The hardest conversation in AI consulting is not ‘your project is going to take longer than planned.’ It is ‘your project should stop.’
Most organisations are not good at this conversation. Sunk cost thinking – the instinct to keep investing because stopping feels like admitting failure – keeps AI projects running long past the point where they are recoverable. Deloitte’s 2025 research found that the average sunk cost per abandoned AI initiative reached $7.2 million, and that 42% of companies abandoned at least one AI initiative that year. The median time to abandonment was 11 months – meaning organisations were typically persisting for nearly a year before acknowledging the project was not working.
Gartner’s June 2025 research predicts that over 40% of agentic AI projects will be cancelled by end of 2027 due to escalating costs and unclear business value. The question is not whether cancellation happens – it is whether it happens at month four or month fourteen.
The Four Signals That Tell You to Stop
1. The data problem is fundamental, not fixable
Every AI project encounters data quality problems. Most are fixable with time and investment. Some are not – and the distinction matters enormously. A fundamental data problem is one where the data required for the AI to work does not exist in usable form and cannot be created within the project timeline and budget. This is different from a data quality problem that requires cleaning and normalisation work. If the audit reveals that three years of transactions were entered in inconsistent formats that cannot be reliably reconciled – and that the AI’s accuracy depends on that historical data – that is a fundamental problem. Continuing to invest in model development does not solve it.
2. The business case has changed materially since approval
AI projects are approved against a specific business context. When that context changes – a market shift, a strategic pivot, a change in the process the AI was designed to improve – the original ROI case no longer holds. The discipline here is to re-evaluate the brief at each project phase gate, not to assume the original business case remains valid throughout an 18-month project. If the use case that justified the investment no longer applies, continuing development is not pragmatic. It is sunk cost thinking with a technical veneer.
3. The project has been ‘nearly ready’ for two consecutive review cycles
The most reliable operational signal that a project is failing: it has been described as ‘a few weeks from production’ or ‘nearly ready for pilot’ for two consecutive monthly reviews without a specific, resolved blocker. This pattern reflects a project that has found its ceiling – a fundamental problem that the team is working around rather than through. The language of progress (‘we’re almost there’) is present; the evidence of progress is not. When this pattern appears, the right response is a direct diagnostic conversation with the technical lead about what the actual blocker is – not another two weeks of the same update.
4. The business owner has disengaged
When the business owner who championed the project stops attending reviews, delegates attendance to a junior team member, and stops asking questions about delivery progress – the project is already at risk. Projects with sustained business owner engagement achieve significantly higher success rates precisely because active ownership catches problems early and maintains the organisational will to resolve them. When that engagement ends, the project usually follows. Address the disengagement directly. If it cannot be recovered, stopping the project while the sunk cost is still manageable is better than drifting to a conclusion everyone can see but nobody will say.
How to Stop an AI Project Without Losing Credibility
Stopping an AI project is not failure. Continuing past these signals is. The organisations that manage AI investment best treat cancellation decisions as evidence of sound governance – not as admissions of error. Document what was learned: about the data, the use case, the integration complexity, or the organisational readiness. That documentation makes the next project more likely to succeed. And reframe the decision internally as capital reallocation, not retreat. The budget and team capacity freed by stopping a stalled project is the resource that funds the one that actually works.
Where We Come In
At DoSystems, project health reviews at defined phase gates are built into every engagement – specifically to catch these signals before they become a $7 million problem. For businesses already in a stalled project, the diagnostic conversation is the right first step. Sometimes the project is recoverable with a scope adjustment. Sometimes it is not. Either way, knowing the answer earlier is better.
Frequently Asked Questions
When should you stop an AI project?
Stop when the data problem is fundamental rather than fixable, when the business case has changed materially since approval, when the project has been ‘nearly ready’ for two consecutive reviews without a specific resolved blocker, or when the business owner has disengaged.
What is the average cost of an abandoned AI project?
Deloitte’s 2025 research found the average sunk cost per abandoned AI initiative was $7.2 million, with the median time to abandonment at 11 months – suggesting organisations persist significantly too long before making the stop decision.
How do you know if an AI project is failing?
The most reliable signals are: a data quality problem that cannot be resolved within budget and timeline, a business case that no longer applies, the same ‘nearly ready’ update for two or more review cycles, and disengagement from the business owner who originally championed the project.
Is stopping an AI project a failure?
No. Stopping a project that has hit a fundamental blocker is sound governance. The failure is continuing past clear signals in order to avoid acknowledging the problem – which is what produces the $7.2 million average sunk cost. Cancellation with documented learnings is a better outcome than a completed project that never delivers value.
#AIStrategy #AIProjects #AIImplementation #DoSystems #AIConsulting #ProjectManagement #BusinessAI
Comments are closed