When to Kill an AI Project: Criteria, Signals, and Salvage

Every AI portfolio has projects that should have stopped a year ago. They continue because stopping was never defined, because sponsors have careers attached, and because the decision arrives in a crisis rather than at a checkpoint. Organizations that stop projects well set kill criteria at approval, review them at scheduled checkpoints, and treat stopping as a portfolio decision that frees money and people for something better. This guide covers the criteria, the signals, the process, and the salvage, drawing on FISTA Solutions' AI enablement practice. The portfolio discipline is in ai portfolio management and the checkpoint structure in ai business case template.

What kill criteria should be set at approval?

Criterion	Example	Checkpoint
Evaluation thresholds	Task success by category must reach thresholds on the golden dataset	End of discovery; end of pilot
Baseline size	Measured baseline must exceed a minimum to justify investment	End of discovery
Data access	Sample data with permissions in hand by a date	Discovery
Cost per outcome	Below a ceiling at realistic adoption with optimization applied	Pilot; production
Adoption	Share of eligible tasks through the system after change management	Production
External change	Vendor, regulatory, or strategic change that removes the case	Any

Criteria are written when nobody has anything to defend, which is the only time they can be written truthfully. Scoring inputs are in the ai use case scoring framework.

What signals that a project should stop?

Thresholds still unreachable after narrowing scope and trying alternative approaches; discovery revealing a baseline too small to matter; data that remains inaccessible or unpermitted after escalation; cost per outcome above ceiling with routing, caching, and model changes exhausted; users declining to adopt after real change management; and external changes that remove the business case. Each is a criterion missed, not an opinion. Failure patterns are in why ai agents fail in production.

Why is narrowing often the better decision?

Evaluation frequently shows some categories reachable and others not. Narrowing to the reachable scope with a revised business case delivers value and keeps the platform work; killing discards both. Narrow when the reachable scope still justifies run cost and maintenance; kill when it does not. The decision is made with the same evidence. Evaluation practice is in what is a golden dataset.

How do you decide without politics?

Decide at scheduled checkpoints against the written criteria, with evidence distributed before the meeting from the same dashboards, and with the steering committee confirming a decision the criteria already made rather than debating one. Sponsors defend projects when the decision is personal; criteria set at approval make it procedural. Committee mechanics are in how to run an ai steering committee.

What should be salvaged?

Golden datasets and labels, which are expensive and reusable; integrations and tool contracts; platform components such as retrieval pipelines and gateway configuration; data readiness work and permissions; documentation; and the lessons about what was unreachable and why. Record them in the portfolio so the next initiative starts further along. Stopping well is a deliverable. Pilot design that produces salvageable assets is in the ai pilot checklist.

How do you communicate a stop?

As a portfolio decision made against criteria, with what was learned and what was salvaged, to the sponsor first, then the team, then leadership. Credit the team for producing the evidence that enabled the decision; the failure was in the hypothesis, not the work. Organizations that punish stops get projects that never end. Board framing is in how to report ai progress to the board.

What does the cost of not stopping look like?

Run cost and maintenance for a system nobody adopts; platform and governance capacity consumed; engineers retained on work without outcomes; credibility spent when leadership eventually asks; and the opportunity cost of the initiative that was not funded. Stopping early is cheap; stopping late is the same decision at a higher price. Risk tracking is in ai risk register and controls in how to de-risk an ai project.

What mistakes keep failing projects alive?

No criteria at approval; checkpoints without decisions; evidence presented as slides rather than dashboards; sponsors as sole judges; funding released in full so nothing is at stake; and a culture where stopping is failure. Each is a governance design flaw with a known fix.

What does a well-handled stop look like in practice?

A financial services firm approves a document summarization project with thresholds by document type, a cost ceiling, and a pilot checkpoint date. Evaluation reaches thresholds on three types and misses two after two rounds of narrowing; the reachable scope is too small to justify run cost against the baseline. The steering committee confirms the stop at the checkpoint, the golden dataset and extraction pipeline move to a contract review initiative, and the sponsor presents the lessons. The next proposal from the same team is approved faster because the evidence discipline was visible.

How FISTA Solutions helps with stop decisions

FISTA Solutions writes kill criteria into every specification and pilot agreement, delivers the evaluation evidence checkpoints require, recommends narrow or stop when the evidence says so, and salvages datasets, integrations, and platform work into the next initiative. The AI enablement practice leads portfolio and evaluation, forward deployed engineers deliver, and AI agents supplies the systems under review. The record behind the approach is 150+ projects for 50+ companies.

To set criteria that make stopping a decision rather than a defeat, message FISTA on WhatsApp, or read ai portfolio management for the discipline around it.