Skip to main content

AI feature documentation plan

info

Method demonstration, not a running engagement

NimbusWiz is fictional. The plan below is how I'd structure cross-functional doc work for an AI feature, illustrated against the Modernization Advisor. The teams, deliverables, and open questions are real moves I'd want owned in a real engagement; the specifics are illustrative.

AI features create documentation input obligations from different teams. This method treats AI feature docs as cross-functional work by default.

The Modernization Advisor (NimbusWiz's upgrade recommendation feature) is a useful illustration. It produces a confidence score, names risks, and recommends one of three actions. Each property triggers a different team's input.

The four lanes of input

Each team owns something the doc team can't write without.

LaneWhat this team ownsWhat docs needs from themWhere it lands
ML EngineeringWhat the model actually doesInputs, confidence-score definition, known failure modes, retraining cadenceConfidence-score explainer, KB articles, in-product Explanation panel
Trust & SafetyWhat can and can't be claimedLimitation language, explainability requirements, trust-boundary disclosuresLimitation copy on the feature page, in-product disclosures
ProductThe intended positioningNaming ("AI-powered" vs "data-driven"), Replace-option emphasis, taxonomy decisionsFeature introductions throughout the docs
MarketingThe external narrativeFeature-page copy, release notes, terminology choicesRelease notes, terminology alignment with the docs

The point of naming the lanes is to surface critical decisions early; everything else compounds from decisions that could block publication.

The hardest question, by example

The current Modernization Advisor docs say the confidence score "reflects how much data the Modernization Advisor has to work with." That sentence is what a doc team writes when ML Engineering hasn't yet committed to a specific definition.

It's also wrong, or at least incomplete. Two questions decide whether it can ship:

  • Is confidence purely a function of data volume, or does it also reflect model uncertainty?
  • Can two systems with identical data quantities produce different confidence scores? If yes, what drives the difference?

The answers change how Technical Managers interpret the score before they act on it. If the current explanation is incomplete, the docs are creating mental models that lead to poor upgrade decisions. The cost of asking ML Engineering to commit to a definition is small. The cost of shipping the wrong sentence to a hero feature is large.

This is the discipline the brief is for: get ML Engineering to write the load-bearing sentence with you, before you publish.

What I'd ship and what I'd hold

DeliverableOwnerHold for
Modernization Advisor user-guide pageDoc leadNone: ships as soon as ML Eng commits to the score definition
Advisor FAQ (KB)Doc leadNone
Confidence-score explainer (KB)Doc leadML Eng definition
Limitation disclosure copyDoc lead + T&ST&S sign-off
In-product Explanation panel copyDoc lead + ProductML Eng input on what the panel actually surfaces
Release notes entryDoc lead + MarketingApproved terminology

The hold-for column matters more than deadlines. A deliverable with no hold-for ships when it's drafted. A deliverable with a hold-for ships when the dependency clears, and not before.

Method observations

The coordination work is the documentation work: The artifacts on this page are outputs. The conversations across the four lanes are the substance. A doc lead who treats AI feature docs as a writing problem will under-resource the part of the work that actually changes the outcome.

Limitation language is the load-bearing copy: Most pages on a docs site have copy where wrongness is annoying. AI feature pages have copy where wrongness is harmful. The limitation paragraph deserves the same review intensity as legal copy, even if it isn't legal copy.

Naming is a taxonomy decision, not a wording decision: If "confidence score" is renamed mid-engagement, the change propagates through every page that mentions it, every API field, and every analytics event. Get Product to commit to the name early, or hold the work that depends on it.