Suppose it failed: why discuss failure before a project starts?
Run a project premortem before launch. Give people time to write down possible causes of failure and decide which concerns warrant a check or a change.
Article language: English

In this article
The project has a name, a deadline and a presentation. At the end of the meeting, the facilitator asks whether anyone sees a problem. For a moment the laptop fan is audible, then everyone moves on to delivery. Doubts may have remained outside a conversation focused mainly on the launch plan.
A premortem changes the question. For the exercise, the team assumes the project has already failed and looks for possible causes. That can make a dependency easier to describe than “I have a bad feeling”. The value is something specific to check or change. A collection of gloomy stories alone does not prepare the project for delivery.
The imagined failure must concern a known plan
Gary Klein describes the premortem method as an exercise that starts after a plan has been presented. Participants imagine its failure and identify possible threats and obstacles. This is the author's description of the technique, not proof that one conversation will prevent a particular undertaking from failing.
Consider a fictional service company planning an internal knowledge assistant to help consultants prepare responses. The company has materials, a chosen scope and an idea for human checking. Before a trial it wants to see what it has missed. The failure discussion should concern that scope, not the entire future of AI or the state of civilisation.
Before starting, the team needs a shared description. What kinds of questions are in the trial? Who will use the result? What would count as failure in this phase? If one person imagines drafts for checking while another imagines autonomous replies to customers, they are discussing different projects. A premortem can reveal that difference, but it is better to name it before assessing threats.
Choose a specific moment and outcome too. “After the trial, consultants still cannot use answers without reconstructing the sources from the beginning” provides a different perspective from “AI was disappointing”. A spectacular disaster is unnecessary. Ordinary ineffectiveness may be important enough to change a plan.
Give people time to write independently first
In the Atlassian Team Playbook premortem guide, participants first write answers silently, then group themes and choose them for discussion. The guide ends by assigning and tracking actions. It is a collaboration-tool provider's own workshop variant, to be assessed in your context.
In our example, a short individual note can gather different perspectives before the first explanation dominates discussion. A consultant thinks about current materials, the person maintaining the tool about access, and the trial lead about checking results. Do not assume everyone knows the same dependencies because they saw the same presentation.
Ask people to describe a cause in relation to work. “There was no time to check drafts during the actual service rhythm” is more useful than “people were not ready for change”. The second statement still needs explaining. Who could not do what, when and why? Without that, the exercise can easily become a ranking of general worries.
Do not immediately reject every idea with an assurance that the company already considered it. First understand the concern. If the plan really contains an answer, point to the material or agreement confirming it. If the answer exists only in one person's memory, the premortem has revealed a need to explain it to others.
For subjects that are difficult to discuss openly, the everyday response to a reported error matters too. The article on how teams report errors (in Polish) examines that response. A single invitation to be frank does not prove participants can safely challenge a plan. The facilitator should address comments on their merits and allow doubts to be explained without judging the person.
Choose participants who know different stages of the work. For our assistant, it helps to hear both the person using responses and the person maintaining materials. People deciding on a purchase may not see the daily checking process. This does not mean inviting the whole company. The aim is knowledge that can change the picture of dependencies in the plan under consideration.
Separate participants' notes from the facilitator's position. If the boss immediately gives their own most likely reason for failure, the group may focus on developing that story. Collect independent comments first, then add your own perspective. The guide to combining individual and group brainstorming (in Polish) describes how to structure these stages. This does not guarantee fully independent judgments, but preserves each person's starting point in the material.
Turn a general concern into a possible sequence
Suppose someone writes, “The results will be out of date.” Develop it into a question for our assistant. The source material changes when the offer changes, but the tool still uses the previous set. The consultant cannot see the date of the underlying material and treats the answer as current. This is already a possible path to failure.
The description helps find a point to check: updating the source, version information or the scope of human review. It does not establish that the problem will occur. It does let the team check whether the plan contains a condition that would reveal it. If not, the team can add a trial using an appropriately prepared example.
Another concern is “nobody will use it”. A concrete sequence is needed. Does the tool require the question to be retyped into another system? Does the answer arrive too late? Does checking take more work than preparing from scratch? Each possibility leads to a different observation. The general recommendation “increase engagement” does not identify what needs fixing.
Keep cause and effect distinct. “A customer receives an incorrect answer” is an effect. Possible causes include the source, use outside the scope or omitted checking. If the group records five similar effects, check whether they lead to a shared condition. A separate action is not needed for every wording.

For each description, ask about its basis. Does the participant know a similar case, see a gap in the current plan, or identify a possibility without supporting material? All three can be useful but should not have the same status. A concern does not become a fact because it was written in the past tense for the exercise.
If someone identifies an existing safeguard, check whether it covers the sequence being considered. A general quality check may not cover updating sources after an offer changes. Conversely, a specific trial may already have clarified part of the concern. Preserve its result and scope so the premortem does not commission the same work again solely because the problem has another name.
Voting does not establish probability
Choose themes for further assessment, but do not treat the number of votes as a measurement of risk. A popular concern may be easy to tell, while an important dependency may be known to one person only. Ask about impact on the objective, the basis for the assumption and whether it can be checked before a greater commitment.
Not every cause can be removed at once. Some need clarification, some a plan change and some a conscious agreement on how to respond. In our example, the trial could be limited to sources that can be maintained, or broader scope postponed until checking is confirmed. These are decisions about specific execution, not a formal safety certificate.
If a concern involves changes in the environment, consider strategic scenarios, as described in the guide to strategic scenarios (in Polish). A premortem examines causes of failure of a chosen plan. Scenarios help examine the plan under different conditions. Combining them may reveal that the plan depends on one external assumption requiring further research.
Check whether the trial's evaluation itself encourages the problem to be overlooked. Measures can encourage unwanted behaviour: a team can hit the reported number without achieving the useful objective. The article on misleading measures (in Polish) examines this risk. If the team counts only generated drafts, the work needed to check them is easily missed.
When assessing a cause, see whether limiting it creates another dependency. If one experienced person must additionally check every draft, confirm they can participate. Adding a checker to the plan is not enough to declare the concern removed. Our example may need a smaller trial or another way to route cases for consultation.
Preserve a rejected concern if it had a specific basis. Record why it does not require action at this stage and what event would justify returning to it. A less popular issue then does not disappear entirely. Not every comment needs its own task, but a consciously assessed one should have a clear status.
When the group disagrees about a fact, first identify the check. A dispute about whether materials are updated daily does not require voting for the best story. It requires reading the actual update process. A premortem can reveal different assumptions; the number of arguments in a meeting does not settle them.
Every selected concern needs a next step
After the discussion record what will be checked, who will do it and when the result returns for a decision. “Look after source quality” is too general. “Walk through three cases after an offer change and compare answers with current material” has a visible scope. Examples should fit the work and available data permissions.
Such a task can go in a business experiment card, as described in the guide to recording a hypothesis (in Polish). Confirming one condition does not establish that the whole project is ready. The result should, however, indicate whether the identified concern still blocks a broader scope, needs correction or has been sufficiently clarified for this particular trial.
Aeneia Audit may help examine a prepared AI application, including tools, sources, data flows and checking. Premortem material can be a useful beginning when it leads to specific questions and examples. The workshop itself does not replace reading the data or checking actual use. Start with an application you can reproduce through examples. Before work begins, discuss the scope through Aeneia Audit.
Do not extend a small exercise to every minor project move. Keep it proportionate to the decision and the possibility of changing the plan. If everything is already settled, participants should know what they can still change. Otherwise the meeting will collect sound comments with no route to further action.
Before a larger commitment, review the fate of the most important actions. Assigning an owner alone does not prove completion. The team needs to see the result of the check or an explicit gap requiring another decision. If the deadline has passed, do not mark a concern clarified merely because the workshop ended long ago. The material should serve a choice still facing the company.
Return to the record when the plan changes
New data may appear during the trial, so preserve links between a concern, a check and a decision. Project stopping criteria are useful when they identify a point for reassessment; the guide to stopping criteria (in Polish) describes how to define them. They should not be added only when a result has become uncomfortable.
After completion, compare earlier assumptions with what happened during a project retrospective. The guide to project retrospectives (in Polish) explains how to turn findings into practice. Which conditions were checked? Which changes helped? What did nobody notice earlier? The absence of a problem should not by itself be treated as proof the workshop worked. It may not have occurred for many reasons.
Pass the result to people carrying out the changed plan. If a smaller source scope was agreed, consultants and the person maintaining materials should know the same boundary. The workshop note should not be the only place where an important change exists. Someone doing the later work must be able to find the current agreement without reconstructing the whole discussion.
A well-finished premortem leaves a few concrete moves before launch: clarify a condition, narrow a scope or try a missing capability. The team may still want to deliver the project. But it should know more about what must work for the plan to rest on more than polite silence after a presentation.










