Complete one record per initiative before its next funding decision. Bring the business sponsor, technical evidence owner, finance partner, and risk/service owners. Use the success criteria agreed before the pilot; if they changed, explain why and show the original result too.
Initiative / sponsor / evidence owner / decision date:
Business outcome / users / non-AI alternative:
Establish the evidence
For each item, record met, not met, or unknown, plus the evidence reference and date. Unknown is a gap, not a passing score. These gates are not a weighted total; unacceptable risk cannot be offset by adoption.
- Outcome: What changed relative to the baseline? Count accepted outcomes at the agreed quality level, not generated outputs. Include rework, human review, and affected users’ experience.
- Economics: What is the next stage’s fully loaded cost? Include committed platform cost, consumption, integration, evaluation, review, security, support, and workforce change. Compare with the alternative on the same basis.
- Risk: Are data use, authority, failure consequences, and residual risk within the agreed boundary? Identify any unresolved release blocker and who can accept an exception.
- Operations: Who will run the service? Show monitoring, support coverage, incident response, fallback, and retirement responsibilities, with evidence that the important controls work.
- Uncertainty: Which assumption could reverse the decision? State the test that would resolve it and the resources it needs. Do not disguise a new use case as a repaired old one.
Evidence met / not met / unknown, with references:
Next-stage cost / accepted outcomes / key assumption:
Release blocker or accepted risk / accountable owner:
Choose one outcome
Scale
Choose scale when the outcome evidence, full cost, risk boundary, and operating ownership justify the next increment. Record the approved scope and budget, the service owner, and the condition that stops further expansion. Passing a pilot does not authorize unlimited rollout.
Repair
Choose repair when one consequential uncertainty remains testable. Record one hypothesis, one owner, a bounded change, a spending limit, and a new review date. Specify in advance what result leads to scale or stop. “Keep exploring” is not a repair contract.
Stop
Choose stop when the need has changed, evidence misses the threshold, economics fail, risk is unacceptable, ownership is absent, or a simpler option is better. Assign retirement of credentials, data copies, integrations, infrastructure, and vendor commitments. Verify completion; a slide announcing a stop does not remove access.
Scale / repair / stop — decision and reason:
Approved scope or repair hypothesis / owner / budget:
Next decision date / stop trigger / required evidence:
Retirement tasks / owner / completion evidence:
Preserve the learning
Record disproved assumptions and reusable work. Separate future spending avoided from money already spent. Report staff capacity released without calling it cash savings unless there is evidence of how it was used. Assign that capacity to the next priority explicitly.
Underlying guidance
Read The AI Portfolio Review Needs a Kill List for the decision framework, Planning the AI Budget for funding buckets, and Enterprise AI Economics for cost per accepted outcome. The articles explain tradeoffs and provide their primary references.