When should a project stop? Agree exit conditions before cost becomes the argument
Set decision conditions before another investment and distinguish a pause, a smaller scope and closure.
Article language: English

In this article
“We've already done so much; it would be a shame to stop now” sounds reasonable until it becomes the only justification for more work. The project is then fuelled by its own history, although it was meant to deliver a future result. Agree stopping conditions before the next investment: what finding would undermine the goal, what obstacle would make delivery impossible and who will decide what happens to the work?
Stopping does not always mean closing everything. It may mean a pause until a condition is met, a smaller scope or giving up one method of solving the problem. The important thing is to name exactly what stops, what is still needed and what should happen to tasks already started.
Separate the goal from the solution the team likes
Imagine a fictional company testing an AI assistant to draft responses to sales enquiries. The goal is to shorten the path to a correct answer. The tool is one possible method. If checking a draft takes more work than writing the answer from scratch, the need for more efficient handling of enquiries does not disappear. The case for the chosen method weakens.
Before the trial, record which category of cases the solution should handle, who will check the result and which current process it will be compared with. An AI pilot evaluation plan (in Polish) (in Polish) should provide the basis. Without a shared definition of a correct result, everyone can declare success in a different contest.
In the UK Service Manual's guide to discovery, completing discovery involves deciding whether further investigation of a solution is justified. The guide permits stopping when findings indicate that is the right choice. It concerns public services, but the principle is useful: discovery should also allow a decision not to enter the next stage.
You therefore do not have to defend the original idea to prove the value of earlier work. Test results may indicate that improving offer material or simplifying the approval flow would have a greater effect. Stopping one method can open the way to a better understood solution to the same problem.
Record the conditions before seeing the results
A stopping criterion should identify an observation and its consequence. “We'll stop if it goes badly” allows the boundary to move at will. “We will not expand the trial if we cannot provide the agreed check for every response in the selected scope” connects an organisational result to a decision about scale. A premortem conversation about possible project failure (in Polish) (in Polish) can help prepare specific conditions: previously identified causes make it easier to name observations that should trigger a review of further work.
You can organise your conditions into four groups:
- Need: does the problem being studied still exist and matter to the identified recipient?
- Feasibility: do the available data, skills and tools allow the task to be done within the agreed limits?
- Quality: does the result meet the previously agreed acceptance requirements?
- Maintenance: can the company control the solution and support its continued operation?
For each group, specify what the next trial will reveal. Do not create conditions that cannot be checked at the chosen scale. A few test cases may expose an error critical to further work, but they cannot confirm every future use. Preserve the limited scope of evidence in the decision record. If there are not enough observations to assess a condition, the result is “undetermined”, not “condition met”. The case for collecting more material then needs its own justification.
If you set a numerical threshold, explain why, how it is calculated and how missing observations are treated. Do not copy a number from another company. A qualitative condition, in turn, needs an unambiguous description of the event and the appropriate assessor. Without context, the word “satisfactory” has remarkably flexible working hours.
Do not confuse a fixable error with a disproved assumption
One error does not always undermine a project's purpose. It may result from missing material, unclear instructions or incorrect input. The response is then a correction and a reassessment within the agreed scope. The situation is different when the underlying need was identified incorrectly or the required quality is beyond the company's capabilities.
For the assistant, one missing product variant in the database may justify completing the source. Having to reconstruct the entire answer from scratch every time points to a deeper problem with the solution's usefulness. The distinction requires looking at the task's progress, not just judging the final text. Establish the cause of the problem using verifiable evidence (in Polish) (in Polish) before choosing between correcting the input and abandoning the method. An automation process audit (in Polish) (in Polish) can help.
Before another correction, agree its scope, the time needed and the point for another decision. “One more fix” repeated without a limit can sustain a project as long as having no criteria. The aim is not to punish iteration. The next trial should test a specific explanation rather than simply prolong hope.
When assessing costs, include checking, corrections and maintenance described in the total cost of an AI process (in Polish) (in Polish). A free additional experiment may in practice use time needed from a person on another task. No new invoice does not mean no new cost.
Name the person who can say “we are stopping”
The project lead should raise reasons to stop, but may not always have authority to close it alone. Agree this before starting. Record who decides, who presents the results and whether effects on other teams require further agreement. Otherwise everyone may wait for a decision from someone who thinks they have already communicated it.
Prepare the conversation using principles for a decision meeting (in Polish) (in Polish). The material should distinguish the condition recorded before the trial, the observed result and the recommendation. If the condition was not agreed earlier, do not present it as a rule in force from the beginning. It can be adopted now, but describe honestly when it changed.
Allow the people doing and receiving the work to raise concerns. Someone maintaining the process may know an effect invisible in the report. Raising it should not automatically restart the whole project. It should help choose how to close or limit the scope.

Record the reasoning in a project decision log (in Polish) (in Polish). Separate a decision about the method from an assessment of people's work. If a sound trial showed a solution was not useful, that is an important finding that can be used. Expecting only good news in reports makes real limitations harder to see.
Choose a pause, a change of scope or closure
A pause makes sense when there is a concrete condition for returning: material becomes available, a skill is acquired or a dependency is completed. It needs a person responsible for checking the condition and a review date. A project left without those arrangements is not resting; it simply becomes harder to tell whether it is still alive.
A smaller scope is appropriate when some uses deliver a useful result and others fail the assessment. Update the service description, instructions and recipients' expectations. It is not enough to know internally that the assistant now answers only one category of questions if users still try all of them.
Closure means ending further work in a particular direction and dealing with the consequences. Account for the time released in prioritising projects with a limited team (in Polish) (in Polish). Do not treat it as immediate, full availability: closing the work may itself require handling cases, documentation and transferring knowledge.
Do not choose a pause solely to avoid an uncomfortable word. If there is no foreseeable condition for return and no current need, that status will mislead the next people planning resources. A clear record of closure is better, with the possibility of proposing a new project when a new reason arises.
Close the work, not just its name on a list
The GOV.UK guide to retiring a service gives an important place to how users' needs will be met after operation ends and to sharing knowledge with teams taking over. This is a public administration context. In a company, it is similarly worth checking where people and cases using the retired solution will go.
For automation, prepare a list of active tasks, dependencies, material and responsibilities. Establish what returns to manual handling, what another process takes over and what ends. A broader description appears in retiring an automation in a company (in Polish) (in Polish). A decision to stop work does not necessarily mean switching off an operating service at the same time.
Review permissions and connections between tools using the principles for assessing AI integration permissions (in Polish) (in Polish). Specific decisions about data and access must reflect the real system arrangement and the company's applicable agreements. Removing a tile from a project board does none of that by itself.
Keep the lessons in a form useful for another trial: the scope studied, assessment material, limitations and reason for the decision. You do not need a monument to failure or a chronicle of every meeting. You need information that prevents the next team from repeating the same unknown without a new reason.
Review example: a mixed result calls for a choice of scope
Imagine our assistant supports simple questions about one product well but makes mistakes on sets covering several variants. The team already has approved material for simple cases and can check the answers. For sets, there is no consistent source of the dependencies. The result justifies neither “everything works” nor automatically abandoning the whole direction. First compare the observations with the scope of the original trial.
If the goal was to support only simple enquiries, the trial may provide a basis for further limited use. Complex sets stay outside scope, which should be visible to users. If the company's main problem was the sets themselves, correct simple answers do not confirm the goal was reached. That is useful knowledge, but it concerns a different use. The team should assess it separately instead of changing the definition of success after the test.
A possible decision is: close the current trial for complex enquiries, keep the results and do not start another iteration until the source has been organised. In parallel, a new, smaller scope for simple answers can be considered with its own justification for the effort. Record the two choices separately. The history will then show that the original assumption was not confirmed and that the new idea follows from observations. It will not look like uninterrupted progress towards a goal that changed address several times on the way.
Before closing, check effects on the people doing the work. Is someone waiting for an integration? Has marketing announced a feature? Has customer support prepared instructions? Each dependency needs information on what changes and when. The decision can save further work only if it reaches the places where that work is planned. A record in the sponsor's folder will not automatically stop an assignment already sent to another team.
Finally, prepare conditions for any new discovery: what material would have to appear, what skill would be needed and who would assess whether the situation has really changed. Do not give the project a permanent place in the queue solely because it was first. A return should compete with other needs on its current justification. Preserved knowledge should help a new assessment, not decide its result in advance.
Try the condition on a project still under way
Choose one initiative and finish the sentence: “If we learn that..., we recommend...”. Ask the person doing the work, the recipient and the sponsor for independent answers. Compare whether they describe the same event and the same consequence. A difference reveals an agreement that is missing before a dispute about the result arises.
Add the source of evidence, the assessor and the next review point. If the condition cannot be checked, change the trial plan or name the limit of available knowledge. A good criterion does not predict the future. It allows the next choice to be made from information that can actually be collected.
When assessing whether to improve, limit or end an automation, see the scope of work at Aeneia Audit. Prepare the original goal, checking results and a list of continuing obligations; that is a better start than “we have already put too much in”.
Aeneia Audit: Show a specific process and the automation trials so far. We can discuss the diagnosis of tools, costs and information flow needed to assess the next stage. Identify the basis for the next automation decision.







