{}Route To AI
Courses
← AI Product Manager
19 min

User Research & Problem Framing

Find the user's job, write the problem down and test ideas cheaply before building.

Lesson 2 of 10
0 done
In plain English

Start from the user's job, not from AI. Learn it with interviews about the last time, observation and data. Write a problem statement with real numbers. Test cheaply with prompt prototypes, a Wizard of Oz and small pilots, and trust what users do more than what they say.

Job to be done
The task a user is trying to get done.
Problem statement
A short, evidence-based description of who has what problem.
Wizard of Oz
A test where a person secretly does the AI's work.
Picture it
From a hunch to a tested problem
1
Learn the user's job
Interviews about the last time
2
Write the problem down
With real numbers
3
Test cheaply
Prompt prototype, Wizard of Oz
4
Decide
Build, change or drop

Good products start with a clear problem. AI makes it tempting to start from the tool instead. This lesson shows how to find real problems, describe them precisely and test ideas cheaply before engineers build anything. It uses simple research methods that work for any product, with a few tricks for AI.

Find the user's job

People do not want AI. They want to get something done. Ask what job the user is trying to do, when it happens, what they use today and what makes it hard. Look for jobs that are frequent, slow, error-prone or disliked.

Research methods that work

  • Interviews: ask about the last time the problem happened. What did you do? What did you try? What was hard?
  • Observation: watch people work. They skip steps in their story that you can see on screen.
  • Data: look at support tickets, search logs, time records and error reports.
  • Surveys: good for sizing a problem you already understand, weak for discovering new ones.
  • Competitor and workaround review: what do people do now, such as spreadsheets and copy-paste?
interview-guide.md
START          Tell me about your role and a typical day.
LAST TIME      Think of the last time you had to {{do the job}}. Walk me through it.
STEPS          What did you do first? Then? What tools did you use?
PAIN           What was slowest or most annoying? What went wrong?
WORKAROUNDS    What do you do to get around that?
COST           What does it cost you: time, money, mistakes, stress?
AI             Have you tried any AI tool for this? What happened? Would you trust it? When not?
SUCCESS        If this were fixed, what would be different next month?

Do not pitch your idea. Listen. Ask why three times.

Write the problem down

problem-statement.md
WHO:        {{the user, in specific words}}
JOB:        {{what they are trying to get done}}
WHEN:       {{how often, in what situation}}
TODAY:      {{how they do it now, how long it takes, what goes wrong}}
COST:       {{time, money, errors; use real numbers}}
SUCCESS:    {{what would be better, and by how much}}
EVIDENCE:   {{interviews, data, observations that support this}}
NOT ABOUT:  {{what this problem is not}}

Test ideas before building

AI lets you test fast. A chat model plus a handful of real examples can tell you in a day whether an idea is promising. Do this before you ask for engineering time.

  • Prompt prototype: paste 20 real cases into a chat tool with a good prompt. Review the answers with a user.
  • Wizard of Oz: show users a screen where a person secretly does the AI's work. See how they react.
  • Concept test: show a mock-up and ask what they would do with it, and what they would worry about.
  • Pilot with a few users and a narrow task, with a person checking all results.

What you want to learn

  • Do users care enough? Would they change their routine?
  • How good must it be to be useful? What error is acceptable?
  • What happens when it is wrong? Who pays?
  • Which part of the job is best for AI, and which should stay human?
  • Would they trust it, and what would build that trust?
⚠Beware of polite yes answers. Users say an idea sounds nice. What matters is what they do. Ask for a commitment: time, a pilot, a switch from their current tool.
Key takeaways
  • ✓Start from the user's job, not from AI. Look for jobs that are frequent, slow, error-prone or disliked.
  • ✓Use interviews about the last time, observation, data, light surveys and a review of workarounds.
  • ✓Write a problem statement with who, job, when, today, cost, success and evidence.
  • ✓Test cheaply with prompt prototypes, a Wizard of Oz, concept tests and small pilots, and watch what users do, not what they say.
Hands-on exercise

Run a mini discovery

  1. Choose a task that a friend does every week. Interview them with five questions from the guide.
  2. Write a problem statement with real numbers.
  3. Run a prompt prototype on ten real examples. Show the results to your friend.
  4. Write what you learned: did they care, how good must it be, what worried them?
Check yourself
Q1Why ask about the last time instead of what users usually do?show answer

A real, recent event gives true details. General answers are tidied up and miss the steps and mistakes.

Q2What is a Wizard of Oz test?show answer

A test where a person secretly does the AI's work behind a screen, so you learn how users react before building anything.

Q3Why be careful with polite yes answers?show answer

People say ideas sound nice without meaning it. Actions and commitments are stronger evidence.

← What an AI Product Manager DoesWhen to Use AI & Judging Feasibility →
0 / 10 finished0%