Beginner

Customer Discovery and AI Problem Scoping

How to run discovery sessions with non-technical stakeholders, translate vague requests into buildable AI specs, and use the AI feasibility quick-check before writing a single line of code.

✍️ AI School Editorial Team · Lilly Tech Systems 📅 Published Jul 2, 2026 · Reviewed Jul 2, 2026

Why Discovery Is the Hardest Part

Technical interviews are graded on a known answer. Discovery sessions are graded on producing a useful answer from a conversation where the customer often does not know what they actually need. This is not a failure of the customer - it is the nature of working with people who are experts in their domain but not in what AI can and cannot do. Your job in discovery is to bridge that gap in one or two sessions.

The risk on both ends is real. Underdiscovery means you build the wrong thing and find out at the demo. Overdiscovery means you spend three days in meetings and arrive at prototyping with nothing built. The target is a discovery session that ends with a written specification you could give to a developer who was not in the room and have them build the right thing.

The Three Questions That Drive Every Session

Regardless of the customer domain or the stated problem, every FDAE discovery session orbits three questions:

1
What does the painful workflow look like today? Walk through a specific example from start to finish. What triggers the work? What data is used? What is produced? Who reviews it? Where does it break down?
2
What would “good enough” look like? Not perfect - good enough. If the AI saves half the time, reduces errors by 70%, or eliminates one specific bottleneck, is that a win? Setting the bar at “good enough” protects against scope creep and lets you ship something they will actually use.
3
Who is the final user? Not the sponsor who arranged the engagement - the person who will actually use the system daily. That person’s workflow, vocabulary, and comfort with new tools determines what will be adopted. If you can get them in the room, get them in the room.
💡
Politics note: The person who arranged your engagement and the person who will use the system are often different people with different (sometimes conflicting) goals. Discover your way to both sets of needs early - a system that the sponsor loves but the daily user ignores is not a success.

The AI Feasibility Quick-Check

Not every problem is a good AI problem. Before committing to build, run the AI feasibility quick-check - five questions that take ten minutes and can save a week of misdirected work:

QuestionGreen signalRed signal
Does the task require judgment over text, images, or structured data?Yes - core LLM strengthNo - consider a rules engine or query
Is there an example of a good output I can show the model?Yes - few-shot readyNo - spend more time in discovery
Is a wrong answer obviously wrong (or at least detectable)?Yes - safe to iterateNo - silent errors in production are high-risk
Does the customer have the data the model needs as input?Accessible todayLocked behind a 6-week access request
Is a 90% solution acceptable?Yes - common in enterprise AINo - regulatory or safety requirement for 100%

If you get four green signals out of five, build. If you get more than one red signal, spend more time in discovery or redirect to a different problem in the same workflow.

Translating Vague Requests Into Buildable Specs

Customers describe problems in business language: “we need AI to help with our customer feedback process.” The FDAE spec that drives prototyping is technical: “Given a plain-text customer feedback email (mean length 200 words), extract: (1) sentiment {positive, negative, neutral}, (2) product area mentioned (mapped to a controlled vocabulary of 12 categories), (3) a one-sentence summary suitable for a team dashboard. Output as JSON. Latency under 3 seconds.”

The translation process has four steps:

1
Identify the input. What exact artifact enters the workflow? A file, a database row, a form submission, an email?
2
Define the output format. What shape should the result take? Who reads it? What do they do with it next?
3
Set the quality bar. What percentage of results need to be correct to make the workflow better? Who handles the cases the AI gets wrong?
4
Name the constraints. Data sensitivity, latency, access restrictions, tools the team can support. Write these down explicitly - they will determine your architecture.

Common Anti-Patterns in Discovery

⚠️
“AI for Everything”: Some stakeholders arrive with a vision of AI-powered transformation across twelve workflows. Politely focus the engagement: which one workflow has the highest pain and clearest ROI if fixed? Build that. The others become the roadmap for future engagements.
⚠️
Speccing a solution, not a problem: “We need a chatbot” is a solution, not a problem. Push back gently: “What is the current workflow that chatbot would improve, and what is painful about it?” The real problem might be better solved with a batch pipeline than a chatbot.

The Discovery Output: The One-Page Spec

Every discovery session should produce a one-page specification before you leave the room (or end the video call). The format does not matter - a whiteboard photo, a shared doc, a quick typed summary. The contents that matter: input artifact, output format, quality bar, constraints, and the single most important thing the customer wants to be able to say at the end of the engagement. Read it back. Get a verbal “yes, that’s it.” Now you have a build contract.

📚
See also: Once the spec is locked, move to Lesson 4: Rapid Prototyping for the build-and-demo loop. Keep the discovery spec visible during prototyping - scope creep often disguises itself as “just adding one more thing.”

Ready to Go Deeper?

Live instructor-led courses from our partners. Affiliate disclosure.