An AI documentation surface for capturing assumptions, evidence, confidence, and open questions.
- Year
- 2026
- Focus
- Inspectable AI documentation
- Role
- Product design, AI documentation patterns, interaction design, prototyping
- Themes
- AI documentation, Explainability, Knowledge work, Design systems
How can people understand what an AI-supported recommendation was based on later?
Case specimen
Inspectable AI documentation
- Messy inputs
- Model output · Source evidence · Assumptions · Reviewer notes
- Working frame
- Treat an AI output as a working note with evidence, confidence, assumptions, and unresolved questions attached.
- Decision enabled
- Help reviewers decide whether to accept, revise, investigate, or reject a generated recommendation.
- What changed
- The pattern reframed documentation as a review surface for judgement, not a passive record of generated text.
The situation
Context
As AI tools become part of everyday product and research work, teams need ways to capture how a recommendation was formed and where uncertainty remains.
Problem
Generated outputs can appear complete while hiding assumptions, weak evidence, or missing human context. That makes later review and accountability difficult.
The work
My role
I explored the information model, designed note structures, and prototyped interaction patterns for reviewing and annotating AI-supported work.
What made it hard
The challenge was making the documentation useful without turning every generated output into a heavy compliance artefact.
Process
I broke the output into layers: claim, source evidence, confidence, assumptions, reviewer comment, open question, and next action.
Key design decisions
The design used lightweight structured notes, visible confidence language, and clear prompts for human review rather than a long freeform audit trail.
What it changed
Outcome
The concept made it easier to discuss what a model produced, why it might be useful, and what still needed human judgement.
What I learned
AI documentation works best when it is close to the workflow. If it lives elsewhere, it becomes a chore instead of a thinking aid.
What I would do differently
I would explore how the pattern scales from individual notes to a searchable project memory without losing precision.