learn.web Back to the curriculum ↗

Context / Tools / Evals / Safety / 01

Find the right product seam

Choose work where ambiguity is useful, review is possible, and the model improves an existing outcome.

Time
60 min
Mode
Learn → Make → Check
Path
AI Product Engineering

By the end, you can…

Start with the workflow, not the model

AI is useful where interpretation, generation, classification, or transformation creates value and a person or system can evaluate the result. It is weaker where exactness is mandatory and errors are hard to detect.

Map the existing workflow, frequency, pain, available context, review point, and cost of failure. Compare the AI concept against a simpler search, rule, template, or interface improvement.

Write a value and risk contract

Define the user outcome, baseline, success measure, unacceptable failure, escalation path, and authority boundary. The model should not quietly gain more power than the user intended.

Prototype the workflow with human-generated outputs before integrating a model. If the interaction is not useful with good outputs, better model quality will not rescue it.

Your studio task

Make — Evaluate three candidate AI features and reject at least one.

  1. List three candidate AI features in an existing workflow.
  2. Score value, uncertainty tolerance, reviewability, and failure cost.
  3. Compare the strongest idea with a non-AI alternative.
  4. Prototype the workflow and reject at least one candidate explicitly.
Definition of done

The selected use case tolerates uncertainty, supports review, and improves a measurable outcome.

Open the interactive lesson with its workspace ↗

Knowledge check

Which task is the strongest AI candidate?
  1. Calculating an exact tax total with no verification
  2. Drafting a support reply that an agent reviews before sending
  3. Silently approving medical treatment
Reveal answer (B)

Drafting supports judgment while preserving human review before a consequential action.

Which AI failure is hardest for a user to detect?
  1. A fluent but wrong answer in an expert domain
  2. A blank screen
  3. An obvious grammar error
Reveal answer (A)

Plausible errors in expert domains are exactly why review and grounding matter.