Key takeaways
- Most enterprises have formal entry criteria for AI investments and almost no exit criteria. This asymmetry is not an oversight — it is a structural incentive problem.
- Three failure patterns account for the majority of AI initiatives that should be stopped but are not: the sunk-cost trap, the narrative substitution trap, and the shared-platform alibi.
- "Needs more time" is the most expensive phrase in enterprise AI economics. The inability to distinguish governance immaturity from structural value failure costs organisations more than most overspend problems.
- A workable stop or redesign discipline requires pre-agreed exit thresholds, named owners, and separation of the investment case from the delivery team's professional stake in it.
The asymmetry nobody talks about
Every serious AI investment in a large enterprise goes through some form of business case approval. The CFO asks about ROI. The CIO signs off on architecture. Procurement reviews vendor terms. A stage gate may exist before a pilot becomes a production deployment.
What happens when that initiative fails to deliver? In most organisations, the honest answer is: not very much.
This is not primarily a measurement problem. It is an incentive problem. The people who sponsor AI investments are the same people who evaluate whether those investments are working. The teams that build AI systems have professional reputations tied to their success. The governance structures designed to challenge poor-performing investments are the same ones that approved them in the first place. Stopping an AI initiative requires someone in authority to say, in effect, that a prior decision was wrong. Very few governance models create real incentive to do that.
The result is an enterprise AI portfolio that accumulates rather than curates. Organisations get better at adding AI investments than removing them. The good cases are diluted by the persistent ones.
What the economics of inaction actually look like
Consider a realistic pattern from a large financial services group. Over 18 months, the organisation approved 34 AI initiatives across customer service, risk, compliance, and back-office operations. By month 18, four had been formally concluded — two as successes, two as completed pilots that were not scaled. The remaining 30 were all described as "in progress," "building toward value," or "awaiting improved data quality."
A conservative estimate of the annual operating cost of those 30 initiatives — including allocated platform time, engineering support, governance overhead, and model consumption — is between £1.8M and £2.4M per year, excluding the original development investment. The question is not whether any of those 30 initiatives might eventually succeed. The question is whether they all deserve continued capital at the same rate, indefinitely, without an explicit decision point.
At Level 3 maturity and below — which is where most enterprises currently sit — the answer is effectively yes, because there is no portfolio mechanism for comparative challenge. Each initiative defends itself against its own original narrative rather than against competing uses of the same budget.
This is the cost of missing exit governance. It is not dramatic. There is no single large failure event. It is the quiet accumulation of small bad decisions sustained past the point where evidence would justify them.
Three failure patterns that survive because they look like something else
The sunk-cost trap
An AI initiative has consumed significant investment in platform setup, data preparation, and integration work. The workflow is live but under-adopted, and early performance data is mixed. The business case assumed 40% productivity improvement in a loan-processing function; actual improvement is running at around 8-12% and appears to be plateauing.
In most governance reviews, this case survives because the sunken cost — the platform work, the integration — is treated as a reason to continue rather than as a measurement of what has already been spent. The argument becomes: "We've invested too much to stop now." This is a logical fallacy, but it is also an institutionally convenient one because it avoids the decision.
The correct question is not what has been spent. It is what additional capital, at what probability of reaching the original value threshold, would be required to continue — and whether that expected value exceeds the next-best use of the same resource.
The narrative substitution trap
A generative AI implementation was deployed to improve the quality and speed of analyst report preparation. Adoption is measurable but the original economic case — 30% reduction in preparation time per report — has not been demonstrated after nine months. In response, the business sponsor has shifted the narrative. The initiative is now described as improving "output quality and consistency," "analyst capability development," and "competitive differentiation."
These may be real benefits. But none of them were in the original business case, none are being measured, and none of them justify the original investment level. The narrative has been substituted to make the initiative look successful without actually demonstrating the return that was promised.
Narrative substitution is the most socially difficult failure pattern to challenge because it requires someone in the governance chain to say explicitly: "That is not what we agreed to measure." This puts the governance function in confrontation with the business sponsor — which most governance functions are designed to avoid.
The shared-platform alibi
A customer-facing AI workflow has weak adoption and unclear ROI. The team argues that the value should not be evaluated at the use-case level because the initiative helped establish shared platform capabilities, common integration patterns, and reusable governance infrastructure. These platform benefits, the argument goes, will support multiple future use cases.
This argument is sometimes valid. Shared platform investment does create option value. But it is also the most durable alibi for underperforming initiatives, precisely because platform value is diffuse and hard to attribute cleanly. When every underperforming use case claims credit for platform-building, the portfolio governance function cannot assess where platform investment is genuinely creating reuse and where it is creating an excuse.
Distinguishing governance immaturity from structural value failure
The most important judgment call in AI portfolio management is whether an underperforming initiative is failing because of governance immaturity — weak measurement, poor adoption infrastructure, insufficient baseline — or because the value case is structurally weak.
These are different problems with different correct responses. Governance immaturity might be worth investing through. Structural failure should be stopped.
Four signals point toward structural failure rather than governance immaturity:
1. The value mechanism has not changed despite repeated redesign. If the initiative has been redesigned, reframed, or re-baselined more than once without improving the core value-creation signal, this is more consistent with structural weakness than with governance immaturity. Governance problems respond to governance investment. Structural problems do not.
2. The theoretical value has been demonstrated but not absorbed. Some AI systems demonstrably improve the speed or accuracy of individual tasks, but the business never realises the productivity gain because workflow, headcount, or decision-making structures did not change around the AI capability. This is not a measurement failure. It is a value-capture failure. The technology worked; the operating model did not adapt. This pattern rarely improves without a fundamental redesign of the surrounding process — which is a different and more expensive investment than continuing to fund the AI capability itself.
3. External conditions that justified the case have changed. Model costs, competitive dynamics, regulatory requirements, or organisational priorities may have shifted materially since the original investment decision. An AI use case that was defensible at 2024 inference prices and 2024 compliance requirements may not be defensible at current costs with current obligations. The original business case is not a perpetual justification.
4. Value ownership is genuinely contested. If the business unit that was supposed to realise value has stopped actively supporting the initiative — delegating it to IT, reducing allocated personnel, or simply deprioritising the workflow change that was supposed to capture the AI benefit — this is a practical withdrawal of the value commitment. Continuing to fund a capability that the consuming organisation has effectively abandoned is governance failure.
The political economy of stopping
Understanding why organisations fail to stop AI initiatives requires understanding who loses from the decision.
The delivery team loses programme credibility and professional capital. The business sponsor loses budget they may not recover. The vendor relationship may be affected. The IT platform team may have to justify why shared infrastructure was built for a use case that was abandoned. The CAIO or AI programme office may be seen as having approved a failure.
These are all real costs, distributed across real people. The benefits of stopping — freed capital, clearer portfolio comparisons, honest evidence base — are diffuse and accrue to the organisation, not to any individual who bears the cost of the decision.
This is why stop decisions in AI portfolios are systematically underprovided. The incentive structure produces continuation bias. Fixing this requires either changing the incentives — for example, rewarding early stopping as evidence of governance discipline rather than treating it as a failure — or separating the evaluation function from the investment function so that the people who judge value are not the same as the people who built the case.
What a workable exit discipline looks like
Exit governance does not require a separate bureaucratic process for every initiative. It requires three things that most organisations do not have.
Pre-agreed exit thresholds. At the point of investment approval, define the conditions under which the initiative will be reviewed for stop or redesign. These should be specific: a minimum adoption rate by a given date, a required delta on a named baseline metric, a maximum cost-per-outcome threshold. Generic conditions like "if performance is unsatisfactory" are useless because they leave the stop decision entirely to subjective judgment.
Named economic owners with real authority. Value ownership must be separated from delivery ownership. The person who owns the economic outcome must have the authority to say the outcome is not being achieved, and the institutional support to act on that judgment without sacrificing their position. If economic owners are politically unable to stop underperforming investments, the ownership is nominal.
Portfolio-level comparison rather than initiative-level self-assessment. The weakest initiatives look most defensible when evaluated only against their own original narratives. Portfolio-level review compares what a continuation decision costs against what the same capital would produce deployed elsewhere. This comparison is the only one that has real discipline because it makes the opportunity cost explicit.
The questions a board or investment committee should be asking
A governance body with serious accountability for AI economics should be able to answer these questions for every material investment in the portfolio:
- What was the original return dimension and threshold, and has it been revised since approval?
- What is the current evidence on that return dimension, referenced against the original baseline?
- Who currently owns the economic outcome, and is that person actively engaged in achieving it?
- What would the initiative need to demonstrate by the next review cycle to avoid a redesign or stop recommendation?
- What is the total forward cost of continuing this initiative for another 12 months, and what is the expected value at that cost?
If the answer to any of these is "we don't know," the initiative has already left the domain of governance and entered the domain of hope. Hope is not a portfolio strategy.
A practical note on timing
The right time to establish exit criteria is before an initiative is approved. Pre-agreed thresholds are less politically fraught than retrospective ones because no specific investment has yet failed to meet them.
Organisations that try to install exit governance only after a portfolio of underperforming investments has accumulated will find the process much harder. The political costs of stopping multiple simultaneous investments are higher than the cost of an ongoing systematic process. Starting with a clean threshold-setting exercise at the point of new investment approvals, while separately conducting a structured portfolio review of existing investments, is a more viable sequencing.
The review of existing investments will still be difficult. Some of them should stop. That conversation will be uncomfortable. It will also be one of the most economically valuable governance actions the organisation can take.