AI Field Manual
FM-PM-03 · Guide

Project Management  ·  Intermediate

Run a pre-mortem on your risk register

Your risk register has fourteen entries. Twelve of them were written at kickoff, and most say things like "resource availability" and "scope creep." Nobody has read it since. Meanwhile the risk that will actually hurt this project is probably not on it, because nobody has had a reason to imagine it yet. A pre-mortem gives them one.

Status
Published
Reading time
About 14 minutes
You'll need
Your current risk register, a one-page project summary or charter, the schedule's key milestones and dependencies, and an approved AI assistant
Last reviewed
November 2026
Before you start

What AI is for here, and what it isn't

A pre-mortem asks the team to imagine that the project has already failed, then explain why. The technique was described by psychologist Gary Klein in "Performing a Project Premortem" (Harvard Business Review, September 2007), and it works because "what went wrong?" is easier to answer honestly than "what might go wrong?" Imagined hindsight gives people permission to name uncomfortable risks.

AI is useful here as a tireless generator of plausible failure stories, including ones your team is too close to the project, or too polite, to raise. It's also useful for the tedious part: comparing those stories with your register and finding what's missing.

What it can't do is know your project. It doesn't know that the integration vendor has missed three deadlines on other work, that the sponsor is leaving in March, or that "finance sign-off" really means one person who's on leave in February. Its failure stories are generic until your team tests them against reality. And it must never set the probability or impact ratings: those depend on context only the team has.

Step 1

Write the failure statement

A pre-mortem starts with a specific, vivid failure, not a vague one. Write it in the past tense, as of a date after the project should have finished:

"It's 30 June. The CRM migration went live six weeks late, 30% over budget, and sales reps are still keeping their own spreadsheets because they don't trust the data. What happened?"

The specificity matters. "The project failed" produces generic answers. A failure with a date, a cost and a symptom forces concrete explanations.

You can run more than one: one for a schedule failure, one for an adoption failure ("it went live on time, and nobody uses it"), one for a quality failure. Different failures surface different risks.

Step 2

Generate failure stories with AI

The pre-mortem prompt
Act as an experienced project reviewer running a pre-mortem. The
project below has already failed in the way described. Write the
explanations a candid post-project review would find.

PROJECT SUMMARY: [scope, objectives, approach, key dates, team shape
by role, major vendors and dependencies]
FAILURE STATEMENT: [your past-tense failure]

PRODUCE 12 distinct failure stories. For each:
- What happened, in two or three sentences, as a chain of events.
- The earliest warning sign that would have been visible.
- Which category it belongs to: scope, schedule, people, vendor,
  technical, data, adoption, governance, external.

RULES: Make each story specific to this project's details, not
generic ("poor communication"). Include at least two stories about
adoption or business change, and at least two about decisions and
governance. Do not rate probability or impact. Do not propose
mitigations yet.

Asking for the earliest warning sign is what makes the stories useful later: it turns each one into something you can monitor.

Step 3

Run the human pre-mortem too

Don't skip this. The AI's stories are a warm-up, not a substitute.

Gather the core team for 45 minutes. Read the failure statement aloud. Give everyone five minutes to write their own reasons for the failure, silently and individually, before showing them the AI's list. Silent writing first matters: it stops the loudest person, or the AI, from anchoring the room.

Then go round the room, one reason per person per round, until the ideas run out. Only then share the AI's stories and ask: "Which of these could actually happen here? Which are impossible, and why?"

The team's reasons will be more specific and more uncomfortable than the AI's. The AI's will be broader and will cover categories the team forgot. You need both.

Step 4

Compare with your register and find the gaps

The register gap-analysis prompt
Compare the failure stories below with the current risk register.

RISK REGISTER: [ID, description, owner, rating, response]
FAILURE STORIES: [the AI's list plus the team's, labelled by source]

PRODUCE:
1. COVERED: stories that an existing risk already addresses, with the
   risk ID. Note where the register entry is too vague to be useful.
2. GAPS: stories with no corresponding risk. Draft a risk statement for
   each in the form: "Because [cause], [risk event] may occur, leading
   to [effect on objectives]."
3. STALE ENTRIES: register risks that no story touches and that may be
   obsolete or too generic to manage.
4. WARNING SIGNS: for each gap, the observable early indicator.

Do not assign probability, impact, ratings or owners.

The cause–event–effect format is worth insisting on. "Vendor risk" can't be managed. "Because the integration vendor is also delivering for two other clients in Q1, interface builds may slip, delaying end-to-end testing and go-live" can.

Step 5

Rate, own and respond, with the team

The team rates each new risk's probability and impact using your project's scale, chooses a response, and names one owner who'll watch its warning sign. This is where judgment matters most, and it's where AI stays out.

Use the five threat responses from PMI's PMBOK® Guide (sixth edition), so your register speaks the same language as the rest of the organization:

  • Escalate: the risk is outside the project's authority, so hand it to the person who owns it, such as the sponsor or a program board, and confirm they've accepted it.
  • Avoid: change the plan so the risk can't occur, or so it can't affect the objective.
  • Transfer: shift ownership of the consequence to a third party, through a contract, warranty or insurance, for example.
  • Mitigate: take action now to lower the probability, the impact, or both.
  • Accept: acknowledge the risk consciously, either with a contingency plan and reserve (active) or with no action (passive).

"Monitor" isn't a response on its own; it's what the owner does while one of the five is in place.

Then delete or merge stale entries. A shorter register that people read beats a long one nobody does.

Step 6

Worked example: the migration nobody used

Fictional project. A CRM migration's register lists: resource availability, scope creep, vendor delay, data quality, budget overrun. The failure statement is the one from Step 1.

Among the AI's twelve stories, three turn out to matter:

  • Adoption: "Sales managers kept running pipeline reviews from their old spreadsheets because the new dashboards didn't match the way they forecast. Reps followed their managers. Warning sign: managers not attending dashboard design sessions."
  • Governance: "Field-mapping decisions waited three weeks for a business owner who'd never been named, compressing testing. Warning sign: open data decisions older than five working days."
  • Data: "Duplicate accounts merged during migration broke ownership rules, so reps lost visibility of their own customers on day one. Warning sign: no agreed de-duplication rules by the end of design."

The gap analysis shows that "data quality" in the register vaguely covers the third story, but nothing covers the first two. In the team session, a sales operations analyst adds the risk nobody else had mentioned: commission calculations run from the old system, and nobody has scoped moving them.

That last risk, found by a person and not the model, became the project's biggest issue. The pre-mortem's job was to create the conditions for someone to say it out loud.

Step 7

Record and revisit

Update the register the same day and send a short summary: the failure statement, the new risks with owners, and the entries you retired. Run a short pre-mortem again at each major phase gate. Risks change as a project moves, and the second pre-mortem is usually faster and sharper than the first.

Risk register entry template
ID:            [PROJECT]-R[number]
Risk:          Because [cause], [risk event] may occur,
               leading to [effect on scope / schedule / cost / benefits].
Category:      [scope / schedule / people / vendor / technical / data /
                adoption / governance / external]
Probability:   [team rating]      Impact: [team rating]
Response:      [escalate / avoid / transfer / mitigate / accept] — [specific action]
Owner:         [one name]
Warning sign:  [observable indicator] — threshold: [value]
Review date:   [date]
Source:        [pre-mortem date; AI-generated or team-raised]

What stays with you: which risks are real for this team, how they're rated, and what you're willing to accept.

Further reading: Gary Klein, "Performing a Project Premortem," Harvard Business Review 85, no. 9 (September 2007): 18–19. Project Management Institute, A Guide to the Project Management Body of Knowledge (PMBOK® Guide), sixth edition, on risk response strategies.

Esc

Saved in this browser

Reading list

Contents

Reference

Keyboard shortcuts

/ or Ctrl K
Search the manual
↑ ↓
Move through search results and index entries
Enter
Open the selected result
← →
Switch discipline tabs; previous / next entry in a workflow card
Esc
Close any dialog
?
Show this list