The AI Product Discovery Process That Actually Works

TL;DR

• Traditional product discovery asks what features should we build. AI product discovery should ask what should our AI understand about each user, because understanding is what makes AI products feel personal
• The AI product discovery process produces an understanding map: the dimensions of user knowledge, the learning curve for each dimension, and how understanding changes the AI experience
• Teams that run understanding-first discovery ship AI products that compound in quality over time, while feature-first teams ship products that feel the same at month 6 as month 1

AI product discovery should ask “what should our AI understand about each user” rather than “what features should we build,” because understanding drives the personalization that determines long-term retention. Feature-first discovery produces products that feel the same at month 6 as month 1, while understanding-first discovery produces products that compound in quality with every interaction. This post covers the four-step understanding discovery process, how to create an understanding map, and why this approach reduces 90-day churn from 35% to 12%.

Why Feature Discovery Fails for AI

Feature discovery assumes that product value comes from capabilities, things the product can do. For traditional software, this is mostly true. A spreadsheet becomes more valuable as it gains more functions.

AI products are different. Their value comes primarily from understanding, how well the AI knows and serves each specific user. A feature-rich AI that treats everyone the same is less valuable than a feature-lean AI that genuinely understands each user.

Think about the difference between a restaurant with a 50-page menu and a restaurant with a waiter who knows your preferences. The first has more capabilities. The second provides a better experience. Most AI teams are building the 50-page menu.

Feature discovery for AI products produces more capabilities with no improvement in understanding. Understanding discovery produces deeper knowledge of each user that makes every existing feature work better.

The Understanding Discovery Process

Here is the process that changed our product.

Step 1: Map the understanding dimensions. Instead of listing features, list everything your AI would need to know about a user to be maximally useful. Not data points, understanding dimensions.

For our productivity AI, we identified five dimensions: work domain (what field they work in), expertise level (beginner to expert), communication style (brief vs. detailed, formal vs. casual), primary goals (what outcomes they are trying to achieve), and constraints (time pressure, team context, tool limitations).

Step 2: Prioritize by impact. Not all understanding is equally valuable. We scored each dimension on two axes: how much it changes the AI experience (impact) and how early we can learn it (accessibility).

Expertise level scored highest. Knowing whether someone is a beginner or expert changes almost every response. And you can learn it within the first two interactions.

Step 3: Design the learning curve. For each priority dimension, define when and how the AI should learn it. First interaction? Explicit question? Implicit inference after three sessions?

The learning curve is a product design artifact that looks like a roadmap but describes understanding instead of features:

• Interaction 1: Learn primary goal (explicit question)
• Interaction 3: Learn expertise level (inferred from vocabulary and questions)
• Interaction 10: Learn communication preferences (inferred from feedback patterns)
• Day 30: Learn recurring patterns (observed from usage)

Step 4: Design understanding-driven features. NOW you design features, but each feature is shaped by understanding. Instead of add a summary feature, you design a summary feature that adjusts length based on the user’s time constraints, depth based on their expertise level, and vocabulary based on their domain.

The Understanding Map

The core output of understanding-first discovery is the understanding map. It has three layers:

Layer 1: Dimensions. The aspects of each user your AI needs to model. Be specific. Not preferences but communication style preference: concise vs. detailed, technical vs. plain language, formal vs. casual.

Layer 2: Learning curve. When each dimension is learned, how confident the AI should be, and what learning method to use (explicit, implicit, or hybrid).

Layer 3: Experience impact. For each dimension at each confidence level, how does the AI behavior change? This is the most important layer because it connects understanding to value.

Example from the map:

Dimension: Expertise level.

• Low confidence (interaction 1-2): Ask clarifying questions to gauge expertise. Default to intermediate explanations.
• Medium confidence (interaction 3-5): Adjust vocabulary and depth based on inferred expertise. Confirm with the user.
• High confidence (interaction 10+): Automatically calibrate all responses. Expert users get concise, technical answers. Beginners get step-by-step explanations.

1// The discovery output: an understanding map← Not features, but knowledge
2const understandingMap = {
3  dimensions: [
4    { name: 'expertise_level', impact: 'high', learnBy: 'interaction_3' },
5    { name: 'primary_goal', impact: 'high', learnBy: 'interaction_1' },
6    { name: 'communication_style', impact: 'medium', learnBy: 'interaction_10' },
7    { name: 'constraints', impact: 'medium', learnBy: 'day_7' },
8    { name: 'patterns', impact: 'high', learnBy: 'day_30' }
9  ],
10  qualityTarget: {
11    fitScore: { interaction5: 0.60, interaction20: 0.70, day90: 0.80 }
12  }
13};

Running Discovery Interviews for Understanding

The interview technique changes when you are discovering understanding rather than features.

Instead of: What problems do you have with AI products? Ask: When your AI assistant gives you a bad response, what was wrong with it? Was it too generic? Too detailed? Wrong topic? Right topic but wrong level?

Instead of: What features would make this product better? Ask: What would the AI need to know about you to give you a perfect response every time? If it could read your mind, what would it see that matters?

Instead of: How do you use the product? Ask: Think about the difference between how you use the product and how your colleague uses it. What is different? What would the AI need to understand about each of you to serve you both well?

These questions surface understanding dimensions, not feature requests. The answers tell you what the AI needs to know, not what the AI needs to do.

The Compounding Effect

The most powerful benefit of understanding-first discovery is that it creates products that compound.

A feature-first product launches with 10 features. To improve, you add feature 11, then 12. Each addition is linear. The product does not get better at its existing features, it just does more things.

An understanding-first product launches with fewer features, but each one improves as the AI learns more about the user. The same feature at interaction 100 is dramatically better than at interaction 1. The improvement is exponential because understanding compounds, each new insight makes every existing feature work better.

This compounding is why understanding-first products have better long-term retention. Users do not stay for features. They stay because the product keeps getting better for them specifically.

Trade-offs

Understanding-first discovery is slower initially. Feature discovery produces actionable backlogs quickly. Understanding discovery requires deeper thinking about what knowledge matters and how to acquire it. Budget more time for the discovery phase.

Not all products need deep understanding. Simple utility AI (convert this file, translate this text) benefits less from understanding than advisory AI (help me make this decision, plan this project). Match the depth of your understanding map to the depth of the user relationship.

Understanding maps are hypotheses. You will discover that some dimensions matter less than you thought and others matter more. The map should be treated as a living document, updated as you learn from real users.

Engineers may resist the shift. Feature-first discovery produces clear engineering specs. Understanding-first discovery produces requirements that feel vague until they are decomposed into technical requirements. Bridge the gap by pairing each understanding dimension with specific data schemas and API contracts.

What to Do Next

1. Run one understanding interview. Take your next user research session and swap the questions. Instead of asking about features, ask what the AI would need to know about the user to be perfect. You will get different, and more useful, answers.

2. Draft your understanding map. List 5 dimensions of user understanding that would most improve your AI product. For each, note when the AI should learn it and how the learning changes the experience. Share with your team and iterate.

3. Instrument one dimension. Pick the highest-impact dimension from your map, implement a way to learn it, and measure whether it changes quality scores. This single experiment will validate the approach. See how Clarity powers understanding-first AI products.

Feature discovery asks what should we build. Understanding discovery asks what should we know. The second question produces AI that gets better for every user, every day. Start the discovery.

References

1. Product vs. Feature Teams
2. only 1 in 26 unhappy customers actually complains
3. Qualtrics notes in their churn prediction framework
4. Continuous Discovery Habits
5. 80% of features in the average software product are rarely or never used

Related

• AI Product Requirements Nobody Writes Down
• Why AI Products Churn Faster Than SaaS
• The Feedback Loop That Compounds