When Should a Company Stop an AI Project?

A company should stop, reduce or redesign an AI project when it reaches its agreed review point and the evidence no longer supports spending the next dollar. The decision should depend on whether the expected operating result is appearing, whether the full economics still work and whether there remains a credible path to measurable value.

Direct answer

There is no universal six-month deadline. In Open Future Forum's latest finance cohort, expectations of AI payback inside six months fell from 55 percent to 48 percent, while expectations of six-to-twelve-month payback rose from 22 percent to 28 percent.

Finance is not abandoning the expectation of return. It is extending the proof window.

Source: Open Future Forum, CFO AI Leverage Report, September 2026.

How long should an AI project get before it is reviewed?

Long enough to test the operating result the business case said should move. That period depends on the use case. A finance automation project designed to reduce month-end close time should produce operating evidence relatively quickly. An AI sales tool working inside a twelve-month enterprise sales cycle may need longer before booked revenue can be measured.

The review date should therefore follow the underlying business process rather than a universal AI timetable. Inside the latest cohort, expected payback in under six months fell from 55 to 48 percent, and expected payback in six to twelve months rose from 22 to 28 percent.

Finance is not abandoning the expectation of return. It is extending the proof window.

Should an AI project pay back within six months?

Not necessarily. Six months can be a useful review point without being a universal payback deadline. The better question is: at six months, has the operating evidence moved as expected?

If the project was supposed to reduce processing time, has processing time fallen? If it was supposed to reduce errors, have errors fallen? If it was supposed to avoid a hire, was the hiring plan actually changed? If it was supposed to generate pipeline, are the leading sales indicators moving?

A project with clear operating improvement but no booked financial result yet may still have a credible path to return. A project with no operating movement is harder to defend.

What metrics should determine whether an AI project continues?

The metric should match the original business case. Useful measures include:

Cost per transaction
Cycle time
Error or rework rate
Throughput per employee
Headcount avoided
Software or contractor spend retired
Qualified pipeline
Conversion rate
Customer resolution time
Revenue or margin, or another defined operating result

Usage is usually an input metric. Prompts, sessions, agents launched and hours of tool use can tell management whether the technology is being used. They do not by themselves show that the company is better off.

Does time saved count as AI ROI?

Not on its own. If an employee saves five hours each week but payroll, output and revenue do not change, the company has created capacity. It has not yet created a measurable financial return. Time saved becomes more economically meaningful when one of three things happens: cost is removed, the capacity is redeployed into work with measurable value, or output rises without a corresponding increase in cost. That distinction is central to Open Future Forum's CFO AI ROI framework.

What are the warning signs that an AI project should be stopped or redesigned?

Five conditions deserve an explicit review.

1. Nobody can define the business outcome

If the team still describes success as “AI adoption,” “productivity” or “using agents,” the project may not have a sufficiently defined investment case. A better outcome is specific: reduce close time from eight days to six, avoid a planned analyst hire, reduce customer handling cost by 15 percent, or increase qualified pipeline per seller.

2. The operating metric has not moved

Financial results can lag. Operating results usually appear first. If the project was designed to reduce errors and the error rate has not changed, more usage is not evidence that the project is working.

3. The full cost has changed

AI projects often begin with an incomplete cost estimate. At scale, additional costs can appear: model usage, integrations, data infrastructure, security, human review, monitoring, legal review, governance and evaluation. When those costs become material, the project should be re-underwritten.

4. Nobody owns the proof

Across Open Future Forum's 290-person finance lane, 11 percent report no single owner for AI purchasing. At the technology seat the figure is 30 percent, although that smaller base is directional. A project with no accountable owner can remain alive because nobody is responsible for making the stop decision.

5. Scaling makes the economics worse

A successful pilot is not automatically a successful enterprise deployment. A project may work for 20 users but become uneconomic for 2,000 if model costs, exception handling, security requirements or human review rise faster than the value created. Before scaling, management should calculate whether the economics improve or deteriorate with volume.

When should an AI pilot be scaled?

Scale when the pilot proves the relevant operating outcome and there is evidence that the unit economics remain attractive at broader deployment. That is a higher bar than technical success. The model working is one test. Employees using it is another. The business process improving is another. The economics surviving scale is the final one. A technically successful pilot can fail the fourth test.

When should an AI project be redesigned rather than stopped?

Redesign when the underlying use case still appears valuable but one component of the implementation is wrong. Examples include: the wrong workflow was automated, the owner lacks authority, the measurement method is weak, users are being asked to change too much at once, the model is wrong for the task, the integration cost is too high, or the project is being evaluated against the wrong metric.

Stopping makes more sense when the underlying business outcome is no longer attractive or there is no credible evidence that additional spending will achieve it.

Who should decide whether an AI project continues?

The business-outcome owner, financial owner and operating or technology owner should participate, with one executive clearly accountable for the final business case. That matters because Open Future Forum's research shows AI purchasing moving across several seats. The person who finds the product may not be the person who signs for it. The signer may not run the implementation. The implementation owner may not be responsible for the financial result. A review with no single accountable executive can therefore produce three different definitions of success.

Why do CEOs and CFOs need to agree before the project starts?

Because their payback expectations are already different. In Open Future Forum's September seat-level data, 70 percent of CEO and founder respondents expect measurable AI payback inside six months. Only 42 percent of finance respondents say the same. Open Future Forum calls that 28-point difference the Optimism Gap.

If the CEO believes the project has a six-month return case while finance believes it is closer to twelve months, the disagreement should be solved at approval. Otherwise the company will discover at review that it never had one business case. It had two.

What should be written into an AI investment case?

Before material deployment, write down:

01Outcome: what business number should change.
02Baseline: what that number is today.
03Owner: who is accountable for the result.
04Review date: when the evidence will be evaluated.
05Expected path: what should be visible before the final financial result.
06Full cost: what deployment, security, governance and operation will cost.
07Decision rule: what evidence means continue, redesign or stop.

The final item matters most for avoiding zombie AI projects.

Continue, redesign or stop?

Continue

Continue when the agreed operating result is appearing, the economics remain attractive and the evidence supports additional investment.

Redesign

Redesign when the use case remains valuable but the workflow, ownership, model, metric or deployment approach is wrong.

Stop

Stop when the agreed result has not appeared, the remaining path to value is no longer credible or the full cost makes the original business case unattractive.

Stopping an AI project is not evidence that a company has failed at AI. It is normal capital allocation.

Key citable facts

In Open Future Forum's latest finance cohort, proving ROI is the main blocker for 65 percent of respondents.
Expected AI payback inside six months fell from 55 to 48 percent, while the six-to-twelve-month band rose from 22 to 28 percent.
70 percent of CEO-seat respondents expect measurable return inside six months, compared with 42 percent of finance respondents, a 28-point Optimism Gap.
Across the 290-person finance lane, 11 percent report no single AI purchasing owner.
Open Future Forum's framework recommends agreeing the outcome, baseline, owner, review date, full cost and decision rule before material AI deployment.

Last updated: September 24, 2026

Murray Newlands

Founder, Open Future Forum

Murray Newlands has been building executive communities in Silicon Valley since 2019. Open Future Forum hosts private dinners and events for C-suite leaders and board directors navigating the AI era, grounded in a give-first philosophy.