How Do You Know When to Kill an AI Project? A Portfolio Decision Framework for Enterprise Leaders

TLDR: Most enterprises cancel AI projects too late, after millions in sunk costs, or too rarely, allowing underperforming initiatives to consume resources that belong elsewhere. This post provides a structured decision framework for evaluating when to stop, pivot, or scale an AI initiative, including the red flags that signal a project cannot recover and the governance structures that make portfolio discipline possible.

Best For: COOs, CTOs, and VP Operations at mid-to-large enterprises managing a portfolio of AI initiatives and facing pressure to demonstrate ROI while avoiding the trap of funding projects that will never scale.

An AI project kill decision is a structured evaluation that determines whether an AI initiative should be stopped, fundamentally redesigned, or continued based on forward-looking evidence rather than sunk cost. Most enterprises don't have a formal process for this. They have budget cycles, steering committee reviews, and an unspoken norm of continuing investments that have already been approved. The result is a portfolio weighted toward persistence rather than performance.

Why Enterprise AI Portfolios Accumulate Underperformers

Enterprise AI portfolios accumulate underperforming projects because the organizational systems designed to evaluate AI investments were built for IT projects, not transformation initiatives. IT projects fail or succeed on technical criteria. AI initiatives fail on a different set of dimensions: misaligned success metrics, governance gaps, organizational resistance, and data readiness problems that surface only after significant investment.

According to Gartner's April 2026 survey of 782 infrastructure and operations leaders, only 28% of AI use cases fully succeed and meet ROI expectations. 20% fail outright. The remaining 52% fall into a middle category: initiatives that produce some results but stall short of the business outcomes they were funded to deliver. These middle-category projects are the portfolio management challenge. They are not clear failures (which would trigger cancellation) and they are not clear successes (which would justify expansion). They consume resources indefinitely while producing modest, hard-to-attribute results.

The Sunk Cost Trap in AI Programs

The sunk cost trap is particularly acute in AI programs because of the nature of the investments involved. Technical infrastructure, data pipelines, vendor contracts, and training time are all real costs that have been incurred before an AI system goes live. By the time operational problems emerge in production, leadership teams have approved significant investment, created visible organizational commitments, and tied personal credibility to the initiative's success.

According to Deloitte's 2025 analysis of enterprise AI initiatives, 42% of companies abandoned at least one AI initiative in the prior year, with the average sunk cost per abandoned initiative reaching $7.2 million. Critically, organizations that abandoned failing projects by month 8 rather than month 18 saved an average of $2.1 million in additional sunk costs and reallocated their resources ten months faster.

Why AI Projects Are Harder to Kill Than IT Projects

Standard IT project governance uses clear failure criteria: the system doesn't work, the vendor can't deliver, or the cost exceeds the approved budget. AI projects are harder. An AI system can produce technically correct outputs that nobody uses, or outputs that are used but have no measurable impact on the business metrics the project was supposed to improve. A pilot can look impressive in a controlled environment and fall apart when it encounters real operational data.

This ambiguity creates a governance gap. Without explicit criteria for distinguishing between a project that needs more time and a project that needs to be stopped, most organizations default to continued investment. The Harvard Business Review's 2025 analysis of AI initiative failures found that AI initiatives don't fail because the models are bad. They fail because everything underneath them is broken, and leadership approved continuation without asking hard questions.

The 5 Kill Criteria That Signal an AI Project Cannot Recover

Not all AI underperformance is terminal. Some projects need a governance fix. Some need a different dataset. Some need a real change management effort rather than a launch email. The question is not whether a project is struggling but whether the root cause of that struggle is actually fixable within a reasonable investment threshold.

Five criteria, when present in combination, indicate that an AI initiative is unlikely to recover and should be stopped or fundamentally redesigned.

1. No Named Business Owner With P&L Accountability

AI projects that cannot name a specific business leader who owns the financial outcome and has the authority to redesign processes around AI are structurally dependent on technical teams to drive adoption. Technical teams can build and deploy AI. They cannot compel operations leaders to change how their teams work. According to research on AI production failures cited in Information Week, this structural gap is the single most reliable predictor of a project that will stall before production.

2. Absence of Pre-Defined Success Metrics and Kill Thresholds

Projects that were approved without specific, measurable success criteria cannot be objectively evaluated for termination. If the original business case said "improve efficiency," there is no threshold below which the project can be declared a failure. The absence of pre-defined kill thresholds is both a governance failure and a signal that the project was funded without sufficient rigor in the business case. The standard for this criterion, drawn from McKinsey Global Institute's 2025 guidance, is a specific metric, a target value, and a date by which the project must hit that value or be reviewed.

3. Data Quality Problems That Vendors Have Confirmed Cannot Be Resolved Within Scope

Data quality is the most commonly cited technical root cause of AI project failure. Gartner reports that 85% of AI projects fail due to poor data quality or lack of relevant data, and that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. A data quality problem is not automatically a kill signal. But a data quality problem that the technical team and vendor have confirmed cannot be resolved within the project's approved scope and budget is.

4. End-User Adoption Below 30% at the 90-Day Post-Launch Mark

AI systems that real users are not using are not delivering business value regardless of their technical performance. If adoption falls below 30% of the intended user base at 90 days post-launch and there is no structured adoption recovery program in place, the project is exhibiting the behavioral symptoms of organizational rejection. This is not primarily a technology problem. It is an indication that the change management investment was insufficient, that the AI system's workflow integration was poor, or that the use case does not actually solve a pain the targeted users experience.

5. Cumulative Investment Exceeding 150% of the Original Budget With No Validated Business Outcome

Budget overruns in AI projects are common and do not by themselves justify termination. However, an initiative that has consumed 150% or more of its original approved budget without producing a single validated business outcome — documented, attributed, and verified by the business owner, not the technical team — has failed to demonstrate the most basic requirement for continued investment. This criterion is deliberately absolute. "Promising results" in test environments do not qualify. Validated business outcomes require production deployment and attributed business impact.

The Portfolio Decision Matrix: Stop, Pivot, or Scale

Rather than treating each AI initiative as an individual investment decision, enterprises that manage AI effectively treat their AI portfolio as a system. The question is not only whether an individual project is performing but how its performance compares to alternatives competing for the same resources.

The framework below provides a structured approach to the stop/pivot/scale decision based on two variables: forward-looking probability of success and the magnitude of the business outcome if the project succeeds.

"Probability of recovery" is assessed using the five kill criteria above. A project that triggers one or two of those criteria may be recoverable. A project that triggers three or more criteria has a low probability of recovery regardless of business outcome magnitude.

The "Pivot" cell deserves particular attention. Pivoting an AI initiative means fundamentally redesigning the use case, the data approach, or the organizational model, not iterating on the existing design. Organizations sometimes use "pivot" as a softer word for "we don't want to cancel this" without actually making the structural changes required. A genuine pivot requires a new business case, new success criteria, and new accountability structures.

Governance Structures That Make Portfolio Discipline Possible

Kill decisions require governance structures that make it safe to stop a project. Without those structures, the organizational incentives all point toward continuation: the project champion's credibility is at stake, the vendor relationship would be affected, and the team working on the initiative has invested months of effort.

The Two-Track Review Board

A two-track review board separates AI initiatives into two oversight categories: high-stakes projects requiring quarterly substantive review, and exploratory projects subject to an 8-week go/no-go gate before significant resource commitment. This structure, described in detail in recent analysis from Virtasant on enterprise AI portfolio management, prevents high-oversight governance overhead from slowing exploratory work while ensuring that scaled investments are reviewed against hard business criteria on a regular cadence.

The 8-Week Kill Gate for New Pilots

For new AI pilots, establishing an explicit kill gate at week 8 creates a forcing function for clear success criteria at the point of initiation. If a pilot cannot demonstrate a leading indicator of business value within 8 weeks, the default decision is termination, not extension. This standard requires pre-defining what a leading indicator looks like before the pilot begins, which in itself forces the business case discipline that most AI pilots lack.

Organizations that implement this governance structure before beginning their current portfolio of AI pilots should also review their AI pilots playbook to ensure pilot design standards support the kind of early measurement that makes 8-week gates viable.

Separating Portfolio Governance From Vendor Relationship Management

A common governance failure is allowing vendor relationship considerations to influence portfolio review decisions. If the business sponsor of an AI initiative is also the primary relationship manager for the AI vendor, their assessment of the project's recovery prospects will be contaminated by the desire to preserve the vendor relationship. Portfolio governance should be led by a leader with accountability for business outcomes, not vendor management.

For enterprises building out their formal AI governance function, understanding how to build AI governance that enables speed is a prerequisite for implementing the kind of portfolio discipline this framework requires. Governance structures that are too heavy slow everything down. Governance structures that are too light allow underperformers to survive indefinitely.

Common Objections Leaders Raise About Killing AI Projects

"We've Already Invested Too Much to Stop Now"

This is the sunk cost fallacy in its most recognizable form. The relevant question is not what you have invested but what the future investment required to reach positive ROI is, what the realistic probability of recovery is, and what alternative uses exist for the resources this project is currently consuming. According to McKinsey Global Institute's 2025 research, the most profitable AI organizations are not those that pick the best models. They are those that kill bad AI projects fastest and reallocate to higher-probability opportunities.

"The Technology Just Needs More Time to Mature"

This framing mistakes a technology problem for what is usually an organizational one. The Gartner finding that 28% of AI projects deliver ROI reflects failures of governance, data readiness, and organizational adoption, not failures of the AI technology itself. An organization that frames its AI project's underperformance as a technology maturity problem will wait for an improvement that will not come, because the bottleneck is not in the model.

"Stopping This Project Will Damage AI Credibility Internally"

The opposite is more often true. Allowing a visibly underperforming project to continue damages the credibility of the entire AI program. When an organization is known for killing failing projects decisively and reallocating to higher-probability opportunities, its AI investments are taken more seriously by the business leaders who are deciding whether to commit their teams to new initiatives. Portfolio discipline is a credibility builder, not a credibility risk.

What to Do With the Resources You Recover

The value of stopping an AI project is not just avoiding further loss. It is the reallocation of technical capacity, operational bandwidth, and change management energy to initiatives with better fundamentals. According to a 2026 analysis of agentic AI project cancellations, 40% of agentic AI deployments are expected to be canceled by 2027 due to rising costs, unclear value, or poor risk controls. That is a lot of reallocation decisions coming. The organizations with portfolio governance in place will move faster than the ones figuring it out reactively.

Where those resources go should be governed by the same logic that should have governed the original investment: a systematic approach to prioritizing AI use cases that weighs business impact, feasibility, and strategic alignment. Not what the vendor pitched. Not what was interesting at the last executive offsite.

The enterprises building the most durable AI programs are not the ones with the most initiatives running. They are the ones that can fail fast on the wrong bets and concentrate on the ones that actually have a chance of delivering what the board was promised.

Frequently Asked Questions

What is an AI project kill decision?

An AI project kill decision is a structured evaluation that determines whether to stop, pivot, or continue an AI initiative based on forward-looking probability of success, not sunk cost. Most enterprises lack formal criteria for this decision and default to continued investment, which is why AI portfolios accumulate underperformers that consume resources without delivering measurable business outcomes.

When should an enterprise consider stopping an AI project?

An enterprise should formally evaluate stopping an AI project when it triggers three or more of the five kill criteria: no named business owner, absent success metrics, unresolvable data problems, adoption below 30% at 90 days, or budget overrun exceeding 150% with no validated business outcome. Individual criteria may be recoverable; a combination of three or more signals structural failure that more investment is unlikely to fix.

What percentage of enterprise AI projects fail to deliver ROI?

According to a Gartner survey of 782 infrastructure and operations leaders, only 28% of AI use cases fully succeed and meet ROI expectations. 20% fail outright and 52% stall in a middle category, producing modest results without reaching the outcomes they were funded to deliver. The middle category is the most costly to manage.

What is the most reliable early warning sign that an AI project will fail?

The most reliable early warning sign is the absence of a named business owner with P&L accountability and authority to redesign processes. AI projects owned exclusively by technical or IT teams stall at adoption because no one with operational authority is accountable for changing how workflows function. This structural gap predicts failure more reliably than any technical metric.

What is the sunk cost trap in AI projects?

The sunk cost trap is the organizational tendency to continue funding an AI initiative because of resources already committed, even when forward-looking evidence indicates it will not deliver its intended outcomes. Sunk costs should not factor into the continuation decision. The relevant inputs are: future investment required to reach positive ROI, realistic probability of recovery, and opportunity cost of continued allocation versus alternatives.

How does the stop/pivot/scale framework work?

The framework maps AI initiatives on two dimensions: business outcome magnitude and probability of recovery. High-impact, high-probability initiatives should be scaled. High-impact, low-probability initiatives should be fundamentally redesigned (pivoted), not just iterated. Low-impact or low-probability initiatives should be stopped. Portfolio discipline requires applying this matrix systematically, not project by project.

What is a genuine pivot versus a relabeled continuation?

A genuine pivot requires a new business case, new success criteria, and new accountability structures. Organizations sometimes use "pivot" to avoid cancellation without making structural changes. A genuine pivot redesigns the use case, data approach, or organizational model from the ground up. If the same business case, the same success metrics, and the same team are in place after the pivot decision, the project has not been pivoted; it has been extended.

Why do 40% of AI projects get canceled before delivering results?

According to Gartner's forecasts and analysis of agentic AI deployments, the primary causes are rising operational costs, unclear business value, and inadequate risk controls. Projects that lack defined success metrics at initiation, or that were approved based on technology capability rather than business need, are disproportionately represented in cancellations. Governance discipline at the start of a project is the best prevention.

What is the 8-week kill gate and how does it work?

The 8-week kill gate is a governance mechanism that requires new AI pilots to demonstrate a leading indicator of business value within 8 weeks or face default termination. It forces success criteria to be defined at the point of pilot initiation, not retrospectively. Organizations implementing this gate must pre-define what "leading indicator" means for each use case before the pilot begins.

How does Deloitte quantify the cost of delayed cancellation?

Deloitte's 2025 analysis found that organizations canceling failing AI projects at month 8 rather than month 18 saved an average of $2.1 million in additional sunk costs and reallocated resources ten months faster. The analysis also found that 42% of companies abandoned at least one AI initiative in the prior year, with average sunk costs per abandoned initiative reaching $7.2 million.

What role does adoption rate play in the kill decision?

Adoption below 30% of the intended user base at 90 days post-launch is a kill signal when no structured recovery program is in place. AI systems that real users are not using deliver no business value regardless of technical performance. Low adoption typically indicates insufficient change management, poor workflow integration, or a use case that does not solve a genuine user pain point — all of which require different interventions than technical iteration.

What is the two-track review board model?

The two-track review board separates AI initiatives into two categories: high-stakes projects requiring quarterly substantive review against hard business criteria, and exploratory projects subject to an 8-week gate before significant commitment. This separation prevents governance overhead from slowing exploratory work while ensuring scaled investments are held to business outcome standards. It is the governance structure most commonly associated with enterprise AI portfolios that consistently deliver ROI.

How should resources recovered from a canceled AI project be reallocated?

Recovered resources should be reallocated using the same prioritization framework that should have governed the original investment: a systematic approach to AI use case prioritization weighing business impact, implementation feasibility, and strategic alignment. The goal is to accelerate higher-probability initiatives, not to simply start the next project on a backlog.

How does portfolio discipline affect internal AI credibility?

Portfolio discipline builds AI credibility rather than damaging it. Organizations known for killing failing AI projects decisively and reallocating resources to higher-probability initiatives earn greater trust from the business leaders whose operational teams are the adoption gatekeepers. Allowing visible underperformers to continue — because stopping them would be politically uncomfortable — is the actual credibility risk.

What does McKinsey say about the most profitable AI organizations?

McKinsey Global Institute's 2025 research found that the most profitable AI organizations are not those that pick the best models but those that kill bad AI projects fastest and reallocate resources to higher-probability opportunities. AI portfolio management discipline, not model selection, is the primary driver of AI ROI at the enterprise level.

How does AI governance connect to project kill decisions?

AI governance provides the institutional structure that makes kill decisions enforceable rather than advisory. Without formal governance, the organizational pressure to continue underperforming projects almost always outweighs the analytical case for stopping them. Understanding how to build AI governance that enables speed means designing review mechanisms that are rigorous enough to catch underperformers but not so heavy that they slow every initiative equally regardless of risk profile.