Phase: Stress-Testing · Complexity: Moderate · Time required: 45–90 min · Tool #45 (2026-native) · Origin: Gary Klein's pre-mortem (1989), extended for the model era
One line: Before you commit, assume the decision has already failed — then use an AI model as a tireless, un-intimidated red-teamer to generate the failure modes your team is too invested, too polite, or too tired to name.
1. What This Tool Does
A classic pre-mortem inverts the usual planning question. Instead of "how do we make this succeed?", the team imagines it is a year later and the project has failed badly — then works backward to explain why. The genius of the original, from Gary Klein, is psychological: it gives people permission to voice doubts they'd otherwise suppress, because they're not predicting failure (disloyal), they're explaining a failure that has "already happened" (just analysis). It converts hidden anxieties into surfaced risks.
The AI-assisted version keeps that psychology and adds a second engine. A model doesn't have your team's blind spots, sunk costs, or social fears. It won't stay quiet to avoid contradicting the boss. It won't get tired after the first five risks. Pointed at your plan, it can generate a broad, unflinching landscape of ways the thing dies — including the boring, structural, and politically awkward ones humans skip. The tool combines the two: the model produces breadth and candor, the humans supply judgment, context, and prioritization. Used together, they cover each other's weaknesses — the model's tendency toward generic or hallucinated risks, and the team's tendency toward motivated blindness.
The output is not a doom list. It's a prioritized set of failure modes, each with an early-warning signal and a mitigation — a plan made antifragile by having stared at its own obituary before it was born.
2. How to Use It — Step by Step
Instructions on the left; a worked example — a startup deciding to launch an AI feature that automates part of its customers' workflow — on the right.
STEP 1 — Frame the decision and the failure precisely. Write the specific decision and a crisp "it's twelve months later and this failed" statement. Vague inputs produce vague risks. Feed the model the real plan — the assumptions, the numbers, the timeline — not a sanitized summary. → Worked: "We shipped the auto-workflow feature to all customers in Q1. Twelve months later it's been rolled back, churn rose, and the feature is a cautionary tale internally. Explain why." The team pastes in the actual launch plan, target metrics, and key assumptions.
STEP 2 — Have the humans pre-mortem first, silently. Before the model speaks, each person privately writes their top 3 reasons for failure. This preserves the original tool's psychological safety and prevents the model's output from anchoring the humans. Collect them. → Worked: Engineers privately flag reliability on edge cases; sales privately flags that a key customer hates automation; the founder privately worries about the base model changing under them.
STEP 3 — Red-team with the model, in several passes. Prompt the model to act as a hostile critic and generate failure modes — then push it: "What would a competitor hope we ignore? What fails silently? What's the boring operational reason this dies? What second-order effects hurt us six months after launch?" Run multiple angles; breadth is the point. → Worked: The model surfaces, among ~25 risks: silent accuracy degradation on a customer segment nobody tests; a support-cost spike from confused users; a base-model update that changes behavior mid-quarter; a compliance exposure in one regulated customer; reputational risk from one viral bad output.
STEP 4 — Merge, dedupe, and — critically — verify. Combine human and model lists. Then check the model's risks against reality — the model will produce some generic, some hallucinated, some inapplicable. Keep the real ones. This verification step is non-negotiable; an unverified AI risk list is noise dressed as rigor. → Worked: The team drops three generic risks that don't apply, confirms the silent-degradation and support-cost risks are real and under-planned, and realizes the base-model-change risk maps exactly to the founder's private worry.
STEP 5 — Prioritize by likelihood × impact, and assign early-warning signals. For each surviving risk, rate likelihood and impact, then define the leading indicator that would tell you it's happening — and the mitigation. A pre-mortem that ends in a list changed nothing; one that ends in monitored signals and owners changes the plan. → Worked: Top risks get owners and tripwires: a per-segment accuracy dashboard with an auto-rollback threshold; a staged rollout instead of all-customers-at-once; a pinned model version with a change-management process; a human-escalation path for regulated accounts.
STEP 6 — Feed it back into the decision. Decide: proceed, proceed-with-modifications, or don't. The staged rollout and tripwires are the modification. Document what would make you reverse. → Worked: The team ships to 10% of customers behind the accuracy dashboard instead of 100% — the single change that later prevents a blowup when one segment does degrade.
3. When It Works Best
| Dimension | Best fit |
|---|---|
| Irreversibility | High-stakes, hard-to-reverse decisions (launches, hires, big bets) where the cost of an unseen failure mode is severe and worth 90 minutes to hunt. |
| Team investment | When the team is emotionally committed to the plan — exactly when human pre-mortems get politest and a model's un-invested candor helps most. |
| Complexity & second-order effects | When failure is likely to come from interactions and downstream effects humans don't trace — the model's breadth shines at generating "and then what" chains. |
| AI-native decisions | When the plan depends on model behavior (an AI feature, an agent, a data pipeline) — the model can reason about its own failure modes (drift, hallucination, base-model change) better than most teams. |
| Time pressure with real stakes | When you can't run a week-long risk review but 90 minutes of structured red-teaming is feasible — the model compresses the breadth-generation step. |
4. When It Breaks Down
| Failure pattern | What goes wrong | What to use instead |
|---|---|---|
| Treating the model's list as truth | The model produces confident, generic, or hallucinated risks; the team logs them as findings without verification and drowns real risks in noise. | Mandatory Step 4 verification; pair with Ladder of Inference to separate the model's claims from evidence. |
| Anchoring on the AI | Running the model first anchors the humans, who then just react to its list and lose their own distinct signal. | Humans pre-mortem silently before the model (Step 2). |
| Theater without tripwires | The team generates a great risk list, feels rigorous, and changes nothing. | Force Step 5: every kept risk needs a leading indicator, an owner, and a mitigation, or it's deleted. |
| Confidentiality / data exposure | Pasting a sensitive plan into the wrong model leaks it. | Use an approved, private deployment; redact identifying specifics; never paste regulated data into a public tool. |
| Low-stakes overuse | Running a 90-minute red-team on a two-way-door decision wastes time and breeds pre-mortem fatigue. | Reserve for irreversible or high-impact calls; for reversible ones, just decide and watch (see Reversible vs Irreversible). |
| Adversarial-prompt drift | Push the model too hard toward doom and it manufactures implausible catastrophes that crowd out real ones. | Ask for plausible failure modes with mechanisms, and rate each for likelihood before keeping. |
The most dangerous failure mode is the first: a model's confident risk list feels like rigor, and unverified it is the opposite. The tool's whole value depends on humans doing the judgment the model can't — deciding which of its many candidate failures are real, here, now. Skip that and you've automated the appearance of a pre-mortem while losing its substance.
5. Visual Explanation
DECISION (pre-commit)
│
▼
┌───────────────┐ silent, first ┌───────────────────────────┐
│ HUMAN team │──────────────────▶│ their private top-3 risks │ (context, judgment,
│ pre-mortem │ └───────────────────────────┘ political nerve)
└───────────────┘ │
│ ▼
│ ┌──────────────┐
└──────────────────────────────────▶│ MERGE + │
┌───────────────┐ breadth + candor │ VERIFY │──▶ prioritized failure modes
│ MODEL as │────────────────────────▶│ (drop the │ each with: leading signal,
│ red-teamer │ (25+ candidate risks) │ hallucinated)│ owner, mitigation
└───────────────┘ └──────────────┘ │
(no sunk cost, ▼
no fatigue, MODIFY the decision
no politeness) (staged rollout, tripwires) → DECIDE
The model supplies breadth and candor; the humans supply judgment and context; the verification gate is what keeps the model's noise out. Neither half is sufficient alone — that's the whole design.
6. Pairs With
- Pre-Mortem (the parent tool) — the AI version extends it; run the human version inside it.
- Inversion — the underlying move: study failure to engineer success. The pre-mortem is inversion with a calendar.
- Second-Order Thinking — prompt the model specifically for downstream, "and then what" failures humans miss.
- Ladder of Inference — the discipline for separating the model's claims from evidence during verification.
- Scenario Planning (use before) — build the plausible futures; pre-mortem the one you're betting on.
- Reversible vs Irreversible (use before) — decide whether the decision even deserves a full red-team.
- Red Team (mental model) — the general principle; this tool is a structured, AI-augmented instance of it.
7. Real-World Application
A product team shipping an AI agent to enterprise customers.
The scenario. The team is a week from launching an agent that will act inside customers' systems. Everyone is excited; the demo is flawless. The founder, aware that flawless demos are where blowups hide, calls a 90-minute AI-assisted pre-mortem instead of a launch celebration.
How the tool applied. The humans pre-mortemed silently first — support flagged confused users, an engineer flagged untested edge cases, an account manager flagged one regulated customer. Then the team red-teamed with a model across several passes, asking not just "how does this fail?" but "what fails silently, what does a competitor hope we ignore, what hurts us six months later?" The model produced roughly two dozen candidate failure modes. The team verified them — dropping the generic ones, keeping the real ones — and found a cluster the humans had under-weighted: silent accuracy degradation on a customer segment nobody was monitoring, and a base-model update that could change the agent's behavior mid-quarter.
What it surfaced. The decisive insight wasn't a single exotic risk — it was that the plan had no leading indicators for the most likely failures. The agent could degrade for weeks before anyone noticed, because success was measured only in aggregate. The fix was structural: a per-segment accuracy dashboard with an automatic rollback threshold, a staged rollout to 10% of customers instead of 100%, a pinned model version with a change-management process, and a human-escalation path for regulated accounts.
The non-obvious factor. When one customer segment did degrade three weeks post-launch — exactly as the pre-mortem predicted — the dashboard caught it, the staged rollout contained it to a fraction of customers, and the auto-rollback fired before it became a churn event or a viral bad-output story. The pre-mortem didn't prevent the failure; it made the failure small, visible, and reversible. That is the entire value of the tool: not clairvoyance, but converting an invisible catastrophe into a monitored, contained event.
8. Analyst's Take
The AI-assisted pre-mortem is the rare case where adding a model to an old technique genuinely upgrades it rather than just decorating it — because the model's weaknesses and the technique's weaknesses are complementary. A human pre-mortem's failure mode is politeness and motivated blindness: the team is too invested to name the real risks. A model's failure mode is confident noise: it names too many risks, some invented. Put them together with a verification gate and each covers the other — the model breaks the human silence; the humans break the model's noise.
Two cautions decide whether you get the upgrade or a shinier version of theater. First, sequence matters: humans before model, always. Run the model first and it anchors the room; you lose the very human signal — the political risk nobody says out loud, the customer relationship only the account manager knows about — that no model can generate. The model is the second voice, not the first. Second, the verification step is the whole tool. An AI risk list you don't verify is worse than no list, because it feels like rigor while being noise. The judgment about which of the model's twenty-five failure modes are real, here, now is the irreducibly human part, and it's the part under time pressure that teams are most tempted to skip. Skip it and you've automated the appearance of caution and lost the substance.
The deepest reason this tool matters in 2026: the decisions most worth pre-morteming are increasingly ones whose failure modes involve AI itself — drift, silent degradation, base-model changes, hallucinated outputs, over-trust. These are exactly the failures human teams are worst at imagining, because they're new, technical, and easy to wave away with "the demo worked." A model reasoning about its own class of failures, checked by humans who own the consequences, is the right shape for the risk. The goal, as always, is not to predict the future but to make your plan cheap to correct when it's wrong — to turn the catastrophe you can't see into the contained, monitored, reversible event you can.
9. Top Resources
- "Performing a Project Premortem" — Gary Klein (HBR, 2007). The primary source for the technique; short, practical, and the foundation everything here extends.
- Sources of Power — Gary Klein. The naturalistic-decision-making research behind why imagining failure beats brainstorming risks cold.
- Thinking, Fast and Slow — Daniel Kahneman. Kahneman championed the pre-mortem as a debiasing tool; Part 3 explains the overconfidence it corrects.
- Superforecasting — Tetlock & Gardner. For the discipline of assigning likelihoods and updating — how to rate the risks your red-team surfaces rather than just listing them.
- Vendor guidance on red-teaming AI systems (2024–2026). The state of the art on using models to adversarially probe plans and other models — the technical complement to the human method.
Note: this tool touches decision-making under uncertainty and AI over-reliance. If a pre-mortem surfaces that a decision's failure would cause real human harm, treat the model's output as input to human judgment, never as the decision itself.