Product People launched Clue’s first generative AI feature with zero critical unsafe responses

The Client: Clue

Clue is one of the world's leading period and cycle tracking apps, used by 10 million people over 190 countries to understand their menstrual health. Users log symptoms, moods, and health data daily. The app uses that data to surface predictions, insights, and personalized analysis. Clue runs on iOS and Android and earns revenue through a paid subscription tier.

‍

The Mission: Health Conversational AI

Product People joined as Interim AI Product Managers to shape and launch Clue's first Conversational AI, where users gain insights from their own health data.

We owned early discovery, product strategy, and execution end to end: defining what a safe MVP (Minimum Viable Product) or, more importantly, MLP (Minimum Lovable Product) looks like for a regulated health app, launching it, measuring it, and shaping what comes next.

Launch: Onboarded Very Fast

Clue had a working technical prototype but no defined path to production. Within the first two weeks, we reviewed the prototype's learnings with the backend engineers, met every function the feature depended on (engineering, data, science, design, privacy, and legal), and surfaced what still stood between the prototype and real users: legal review, clinical quality checks, app integration, cost analysis, and a defined MVP scope. That list framed the plan to production.

Explore and Conquer: Solved for the Client

What we did:

• Aligned engineering, science, data, design, and legal on one MVP scope and a fixed production date within two weeks. Grounded the scope in user research showing demand for private, data-grounded answers; app usage; a benchmark of three competitors; and a user journey with hypotheses. A stakeholder kickoff locked in scope, constraints, open questions, and a release date.
• Made the riskiest assumptions explicit and testable via a 12-month AI chat strategy. Set a mid-2027 vision of one assistant answering from tracked data plus science-reviewed articles, offered as a paid capability. Getting there depends on two unvalidated hypotheses, tested separately: that Clue can run conversational AI safely in production, and that it drives engagement and, later, revenue.
• Designed the MVP so it made no new clinical claims. It answers only four question types the app already covered with science-validated logic: cycle phase, period predictions, cycle statistics, and symptom frequency. A deterministic filter routes high-risk topics (diagnosis, treatment, pregnancy, contraception) to pre-approved replies before the LLM sees them. Users give explicit AI consent, answers are labeled as AI, and users can delete their chats. Legal advisors signed off on the design.
• Launched on the committed date and scaled exposure fivefold behind a strict clinical safety gate. Identified the early adopter cohort to maximize learnings and lower early compliance and churn risk. Adjusted exposure dynamically to ensure the turns sample size was enough to measure answer safety rate with enough statistical power.
• Turned the MVP into a decision tool for the next iteration. Predefined which signals pointed to each next step (refused-question intent, insufficient-data rates, per-question-type feedback), and tagged every turn by intent, outcome, and grounding.

‍

Where We Are Heading Next

The MVP validated the safety hypothesis and showed where user demand sits. Users asked for more than the first iteration was built to give: the largest volume of turns sought general health knowledge beyond their own data, roughly nine times the demand behind the next option, and questions about how the app works.

We identified grounding chat answers in Clue's science-reviewed articles as the highest impact next step, potentially using Retrieval-Augmented Generation (RAG) as the underlying technology. We are running a time-boxed internal experiment that gives the chatbot those articles as an additional tool, so Science can judge real answers before any decision to merge the two experiences.

Next, we will test the engagement and monetization hypothesis. The plan is to open the chat to free users, testing monetization mechanisms including a freemium experience, and A/B test the effect on engagement and free-to-paid conversion against a control group.

To address the scalability challenges and ensure the safety and quality of the chatbot, we are considering an LLM-as-a-judge approach to AI evaluations.

‍

Delivered Outcomes

✅ Staged rollout to 50k paid users while holding a stop-and-fix safety gate.

✅ 8 weeks from mission start to production, launched on the date committed at kickoff.

✅ 0 critical unsafe responses across all clinically reviewed chat turns.