An AI initiative should not continue simply because it has already consumed time and budget. The decision is about the next commitment: whether the expected value of more work still justifies its cost and risk compared with stopping, narrowing the scope, or trying a different approach.
This article proposes a decision method, not a validated scoring model. Set the evidence and decision owner before the next funding period, then use them to decide what happens next.
Define the decision before the next commitment
Write down the outcome the initiative was funded to produce, who uses that outcome, and how the team will observe it. Separate the intended business result from activity measures such as prompts written, integrations completed, or requests processed. Activity can show progress, but it does not by itself show that the initiative is producing the intended value.
Choose evidence that fits the outcome. That may combine a baseline, operational measures, review of representative outputs, user feedback, and risk signals. Record what is not measured and what evidence would be too weak to support the decision. The NIST AI Risk Management Framework treats measurement and risk management as ongoing functions; it is voluntary guidance, not a stop rule. The decision method here adapts that evidence-first idea to a continue, redesign, or stop choice.
Set a review window that fits the risk
At approval, name the next review date or event, the evidence window, and the person accountable for the decision. Set the window according to how quickly the relevant outcome can be measured and how costly or risky it is to keep operating without a decision. A high-impact workflow may need earlier review and interim safeguards; an outcome that changes slowly may require a longer observation period. There is no universal number of weeks or review cycles that fits both.
The initiative lead should provide the evidence and explain gaps. The accountable decision maker should be able to stop or redirect the work and should not rely only on the sponsor's assessment. Invite the people who own the outcome, operations, and affected users to challenge whether the evidence answers the funded question. If the evidence is inconclusive, authorize one bounded follow-up test with a named owner, a decision date, a measurable result, and a limit on additional exposure or spend.
Compare future options, not past spend
Money and effort already spent cannot be recovered by continuing. Exclude sunk cost from the forward comparison. Include obligations that still affect the choice: shutdown or migration work, contractual commitments, transition risks, and assets that can actually be reused.
Compare the next increment of work against the best available alternatives: continue as planned, reduce scope, redesign a specific part, use a different approach, or stop. For each option, estimate remaining cost and time, expected outcome, operational risk, and what must be true for it to work. Use ranges and identify assumptions when estimates are uncertain. The purpose is to expose the choice, not to make uncertain estimates look precise.
Use evidence to choose the next action
An illustrative example: a team funded an assistant to reduce manual handling while preserving answer quality. At review, compare the agreed baseline with results from a representative set of cases and the quality checks set before the test. If both the handling outcome and quality guardrail meet their agreed conditions, continuing may be justified. If the outcome misses but a specific retrieval or workflow defect could plausibly explain the gap, redesign only that part and test it within a bounded period. If the outcome still misses, the guardrail fails, or the remaining cost exceeds the best alternative, stop or reduce the initiative.
These are proposed tests for that example, not universal thresholds. Set acceptable results, sample selection, evidence quality, and review timing before observing the result. If evidence is noisy or incomplete, record the uncertainty and decide whether a limited test can resolve it at acceptable risk. Do not turn a few weak signals into a score whose sum pretends to decide for you.
Record the decision and what can be reused
The decision record can be short: the funded outcome, evidence reviewed, remaining uncertainty, options compared, accountable decision maker, chosen action, and next review condition. If work stops, record which components are safe and useful to retain, such as a tested evaluation set or reusable integration, and who owns them. Do not assume that an asset is reusable until its dependencies, maintenance, and data permissions are understood.
The AI initiative portfolio review guide covers comparisons across multiple initiatives. For a financial framing of a data project's expected return, see Calculating Data Project ROI. Both are supporting context; neither replaces evidence about the initiative under review.
A sound stop decision is not proof that the original investment was right or wrong. It is a decision about the evidence available now, the remaining options, and the cost of waiting for more evidence.