Discovery: the right problem
The most important AI product decision is whether to use AI at all. This module helps you find problems where AI adds real value.
Learning objectives
- Assess whether AI adds unique value
- Decide between automation and augmentation
- Write an AI problem framing
- Prototype quickly with real examples
2.1 Start with the user problem
"Use AI" is not a user need. Discovery for AI products is the same as for any product — interviews, observation, data — with one extra question: is AI the best way to solve this? Google's People + AI Guidebook recommends asking whether AI adds unique value before committing.
2.2 When AI helps — and when it doesn't
AI is likely to help when…
- The task involves language, images or patterns humans find tedious
- Personalisation at scale matters
- Rules would be too many or too brittle to write
- Some errors are acceptable and recoverable
Prefer simpler solutions when…
- Clear rules already work well
- Users need total predictability
- Errors are costly and hard to detect
- There's no data, or no way to judge quality
2.3 Automate or augment?
Should AI do the task for the user, or help the user do it better? Users often want to keep control of tasks they enjoy, are accountable for, or where personal judgement matters; they're happy to hand off tasks that are boring, repetitive or that they lack skill for. Many products start by augmenting (suggestions a person approves) and move towards automation as trust and evidence grow.
2.4 Write the problem framing
| User & job | Who, and what they're trying to get done |
| Current pain | Time, errors, frustration — with evidence |
| AI's role | Generate, transform, classify, answer, predict or act; automate or augment |
| Inputs available | What data or context the AI will have at that moment |
| Quality bar | What "good enough" looks like, and which errors are unacceptable |
| Outcome metric | How you'll know users and the business are better off |
2.5 Prototype early with real inputs
With today's models you can test an idea in hours: write a prompt, run it on 20–50 real examples (anonymised where needed), and read every output. This "vibe check" reveals feasibility and failure modes long before engineering investment. It is not an evaluation — that comes in Module 4 — but it's the fastest way to kill weak ideas and sharpen strong ones. See Prompt Engineering.
Fill in the six-row problem framing for one AI feature idea. Decide automate or augment and justify it.
See a worked example
Example — expense categorisation: User: employees submitting expenses. Pain: 4 minutes per claim choosing categories; 12% miscategorised. AI's role: classify receipts and suggest a category (augment — the employee confirms). Inputs: receipt image, merchant, amount. Quality bar: ≥ 95% correct top suggestion; never auto-submit. Outcome: time per claim, finance correction rate.
Key takeaways
- 'Use AI' is not a user need.
- Prefer rules when they work and predictability matters.
- Augment first; automate as trust and evidence grow.
- Frame user, pain, AI's role, inputs, quality bar and outcome.
- Prototype on real inputs and read every output.