Short answer. Roughly 1 in 5 AI projects never reaches production, and a bank pays the cost of a dead project every month it stays funded. Kill an AI project when it misses two consecutive stage-gates on the metric that justified it, when the model-risk or compliance work it still needs costs more than the value it returns, or when the sponsor cannot name what breaks if the system is switched off tonight. Make the call at a defined gate against criteria set before the work started, not in the annual budget meeting.
Killing an AI project is a scheduled decision, not an emotional one. A bank that runs its AI portfolio through stage-gates decides at each gate whether the project advances, pauses, or stops. The gate compares one number, the metric that made the business case, against a threshold agreed before any code was written. When the number misses twice in a row, the default is to stop, and the burden shifts to the sponsor to argue why it should continue.
Ableneo shipped 34 production AI projects in 2025 at roughly an 80% delivery rate, so 4 of 5 reached production. That rate comes from stopping the weak projects early, before they consume a full year of engineering and validation, so the strong ones get the capacity. The discipline that produces a high delivery rate is the same discipline that kills fast.
In practice, the projects a bank should stop look like this:
In a bank, an AI project that will never reach the risk bar is not free while it waits. A high-risk system under the EU AI Act, whose obligations for high-risk use cases apply from 2 August 2026, carries a fixed load: a conformity assessment, technical documentation, human-oversight design, and post-market monitoring. Every month a doomed project stays open, it holds a slot in the model-risk validation queue and, where a third party is involved, an entry in the DORA register of information that the bank has maintained since DORA became applicable on 17 January 2025.
That scarce capacity is the real cost. Model-risk teams, compliance reviewers, and the audit evidence they produce are limited. Keeping a project alive out of sunk-cost loyalty starves the projects that could pass. Killing on a clear signal is a governance act, not a failure, and a well-run AI committee treats a documented kill the same way it treats a documented go-live.
Roughly 1 in 5 AI projects never reaches production, so a portfolio needs a way to stop the weak ones fast.
Four signals should trigger a stop review. First, the value metric misses its threshold at two consecutive gates, and the trend is flat or down, not climbing toward the bar. Second, the cost to close the remaining gap, more data, more compute, a heavier compliance build, exceeds the value the project would return in its first year live. Third, the risk profile has changed: the use case has drifted into high-risk territory the bank is not staffed to govern, or the data source has become non-compliant. Fourth, the sponsor cannot answer what the business loses if the system is turned off tonight. When the honest answer is “not much,” the project is already dead and only the budget line is still breathing.
Kill criteria written at the end are an argument. Kill criteria written at the start are a contract. Before funding a project, name the single metric that defines success, the threshold it must clear, and the gate dates where it will be checked. Attach a kill clause to every tranche of budget, so the next block of funding is a decision the sponsor has to earn, not a default the project inherits. Fund AI in stages, release each stage against the last gate’s result, and require a written post-mortem when a stage fails. This turns “should we keep going” from a debate into a check against numbers everyone agreed to in advance.
The kill decision sits with the AI steering committee or its equivalent, informed by three inputs the committee cannot overrule lightly. The business sponsor states the value the project has delivered against the case. Model risk states whether the system can meet the validation and monitoring bar. Finance states whether the next tranche is defensible against the return. In a bank, no single engineer or product owner should carry the call alone, because the decision touches risk appetite, regulatory exposure, and capital. The committee decides, records the rationale, and the record becomes evidence the next audit will value.
Stopping a project is not deleting its work. Most of what a dead AI project produced is reusable: the cleaned data, the labelled examples, the pipeline, and the specific reason the model failed. Run a short post-mortem that captures why the project missed, what data or capability was absent, and which assumption in the business case proved wrong. Feed that into the next project’s gates. A bank that kills ten projects and learns from each builds a sharper use-case filter than one that finishes ten mediocre ones. The learning is the asset the kill preserves.
Across 34 production AI projects in 2025, spanning 20 clients in financial services, insurance, telecom, and energy, Ableneo’s delivery rate held near 80% because weak projects were stopped early and strong ones were driven to production. Repeat clients run an average of 1.7 projects with the team, which is what happens when a partner is willing to say stop before a budget is wasted. Our AI transformation work builds the stage-gates, the kill criteria, and the model-risk evidence into the program from the first tranche, so the kill decision is a number the board can see rather than a fight it has to have.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.