How to decide which AI investments to stop

The organizational infrastructure around AI investment is heavily weighted toward the beginning of projects. There are processes for proposing AI initiatives, frameworks for evaluating them, approval gates for funding them, and launch ceremonies to announce them. For ongoing projects, there are status reports and quarterly reviews.

There is almost no infrastructure for stopping AI investments that are not working.

The asymmetry is not unique to AI. Organizations generally find it easier to start things than to stop them. But AI has a specific vulnerability to this pattern because the failure modes of AI projects are subtle enough to obscure underperformance for extended periods. A customer support AI that handles 40% of tickets but handles them poorly looks like a success on one metric and a significant problem on another. An AI-assisted forecasting system that produces confident but inaccurate forecasts continues to be used long after its inaccuracy should have triggered a reassessment, because the alternative, returning to manual forecasting, feels like a step backward.

The result is an AI portfolio that expands steadily with new initiatives while underperforming ones are maintained indefinitely, never explicitly stopped but never improved enough to justify their cost.

Why AI projects are hard to stop

Several factors make AI projects more persistent than their performance warrants.

Sunk cost attachment, amplified. AI projects often require significant upfront investment in data infrastructure, model development, and integration work. By the time underperformance is clear, the sunk costs are large, and the organizational psychology around large investments makes it difficult to stop. The argument “we have already spent so much” is not rational, but it is powerful, and AI’s large upfront costs make it more powerful than for smaller investments.

Ambiguous performance signals. Traditional software either works or it does not. AI systems produce outputs that are sometimes correct and sometimes not, and the distribution of correct to incorrect is often hard to measure cleanly. An AI system with mediocre overall performance might have pockets of genuine value that advocates point to when stopping is proposed. The argument “it works for some cases” is always available for AI systems, because some cases are almost always handled well.

Improvement promises. AI systems can often be improved with more data, better prompts, or model updates. When stopping an AI investment is proposed, the response is frequently “if we just had more time and resources to improve it, it would perform better.” This argument may be true, which makes it hard to refute. It may also be a way to defer a difficult decision indefinitely.

Lack of explicit stopping criteria. Most AI investments are launched with success criteria but not with stopping criteria. The question “under what conditions would we stop this investment?” is rarely asked at the beginning, which means there is no agreed framework for evaluating whether those conditions have been met. Without stopping criteria, every review becomes a negotiation about whether performance is good enough, with no baseline to reference.

Building a stopping framework

The organizations that manage AI portfolios effectively address the stopping problem by building explicit frameworks before they need them.

Define stopping criteria at launch. Before an AI investment is funded, specify the conditions under which it would be stopped. What performance level, sustained over what time period, would trigger a stop decision? What would constitute evidence that the fundamental approach is not viable? These criteria do not need to be rigid; they can be revisited as the project develops. But the practice of asking “when would we stop this?” forces the kind of clear thinking about expected outcomes that makes subsequent evaluation tractable.

Distinguish between underperformance and fundamental unviability. Some AI investments underperform because they need improvement: better data, more refined models, clearer product integration. These are worth investing in if the improvement path is clear and the expected performance after improvement justifies the additional investment. Others underperform because the fundamental premise is wrong: the AI approach is not well-suited to the problem, the data required to make the system work is not available, or the product integration was designed in a way that limits AI’s usefulness. These are not worth improving; they are worth stopping. Distinguishing between the two requires honest assessment of why performance is where it is, not just whether it is below target.

Apply the new-project test. A useful exercise: if this AI investment did not already exist and someone proposed it today, with the performance data you now have, would you fund it? If the answer is no, the sunk cost logic is the only reason to continue, which is not a sound basis for resource allocation. This reframe does not make the stopping decision easy, but it clarifies whether continuing is actually justified or merely comfortable.

Evaluate opportunity cost explicitly. Resources allocated to underperforming AI investments are not available for better-performing alternatives. Making this explicit, quantifying the cost of continuing relative to the expected return from reallocating those resources, converts a vague sense that something is underperforming into a concrete business case for stopping. Organizations that evaluate opportunity cost explicitly make stopping decisions more consistently than those that evaluate projects in isolation.

The portfolio view

Individual project evaluations, no matter how rigorous, do not substitute for a portfolio view. A portfolio of AI investments that includes too many underperformers does not just waste the resources directly consumed by those investments. It also diffuses organizational attention, creates confusion about what AI is capable of (as underperforming systems generate skepticism that spreads beyond the specific project), and prevents the organization from concentrating resources on the investments that could produce real returns.

A portfolio review, conducted at least annually, should explicitly address the question of which investments to stop, not just which to grow and which to sustain. The review should include stopping decisions as a normal output alongside scaling decisions and new investment decisions.

The stopping decisions that come out of a portfolio review are different from project-level stopping decisions. They can address portfolio-level patterns: too many investments in a particular category, too many early-stage investments without enough resources to develop any of them fully, too many projects supported by teams without adequate AI expertise. These patterns are not visible at the project level but are clear at the portfolio level and require portfolio-level action.

Making the stop decision stick

Stopping an AI investment that has vocal internal advocates is politically difficult. The advocates may be the same people who proposed the investment, which means stopping it implies their judgment was wrong. They may have users who depend on the system and will complain if it is discontinued. They may have reasons to believe the system could be improved.

The organizations that make stopping decisions stick have several practices in common.

They make the decision explicitly and accountably, with a named decision-maker, rather than through gradual defunding that nobody formally owns. This prevents the investment from being quietly continued through informal channels.

They communicate the stopping rationale clearly: what the stopping criteria were, why the project failed to meet them, and what will be done differently in future investments. This converts the stopping decision from a failure into an organizational learning moment.

They give users of the stopped system a clear transition path: how they will accomplish the tasks the AI system was performing, what the timeline for the transition is, and who is responsible for supporting them through it.

None of this makes stopping decisions comfortable. They require acknowledging that a bet did not pay off, managing the disappointment of advocates and users, and absorbing the sunk cost. But a portfolio that never stops anything accumulates exactly the wrong investments over time: the ones that have been around long enough to build constituencies, regardless of whether they produce value. The ability to stop investments that are not working is a prerequisite for a portfolio that does.