
FinancialDebt management, repayment, home purchase, mortgage decisions, retirement, and wealth planning.
Personal DevelopmentStudy planning, exam prep, self-improvement, goal setting, digital addiction, and compulsive habits.
SpiritualityDivination and astrological guidance, religious reflection, and spiritual practice.
LegalReal estate, rental disputes, employment questions, and financial-legal disputes.
ConsumerProduct comparison, purchase decisions, rental and housing decisions, vehicle purchase, and repair.
ParentingInfant care, child caregiving, school-behavior parenting advice, education planning, and life logistics.
OtherTravel and leisure planning, immigration, geopolitical relocation, and other personal decisions.

Anthropic's guidance conversation map shows demand clustering around health/wellness, career, relationships, and finance, but the strategic lesson is trust, not volume.

The winning product is not "AI for everything."The winning product is AI for one painful decision, with the right workflow, context, memory, output, and escalation path.

Trusted AI guidance products are not thin chatbot wrappers. They are domain-specific decision systems that help users prepare, decide, act, and follow through.

Opportunity map

Where I would look for wedges.

Here is how I would think about the wedge: pick a painful decision, produce a useful artifact, and make the boundary obvious.

Health and wellness

Wedge I would test: Pre-visit decision prep
Why it works: Helps users organise symptoms, questions, and lifestyle context before speaking to a professional.
What not to build: Diagnosis bot or emergency triage tool.

Career

Wedge I would test: Offer and manager-conflict copilot
Why it works: Converts uncertainty into scripts, tradeoffs, and decision memos.
What not to build: Generic resume wrapper.

Relationships

Wedge I would test: Conflict debrief and message planner
Why it works: Adds perspective, slows emotional escalation, and helps users communicate clearly.
What not to build: Validation bot that always agrees.

Personal finance

Wedge I would test: Cash-crunch and bill-shock planner
Why it works: Helps users prioritise, organise, and understand options.
What not to build: Investment picker or guaranteed financial advice.

Legal

Wedge I would test: Issue intake and evidence organiser
Why it works: Helps users prepare facts, timelines, documents, and questions before legal support.
What not to build: Lawyer replacement or court-outcome predictor.

Parenting

Wedge I would test: Routine and school-support planner
Why it works: Helps parents structure recurring challenges and communication.
What not to build: Medical or child-safety emergency adviser.

Personal development

Wedge I would test: Decision journal and pattern review
Why it works: Memory and reflection compound over time.
What not to build: Generic motivation chatbot.

Ethics

Wedge I would test: Structured dilemma memo
Why it works: Helps users reason through tradeoffs without pretending to be morally authoritative.
What not to build: One-answer moral oracle.

Spirituality

Wedge I would test: Reflection companion with boundaries
Why it works: Supports meaning-making while being explicit about limits.
What not to build: Fortune-teller or certainty machine.

Signals vs false signals

Validation has to travel.

If users say "this is amazing" but never return with the next real-world step, you may have a demo, not a company.

True signals

Users share real context.
Outputs travel outside the app.
Users return with outcomes.
Users ask for tradeoffs, not just validation.
The workflow creates a useful artifact.

False signals

Screenshot virality.
Gratitude-heavy one-offs.
Prompt obsession.
Compliance theatre.
Wrapper comfort.

Why thin wrappers are weak

The prompt is not the moat.

A thin wrapper is a product where most of the value comes from calling a general-purpose AI model and showing the answer in a nicer interface.

The new moat is not the prompt. The new moat is the workflow, the data loop, the trust layer, and the distribution channel.

The base model can add the feature.
Competitors can copy the prompt.
Users can recreate the workflow in general chat.
There is no proprietary context.
There is no domain-specific evaluation.
There is no trusted workflow.
There is no durable distribution advantage.

Trust layer

Trust is the product.

Define what the product helps with.

Define what the product does not help with.

Ask for context when context is missing.

Show assumptions.

Show tradeoffs.

Push back when the user is seeking validation.

Escalate when stakes are high.

Make memory opt-in.

Make deletion simple.

Measure harm, drift, and follow-through.

Execution plan

How I would de-risk it in 90 days.

Days 1-30

Pick the wedge

Choose one domain.
Choose one narrow job-to-be-done.
Define user, trigger, context, output, and escalation boundary.
Build a manual concierge prototype.
Create a small hard-case evaluation set.
Interview users after the output.
Measure whether the artifact is used in the real world.

Days 31-60

Ship the workflow

Convert the manual workflow into a narrow product.
Add structured intake.
Add consented memory.
Add uncertainty and assumptions.
Add exportable artifacts.
Add escalation logic.
Run five GTM experiments.
Track repeat use and follow-up outcomes.

Days 61-90

Prove the trust loop

Double down on the highest-trust workflow.
Improve the evaluation set.
Add human review or partner handoff.
Publish internal quality gates.
Improve onboarding around repeat behaviour.
Measure retention by outcome.
Document provider risk and fallback options.

Templates

Five working templates for the founder.

If I were building in this category, I would keep these artifacts close and force every product decision through them.

Wedge Selection Scorecard

Pain intensity
Frequency or replayability
Willingness to share context
Quality of artifact you can generate
Ability to measure a good outcome
Clear escalation path
Distribution adjacency
Revenue trigger strength
Provider independence
Liability surface clarity

Trust Wedge Design Checklist

The problem statement is narrow.
The scope boundary is visible.
The product asks for minimum viable context.
The output shows assumptions.
The output shows tradeoffs.
The system can respectfully challenge the user.
There is a defined escalation path.
Memory is opt-in.
Deletion is understandable.
Important outputs are auditable.
The workflow is tested with hard edge cases.
The product does not reward dependency.

GTM Experiment Backlog

One search-intent landing page for one wedge.
One answer-style page designed for AI summarisation.
One exportable artifact users can send to another person.
One partner pilot through an existing trust holder.
One creator-led walkthrough of the workflow.
One referral ask at the moment of user relief.
One reactivation prompt timed to the next likely trigger.
One paywall test after the first real artifact.
One challenge my thinking mode versus validation mode.
One assisted handoff test to a human or outside resource.

Model and Provider Risk Checklist

Can the workflow survive a weaker model?
Are prompts, policies, and evaluations portable?
Can you switch models without rewriting product logic?
Do you know what is model-derived versus workflow-derived?
Could a provider policy change break conversion?
Are unsafe cases caught by your system or only by the provider?
Do you log regressions after model updates?
Do you have a fallback experience if quality drops?
Are your economics resilient to price changes?
Does the moat still exist if the base model adds a similar feature?

Launch Readiness Gate

The wedge is narrow.
The artifact is useful.
The user understands the product boundary.
The system performs acceptably on hard edge cases.
Escalation works.
Memory and deletion are understandable.
The product improves from follow-up outcomes.
The workflow is better than freeform chat.
The team can explain why users should trust it.
The team can explain when users should not trust it.

Closing note

Trusted AI guidance products will not win by pretending to be universal experts.

They will win by being narrow, useful, honest, and behaviour-changing.

The best founder move is not to build an AI therapist, lawyer, doctor, coach, or adviser.

It is to build the trusted preparation, decision, and follow-through layer around one painful human situation.

Want to build this kind of AI product?Explore my work on AI fluency, startup strategy, and product design.Visit Rushi's Portfolio