The Ultimate Guide to AI Product Discovery
AI product discovery is the work of deciding what to build before you build it. That sounds obvious. It is also the step most teams skip, because the tools now turn an idea into a working prototype over a weekend.
I have spent 18+ years building products in startups, scaleups, and S&P 500 companies. The pattern today is always the same: teams produce demos faster than they produce understanding. Before I write a line of code, I want to know three things. Who else is solving this problem. What technology already exists. Whether the problem is worth solving at all.
This guide is how I run discovery so that fast building becomes an advantage instead of a trap.
What AI product discovery really means
Classic discovery asks four questions. Is the problem real? Can people use the solution? Can we build it? Can we make money from it? AI adds a new layer to each one.
AI does not give the same answer every time. Results change with the prompt, the model, and the data. So feasibility is no longer a simple yes or no. It becomes a question of how good, how fast, and how expensive the answers are.
Value depends on the model being right often enough for people to trust it. · Usability depends on how the product behaves when the model is wrong, not when it is right. · Feasibility depends on quality, speed, and cost on real data. · Viability depends on the math per user, which a single model price change can rewrite overnight.
AI product discovery is not classic discovery with an AI feature added on top. It is the practice of reducing four kinds of risk while the ground keeps moving.
Know the market before you build
Understanding the problem starts with understanding the market. Before I commit to a direction, I want to know who already owns this problem and how they talk about it.
Direct competitors are easy to find. The more useful insight usually sits somewhere else: spreadsheets, manual workflows, agencies, or people simply doing nothing about it. That is your real competition.
Then I look for signals that show whether the space is crowded, new, or ignored. Funding news, product reviews, pricing pages, integration lists, and sales messaging all tell you what buyers care about and where the incumbents are weak. If everyone is selling speed, there may be room for quality. If everyone is selling quality, there may be room for simplicity.
This is not a slide deck exercise. It sharpens the problem statement and keeps you from shipping another copy of what already exists.
Map the direct competitors and the tools they compete against. · Find the indirect ones: spreadsheets, services, workarounds, and inaction. · Read reviews, forums, and sales pages to spot the real pain points. · Compare pricing, positioning, and integrations to find the gaps. · Write your difference in one sentence before you design anything.
The four risks to reduce
With the market clear, the work turns to risk. These are the four I keep on the wall.
Value: Is the problem real, Is it painful and frequent, Would users pay or switch
Usability: Can users reach the right output, Do they trust it, Can they recover from wrong answers
Feasibility: Model quality on real data, Speed under load, Cost per successful task
Viability: The math per user at scale, Exposure to model price changes, Legal, privacy, and compliance
If you cannot say something concrete about all four, you are not ready to build. You are ready to run one more loop.
A simple discovery loop
I run discovery as a short loop that repeats. Each pass takes days, not months.
Problem. Write it in one sentence, in the user's own words. · Signal. Collect at least five pieces of evidence that it is real. · Prototype. Build the cheapest thing that could deliver the value. · Evaluation. Test it on real inputs and real users, not curated ones. · Decision. Continue, pivot, or park it. Write down what you learned.
Discovery is not a phase you finish. It is a loop you get better at running.
Techniques that work for AI products
A few techniques earn their place in almost every cycle I run.
Interview users about what they do today, not what they think about AI. · Run Wizard of Oz tests where a human plays the model to check the workflow. · Prototype prompts in a spreadsheet or notebook before you touch the UI. · Build a small labeled test set so you notice when quality drops. · Run the AI in shadow mode beside the current workflow, without acting. · Open a narrow beta for the users who feel the problem hardest.
Choose the technology on evidence, not hype
Once the problem and the market are clear, the next question is what technology solves it most efficiently. AI moves fast, so the best answer from six months ago is rarely the best answer today. Building on the wrong foundation can cost you months.
Model quality is only part of the comparison. I also look at cost per task, speed at scale, integration effort, how data is handled, and how often the vendor changes things. Sometimes a frontier model is overkill. Sometimes a smaller open model, a set of rules, or a mix of both is the better design. Sometimes the right answer is not to build AI at all, because a product already on the market solves it well.
The trap is falling in love with one model, framework, or vendor. Keep the problem first and let the evidence pick the technology.
List the models, APIs, and platforms that could solve this today. · Compare them on quality, cost, speed, scale, and integration effort. · Consider hybrids: rules plus AI, retrieval plus generation, human in the loop. · Check whether an off the shelf product removes the need to build at all. · Pick the option the evidence supports, not the one in the headlines.
Common traps
Building before checking who already owns the problem and how they solve it. · Treating building as progress, because the tools make it feel that way. · Demo driven roadmaps that impress stakeholders and confuse real users. · Ignoring speed and cost until they break the business model. · Skipping evaluation, then finding every regression in production. · Obsessing over the model, when the leverage sits in data, workflow, and UX.
Why discovery compounds
AI made building cheap. That made discovery more valuable, not less. The teams that win are not the ones with the best model access or the most polished demos. They are the ones who read the market clearly, choose technology on evidence, and learn faster than everyone else.
If you want to connect this to a longer range plan, pair it with my 3x3x3 framework for vision, strategy, and evidence, and my thesis on why product management still matters in the AI era.