When Should You Stop an AI Project? A Framework for Scale, Change or Stop Decisions
Organisations are being encouraged to experiment with AI.
Launch pilots.
Test use cases.
Give employees access to new tools.
Explore AI agents.
Learn quickly.
That experimentation is necessary.
Not every AI opportunity can be understood through a spreadsheet before anything is built.
But experimentation creates another responsibility that receives far less attention:
knowing when to stop.
An AI project can be technically impressive and commercially weak.
It can work perfectly but solve a problem that isn't valuable enough.
It can save time without creating usable capacity.
It can create value but cost too much to scale.
Employees may refuse to adopt it.
Risk may increase.
A better alternative may emerge.
Or the evidence may simply show that the original business case was wrong.
Yet organisations can find stopping difficult.
Money has already been spent.
Senior leaders have sponsored the initiative.
Teams have invested months of work.
The project has been described internally as strategic.
Nobody wants to label it a failure.
So investment continues.
This creates an important principle for AI leadership:
The objective of AI experimentation is not to prove every idea right. It is to generate enough evidence to make a better investment decision.
Sometimes that decision should be:
Scale.
Sometimes:
Change.
Sometimes:
Pause.
And sometimes:
Stop.
Why is stopping an AI project difficult?
AI projects are vulnerable to the same sunk-cost behaviours as other investments.
Once an organisation has invested:
money;
time;
technology;
political capital;
employee effort;
executive sponsorship,
the decision can subtly change.
Instead of asking:
Does the evidence justify the next £100,000?
leadership starts thinking:
But we've already spent £500,000.
The £500,000 has already gone.
The relevant investment decision is whether the next pound is justified by the expected future value.
This distinction is fundamental to good AI portfolio management.
Stopping an AI project is not necessarily failure
Suppose an organisation tests ten AI opportunities.
Three create significant value.
Four need further development.
Three demonstrate that the original hypothesis was wrong.
If the organisation stops those three early, it may have operated an excellent innovation process.
The pilots produced information.
The information prevented further poor investment.
That is valuable.
The failure would be continuing to fund all ten simply because stopping feels uncomfortable.
For more on why experimentation and commercial value are different things, read Why AI Pilots Fail: The Gap Between Experimentation and Commercial Value.
Every AI initiative is an investment hypothesis
Before an AI project begins, leadership is effectively making a series of assumptions.
For example:
We believe this problem is valuable enough to solve.
We believe AI can improve it.
We believe the technology will perform sufficiently well.
We believe employees will adopt it.
We believe the organisation can implement it.
We believe the benefit can be realised.
We believe the value will exceed the cost.
Together, those assumptions form an investment hypothesis.
The purpose of implementation is partly to test them.
This changes how leadership should view AI projects.
Instead of:
Approved = we are committed to delivering this.
think:
Approved = we are prepared to invest this amount to test whether the evidence supports the next decision.
That creates much stronger investment discipline.
For more on establishing those assumptions before approval, read How to Build an AI Business Case: From AI Idea to Measurable Business Value.
The Stratify Evidence-to-Investment Model
AI investment should progress through evidence.
HYPOTHESIS → TEST → EVIDENCE → DECISION → INVESTMENT
Then repeat:
MORE INVESTMENT → MORE EVIDENCE → NEXT DECISION
Capital therefore follows evidence.
Not enthusiasm.
Not sunk cost.
Not executive sponsorship.
Not technological novelty.
This is particularly useful for AI because uncertainty can be high at the beginning.
Leadership may not know exactly:
how well the technology will perform;
how employees will respond;
how expensive production will become;
how workflows will need to change;
how much value will actually be realised.
Staged investment allows those uncertainties to reduce before larger commitments are made.
The Stratify AI Investment Gates
A material AI initiative can move through six investment gates.
GATE 1 — PROBLEM
Is the problem valuable enough to justify attention?
Ask:
What happens today?
What does the problem cost?
Who experiences it?
Which strategic outcome does it affect?
What would improvement be worth?
If the problem itself is insignificant, AI capability is irrelevant.
GATE 2 — FEASIBILITY
Can AI realistically improve the problem?
Ask:
Does the required technology exist?
Is appropriate data available?
Can the AI integrate with the workflow?
Can security and governance requirements be met?
Are the expected outputs achievable?
The objective is not perfect certainty.
It is enough evidence to justify testing.
GATE 3 — PROOF
Does the AI actually work in realistic conditions?
This is where pilots and controlled experiments matter.
Measure:
output quality;
reliability;
error rates;
operational fit;
technical performance;
exceptions.
A successful demonstration is not sufficient.
The capability needs to work in something approaching the real business environment.
GATE 4 — ADOPTION
Will the intended users actually use it?
This gate is frequently underestimated.
A technically successful AI solution can still create almost no value if employees do not adopt it.
Ask:
Are employees using it repeatedly?
Does it fit the workflow?
Do managers support it?
Do employees trust it?
Are skills sufficient?
Are usage barriers understood?
For more on the barriers that prevent adoption, read Why Employees Aren't Adopting AI: 7 Barriers Business Leaders Need to Address.
GATE 5 — VALUE
Is the AI changing a meaningful business outcome?
This is where productivity claims need to become evidence.
Ask:
What capacity has been created?
Has that capacity been redeployed?
Has output increased?
Has cost fallen?
Has revenue changed?
Has customer performance improved?
Has risk reduced?
For more on why employee time savings are not automatically financial value, read AI Productivity vs AI Value: Why Time Saved Doesn't Automatically Become ROI.
GATE 6 — SCALE
Does the evidence justify further investment?
Now leadership considers:
total cost;
realised value;
future value;
risk;
organisational capacity;
alternative investments.
The question is no longer:
Does the AI work?
It becomes:
Is this one of the best places for us to put additional capital and organisational effort?
The four AI portfolio decisions
At each material review, leadership should make one of four explicit decisions.
SCALE
Evidence supports increasing investment.
Potential actions include:
more users;
more workflows;
greater integration;
greater geographic coverage;
additional capability;
greater AI autonomy.
Scale should be earned through evidence.
CHANGE
The opportunity remains attractive, but something in the current approach is preventing value.
For example:
the technology needs changing;
workflow redesign is required;
adoption is weak;
training is insufficient;
the business case needs revising;
governance is creating unnecessary friction.
The answer is not necessarily more investment in the existing approach.
It may be a different approach.
PAUSE
The organisation lacks enough evidence or capability to make the next commitment.
Perhaps:
critical data is unavailable;
integration is not ready;
regulatory clarity is needed;
another technology decision is pending;
organisational capacity is constrained.
Pausing preserves the option without automatically continuing expenditure.
STOP
The evidence indicates that expected future value no longer justifies additional investment.
This should be treated as a legitimate portfolio decision.
Not an embarrassment.
The Stratify SCALE Test
Before continuing investment, I would test seven dimensions:
STRATEGIC FIT
Does this still support an important organisational objective?
CAPABILITY
Does the AI perform sufficiently well?
ADOPTION
Are intended users actually adopting it?
LEVERAGE
Is it changing productivity, capacity or performance meaningfully?
ECONOMICS
Does achievable value justify total cost?
RISK
Is the risk acceptable and governable?
EVIDENCE
Is there enough evidence to justify the next investment?
Conveniently, these create:
S → C → A → L → E
The first five dimensions form a memorable SCALE lens, with Risk and Evidence acting as decision conditions.
I would not force the acronym too heavily publicly, but the underlying logic is useful.
When should you stop an AI project?
There is no universal threshold.
But several signals should trigger serious review.
1. The business problem is no longer important
Priorities change.
A problem that justified investment six months ago may no longer deserve capital today.
2. The AI cannot achieve acceptable performance
Repeated testing shows the capability is not sufficiently reliable.
The correct response may be to stop rather than endlessly optimise.
3. Adoption remains structurally weak
Employees are not using the solution despite appropriate intervention.
This may indicate poor fit between the technology and the work.
4. Value cannot be converted
AI creates productivity, but the organisation cannot turn that capacity into meaningful economic value.
5. Total cost has changed materially
Integration, usage, governance or operating costs may make the original business case unrealistic.
For more on understanding the full economic commitment, read The True Cost of AI: What Businesses Forget When Calculating AI Investment.
6. Risk exceeds the value
The use case may technically work while creating risk that is too expensive or difficult to control.
7. A better alternative exists
AI changes quickly.
The best solution six months ago may no longer be the best solution today.
Stopping one investment to redirect capital towards a stronger alternative can be good management.
8. The original hypothesis was wrong
This is perhaps the healthiest reason to stop.
The organisation tested something.
It learned.
The evidence contradicted the assumption.
That is what experimentation is supposed to do.
Stop criteria should be defined before the project starts
This is one of the strongest ways to reduce emotional decision-making later.
Before approving an AI initiative, leadership should define:
What evidence would make us stop?
For example:
Adoption below X after Y months
Accuracy below X
Cost per outcome above £X
Benefit below £X
Implementation cost exceeds threshold
Risk classification changes materially
Customer complaints exceed threshold
Human override rate remains above X
The exact measures depend on the use case.
The principle is more important:
Define failure conditions before everyone becomes emotionally invested in success.
The Stratify AI Stop Contract
For significant AI investments, I would capture five things before approval:
HYPOTHESIS
What do we believe will happen?
EVIDENCE
What will demonstrate that it is happening?
INVESTMENT LIMIT
How much are we prepared to spend before the next review?
SUCCESS THRESHOLD
What evidence justifies continuation or scale?
STOP THRESHOLD
What evidence indicates that further investment is no longer justified?
This creates an AI Stop Contract.
Again, not necessarily a legal document.
A management discipline.
It makes the future decision easier because leadership agreed the criteria before knowing the outcome.
Who has authority to stop an AI project?
This is another accountability question.
A project may have:
executive sponsorship;
business ownership;
technology ownership;
value ownership;
governance oversight.
But who can actually say:
We're not funding the next stage.
Finance may challenge.
Technology may identify technical problems.
Risk may identify unacceptable exposure.
The business owner may conclude the value isn't materialising.
Executive leadership may redirect strategic priorities.
Decision rights should therefore include investment stop authority.
For more on defining approval, escalation and stop authority, read AI Decision Rights: Who Should Be Allowed to Approve, Override or Stop AI?
The value owner should be central to the stop decision
Every material AI investment should have someone accountable for the business benefit.
That person should provide evidence on whether expected value is emerging.
If the value owner cannot demonstrate:
adoption;
operational change;
business outcome;
credible future benefit,
then continued investment should be challenged.
For more on assigning this responsibility, read Who Owns AI ROI? Why Every AI Investment Needs a Business Value Owner.
Finance should ask about the next pound, not the last pound
This is where CFO discipline becomes particularly valuable.
Finance should challenge:
What have we learned?
What future investment is required?
What future benefit is realistically achievable?
What alternatives exist?
What happens if we stop?
What happens if we continue?
The decision should be based on future economics, not historical expenditure.
For more on Finance's wider role in AI investment, read AI for CFOs: 10 Questions Finance Leaders Should Ask Before Approving AI Investment.
The AI Project Review
A Fractional CAIO or internal AI leader should be able to run a repeatable review across material initiatives.
For each use case:
ORIGINAL HYPOTHESIS
What did we expect?
INVESTMENT TO DATE
What has been committed?
PERFORMANCE
Does the capability work?
ADOPTION
Are people using it?
VALUE
What measurable outcome has changed?
TOTAL COST
What is it really costing?
RISK
Has exposure changed?
FUTURE CASE
What additional value is achievable?
DECISION
Scale / Change / Pause / Stop
This turns portfolio management into a leadership discipline.
Stop decisions improve the AI portfolio
Imagine two organisations each have £5 million available for AI.
Organisation A never stops anything.
Its capital gradually becomes trapped in legacy pilots, underperforming tools and politically protected programmes.
Organisation B routinely reviews evidence.
It stops weak investments.
Capital is redirected towards stronger opportunities.
After several years, Organisation B may have spent exactly the same amount.
But the quality of its AI portfolio could be dramatically higher.
That is why stop discipline is ultimately a capital-allocation capability.
The Stratify AI Capital Loop
We can describe the process as:
INVEST → TEST → EVIDENCE → DECIDE → REALLOCATE
Then repeat.
The important word is:
REALLOCATE.
Stopping an initiative does not necessarily mean reducing AI ambition.
It can mean moving resources towards opportunities with greater expected value.
That is a much more commercially mature approach to AI.
AI portfolios need expiry dates
There is another useful discipline leadership can introduce.
Every experimental AI initiative should have a decision date.
Not necessarily an end date.
A decision date.
For example:
On 30 June, leadership will review the evidence and decide whether this initiative scales, changes, pauses or stops.
Without a decision date, pilots can drift.
Temporary experiments become permanent costs.
Licences renew.
Teams continue working.
Nobody consciously decides to continue.
An expiry/decision mechanism forces the investment back into view.
What should happen when an AI project stops?
Stopping should produce learning.
Capture:
What hypothesis were we testing?
What happened?
Which assumption was wrong?
What did we learn about users?
What did we learn about technology?
What did we learn about economics?
Could the capability be useful elsewhere?
Should the vendor or platform still be retained?
What should future teams know?
A stopped AI initiative should therefore create an evidence asset.
This reduces the chance that another team repeats the same failed experiment six months later.
Psychological safety matters for AI investment discipline
If stopping a project damages careers or reputations, employees will hide bad news.
Project teams will naturally defend their initiative.
Metrics become optimistic.
Problems are reframed as temporary.
Investment continues.
Leadership needs to distinguish between:
poor execution
and
a well-run experiment that disproved a hypothesis.
The second should be viewed as learning.
This is especially important in AI, where experimentation is unavoidable.
Ten questions before continuing an AI investment
At every major funding gate, ask:
Is the original business problem still strategically important?
Does the AI perform sufficiently well?
Are intended users adopting it?
Has the workflow actually changed?
Is measurable business value emerging?
What is the true total cost?
Has the risk profile changed?
What evidence supports further investment?
Is this still a better use of capital than alternative AI opportunities?
If we had not already invested anything, would we approve the next stage today?
That final question is particularly powerful.
It removes much of the emotional weight of sunk cost.
Stopping AI can be a sign of maturity
An organisation where every AI project succeeds should raise questions.
Either:
the organisation has discovered an extraordinary ability to predict uncertain technology investments,
or:
it is not experimenting very much,
or:
it is not stopping enough.
A healthy AI portfolio should contain:
experiments;
successes;
changes;
failures;
stop decisions.
The objective is not a 100% project success rate.
It is a high-quality capital allocation process.
Because AI investment maturity is not demonstrated by how many initiatives an organisation launches.
It is demonstrated by how effectively leadership learns where to keep investing — and where not to.
Frequently Asked Questions About Stopping AI Projects
When should you stop an AI project?
An AI project should be reconsidered when evidence shows that achievable future value no longer justifies the additional investment, risk or organisational effort required.
Is stopping an AI pilot a failure?
Not necessarily. A well-designed pilot should generate evidence. If that evidence shows the original hypothesis was wrong, stopping can prevent larger future losses.
What are AI stop criteria?
AI stop criteria are predefined conditions that trigger reconsideration or termination of an initiative, such as poor adoption, insufficient performance, unacceptable economics or increased risk.
How long should an AI pilot run?
There is no universal duration. A pilot should run long enough to generate the evidence required for a defined investment decision rather than continuing indefinitely.
Who should decide whether an AI project stops?
The appropriate authority depends on the initiative, but business/value owners, Finance, Technology, Risk and executive leadership may contribute. Stop decision rights should be defined
Is Your AI Investment Portfolio Creating Value?
AI portfolios can accumulate quickly.
The challenge is knowing which investments deserve more capital and which should change, pause or stop.
Stratify's Executive AI Value & ROI Assessment helps leadership teams examine whether AI investment is connected to measurable outcomes and disciplined portfolio decisions.
Assess areas including:
Investment visibility
Business-case quality
Total cost
Value ownership
Benefits realisation
ROI
Scale and stop decisions
Test Your AI Value & ROI
Because the question is not:
“How much have we already invested?”
It is:
“Does the evidence justify investing the next pound?”
Comments