Fund, Fix or Stop: A Decision Rule for AI Initiatives

Every AI initiative deserves exactly one of three verdicts: fund, fix, or stop — and the verdict comes from three facts, not enthusiasm: readiness score, a defensible case, and monthly proof. This page gives the decision table and the evidence each verdict requires.

The decision table

You have...	Verdict	What happens next
Readiness gaps AND no case yet	FIX	Fix the weakest factor first (data, people, approval, sponsor); re-score before any spend
Readiness OK + case signed + no proof yet	FUND	Commit, with the sponsor and the monthly check date named in the same breath
Funded + 3 months of monthly proof	FUND (continue)	Continue while the line moves inside the normal range or improves
Funded + proof outside normal range for 2+ months	FIX or STOP	Diagnose: workflow problem = fix; value problem = stop
No sponsor + no baseline + pressure to buy	STOP	Buying more AI does not move the number (/pillar/ai-pays-off, live 11 Oct 2026)
Case shows payback beyond the horizon in the conservative scenario	STOP or RESCOPE	The property case on /examples shows a first-year loss of RM30,300 — the model said stop, the page says stop (/examples, fetched 11 Oct 2026)

The three facts each verdict rests on

1. Readiness (0-100): ten questions score data, people, approval path and sponsor; the weakest factor is named first (Scan's readiness map, skillpilotadvisory.ai/scan, fetched 10 Oct 2026).
2. The case: hours x loaded cost, three scenarios, payback period, worst case stated plainly — in the approver's format, before the ask.
3. The proof: two consecutive months of CSV judged monthly against a normal range you defined in advance; one sentence, one accountable name.

Why a rule, not a feeling

• Most AI decisions are made twice: when the money is committed, and every month after, when somebody asks what it returned (skillpilotadvisory.ai/scan, fetched 10 Oct 2026). The rule exists for the second decision, which most teams never prepare for.
• The failure pattern is documented: spend rises, pilots multiply, the numbers barely move (skillpilotadvisory.ai/about, live 11 Oct 2026).
• The model publishes negative outcomes on purpose — the telco case shows first-year net of only RM1,485 with payback at 11.4 months; the healthcare case shows zero cash benefit and the line "do not claim cost savings" (/examples, fetched 11 Oct 2026). A rule that only ever says fund is not a rule.
• Governance context: MOSTI's National Guidelines on AI Governance and Ethics (Nov 2024) expect accountability and human oversight — a fund/fix/stop rule with a named sponsor is the operational form of both.

Where the human fits

Self-serve: Scan runs readiness, case and monthly proof in one login, 365 days. Human-driven: the advisory founding sprint runs two weeks on one initiative and ends in exactly this decision — fund, fix, or stop (skillpilotadvisory.ai/advisory, fetched 10 Oct 2026).

FAQ

Who applies the rule? The named sponsor, with the readiness score and the monthly sentence as evidence.
How often is the verdict revisited? Monthly, on the proof; the check date is part of the fund decision, not an afterthought.
Is "stop" a failure? It is the rule working; stopping an unprovable spend is the cheapest possible outcome.
Does the rule apply to pilots already running? Yes — restart the proof chain from the current month.