Product & UX

Designing for Non-Deterministic Outputs

Traditional software is a promise: this button always does this thing. AI is a probability. Good AI interfaces set that expectation honestly — they present output as a draft, make it trivially easy to edit or regenerate, show confidence or sources where possible, and never present a guess with the visual authority of a fact. Users forgive a wrong suggestion; they do not forgive a wrong answer that was dressed up as certain.

Evaluation and Testing

You can't unit-test a language model with exact string matching. Instead you build an evaluation set: a fixed collection of real inputs paired with what a good answer looks like, run after every prompt or model change to catch regressions. Track a small number of honest metrics — accuracy on that set, escalation rate, edit rate, time saved — and re-run them continuously, because model providers update the underlying models beneath you.

• Keep a fixed evaluation set of real, representative cases
• Re-run it on every prompt, model, or data change
• Track edit rate — how often humans fix the output — as your quality signal

Human-in-the-Loop Patterns

The question is never whether a human is involved, but where. Review-before-send suits customer-facing output. Approve-the-exception suits high-volume work where the model handles the routine and flags what it isn't sure about. Spot-check-a-sample suits mature, well-measured systems. Match the pattern to the cost of an error: the higher the consequence, the earlier and more mandatory the human checkpoint.