Discovery: Finding the Real Problem
How to talk to a customer, watch the work, and describe the problem before you build.
Customers often ask for a tool when they need a result. Discovery finds the real problem before you build. Talk to the sponsor and the daily users. Watch the work. Ask for real examples. End by writing the problem in a few sentences and reading it back to them.
- Discovery
- The early step of learning how the work really happens.
- Baseline
- How things are today, measured before any change.
- Gatekeeper
- A person in security, legal or compliance who must approve.
Customers often ask for a tool. What they need is a result. Discovery is how you find the difference. Good discovery saves weeks of building the wrong thing. This lesson shows how to run discovery with an AI project in mind.
Start with the person, not the tool
A customer may say, we want a chatbot. That is a solution, not a problem. Ask what happens today. Who does the work? How long does it take? What goes wrong? What would change if it were fixed?
Your goal is to describe the problem so well that the customer says, yes, exactly. Only then do you talk about AI.
Who to talk to
- The sponsor: the person who pays and wants a result. Ask what they will report upward.
- The users: the people who do the work every day. They know the real steps and the tricks.
- The owners of data and systems: they decide what you may touch, and when.
- The gatekeepers: security, legal and compliance. Meet them early, not at the end.
Questions that find the real problem
THE WORK TODAY - Walk me through the last time this happened. What did you do first? - How long does it take? How often does it happen? - Where do mistakes or delays come in? - What do you do when it goes wrong? THE PAIN - What does a bad day look like? - What does this cost you in time, money or customers? - What have you already tried? Why did it stop? THE DATA - Where does the information live? Who owns it? - Can I see 20 real examples, with private details hidden? - How clean is it? Who fixes it when it is wrong? THE RESULT - If this worked perfectly, what would be different in 3 months? - Who would notice first? What number would move? THE LIMITS - What must never happen? What needs a person to approve? - Which rules apply: privacy, law, company policy?
Watch the work
If you can, watch someone do the job. Sit next to them for an hour. You will see steps that nobody mentions in a meeting. You will see copy and paste between screens, side notes on paper and small workarounds. Those are the places AI can help.
Is this a good AI problem?
Not every problem needs AI. Ask four things. Is the work repeated often? Is the input messy text, documents or images? Can a person check the result quickly? Is a small mistake survivable?
If the answers are yes, AI may fit. If the work is rare, high-stakes and impossible to check, be careful. Sometimes the honest answer is that a simple rule or a spreadsheet is better. Saying so builds trust.
Write it back
End discovery by writing the problem in two or three sentences. Read it back to the customer. Add who is affected, what it costs and what success would look like. Ask them to correct it. A short summary they agree with is worth more than a long report they never read.
- ✓Start with the person and the result, not the tool. A request for a chatbot is a solution, not a problem.
- ✓Talk to the sponsor, the users, the data owners and the gatekeepers. Users know how the work really goes.
- ✓Watch the work and ask for real examples. The best AI opportunities hide in small workarounds.
- ✓Test the fit: repeated work, messy input, quick checking, and survivable mistakes. Be willing to say that AI is not needed.
Run a mini discovery
- Choose a real task a friend or colleague does every week. Ask them five questions from the discovery list.
- Ask to see three real examples of the work, with private details hidden.
- Write the problem in three sentences and read it back to them.
- Score the task on the four fit questions. Decide whether AI is a good fit, and say why.
Q1Why not talk only to the sponsor?show answer
Sponsors describe the work as they imagine it. The daily users know the real steps, the workarounds and where it breaks.
Q2What are the four questions for AI fit?show answer
Is the work repeated, is the input messy, can a person check the result quickly, and is a small mistake survivable?
Q3How should discovery end?show answer
With a short written summary of the problem, who is affected, the cost and the success picture, read back to the customer for correction.