Open resource · AI value and governance

Govern AI close enough to the work to improve it.

A practical toolkit for moving from an interesting AI possibility to an accountable product or workflow with observable value.

Download the toolkit

Why this exists

Possibility is not yet value.

AI work often begins with model access and a demonstration. Useful adoption begins somewhere else: a meaningful outcome, a person accountable for the consequences, controls shaped around real failure modes and evidence strong enough to guide investment.

The tools are intentionally lightweight. Use only the sections that can change a decision. Increase rigor as impact, uncertainty and potential harm increase.

Six connected tools

From value hypothesis to operating evidence.

Start with the value canvas. Return to the accountability map and guardrails whenever the use case, affected group, data or operating context changes.

01Working tool

AI use-case value canvas

Frame the outcome before selecting a model or workflow.

  • Whose work or decision should improve?
  • What observable outcome matters?
  • Why is AI appropriate for this constraint?
  • What is the credible non-AI alternative?
02Working tool

Human-accountability map

Make ownership visible across design, use, review and consequences.

  • Who owns the product outcome?
  • Who can approve, override or stop the system?
  • Who reviews affected-person and domain risk?
  • Who responds when the system fails?
03Working tool

Guardrail design checklist

Place controls near the decisions and failure modes they must influence.

  • Which data must not enter the model?
  • Which outputs require human review?
  • What limits, fallbacks and kill switches are needed?
  • What must be logged without exposing sensitive content?
04Working tool

POC-to-MVP decision gate

Separate technical possibility from a product worth operating.

  • Which risky assumption did the POC reduce?
  • What remains unknown about real use?
  • Can the MVP be operated safely end to end?
  • What evidence would justify further investment?
05Working tool

Outcome and evidence tracker

Measure behavior, quality, risk and economics before scale.

  • What baseline are we changing?
  • Which quality and risk signals can be observed?
  • What is the cost per useful outcome?
  • What evidence would cause us to stop or redesign?
06Working tool

Governance review agenda

Run a proportionate review that changes the work, not only the paperwork.

  • Is the value hypothesis still credible?
  • Are failure modes and affected groups understood?
  • Are controls operating as intended?
  • What decision and owner leave the review?

POC to MVP to scale

Progress only when the next commitment is justified.

A POC reduces uncertainty. An MVP tests a usable outcome. Scale is an operating commitment. They need different evidence.

GateEvidence to progressReason to pause
Problem

A real user or operating constraint is evidenced.

Do not progress from an AI-first idea.

Value

A measurable outcome and baseline are defined.

Do not use activity as a proxy for value.

Control

Ownership, review, fallback and stop conditions are clear.

Do not rely on policy alone.

Evidence

Representative evaluation covers quality, risk, cost and behavior.

Do not scale from a compelling demo.

Operation

Monitoring, incident response and change ownership are ready.

Do not treat launch as completion.

A 45-minute review

Make one decision and leave with one accountable owner.

  1. 05 min
    Restate the value hypothesis

    What should improve, for whom and against which baseline?

  2. 10 min
    Examine current evidence

    Review quality, risk, behavior, cost and operational signals.

  3. 10 min
    Test accountability

    Confirm who decides, reviews, overrides, responds and can stop the system.

  4. 10 min
    Inspect guardrails

    Check whether controls are working in the product and workflow.

  5. 10 min
    Make the next decision

    Continue, adapt, constrain or stop. Name the owner and next evidence date.

Boundaries

A working aid, not a substitute for specialist judgment.

01

Proportionate

Increase review depth as impact, uncertainty and potential harm increase.

02

Evidence-led

Mark assumptions and unknowns. Do not invent confidence where evaluation is incomplete.

03

Accountable

Keep a named person responsible for decisions and consequences throughout the lifecycle.

Printable toolkit

Use the worksheets in a product, engineering or leadership review.

Download the PDF