The pre-mortem: assume the plan already failed, then ask why

Overview

This is about the cheapest moment to catch a bad decision: before you commit to it. The technique is the pre-mortem, and it is old, but it just got newly useful.

Why now. In July 2026 Anthropic moved its Cowork agent to the cloud — it now takes a plan and keeps working with your laptop shut, on a schedule, unattended. That is the direction of the whole field: less chatting with an assistant, more handing an agent a plan and letting it run. The more of the doing you delegate, the more your real leverage collapses onto the one thing you still own outright — the decision to hit go. The point you’ll be able to act on: a single prompt that stress-tests that decision before you make it, on a task you actually have this week.

The content

The obvious move, when you want an AI to sanity-check a plan, is to ask it what could go wrong. It will oblige — a tidy list of risks, hedged and generic. It reads as diligence. It rarely changes your mind.

Overturn it with a flip that decision researchers have studied for decades. Don’t ask what might go wrong. Assume it already has. Gary Klein named this the pre-mortem in a 2007 Harvard Business Review piece: unlike a critique, “in which project team members are asked what might go wrong, the premortem operates on the assumption that the ‘patient’ has died, and so asks what did go wrong.” Same decision, one flipped assumption — the failure treated as fact, not risk — and the answers come loose.

That certainty is doing real work. In a 1989 study by Deborah Mitchell, Jay Russo and Nancy Pennington, treating an outcome as already certain led people to generate about 30% more reasons for it — and more of them concrete and specific — than when they merely considered that it could happen. Be precise about what that shows, because the popular retellings are not: it is more reasons surfaced, not more accurate ones. The study never graded whether the reasons were right. What the flip buys you is a fuller list of candidate failures, not a verdict on which will bite.

Which is exactly where an AI earns its place. Klein’s original problem was that people stay quiet — they are reluctant to voice doubts about a plan the room has already bought into. A model has no such reticence. Ask it to assume your plan has failed and it will write you the uncomfortable, complete version of how, with none of the politeness that makes a human colleague trail off. You are not outsourcing the judgement. You are using the machine for the one part humans reliably flinch from: saying the failure out loud.

Try it

Take a real decision you are close to committing to — a launch, a hire, a proposal, a reorganisation, a big spend. Paste it in and run this:

It is [3 / 6 / 12 months] from now. The decision below has clearly failed —
not a mild disappointment, a real failure everyone can see.

Decision: [describe your actual plan, in a few sentences]

Write the post-mortem from that future:
- Tell the story of how it failed, most plausible path first.
- List the specific failure modes. For each, note roughly how likely it
  felt and how much damage it did.
- Name the assumptions I'm currently treating as facts.
- Give me the two or three earliest warning signs I'd have seen first.

Where it breaks: the model will generate fluent, plausible failures whether or not they are real, and it will happily rank them — but those rankings are invented, not measured, so don’t mistake a confident-looking likelihood for a probability. Read the output as a list of hypotheses to weigh with your own knowledge, not a forecast. Its value is narrow and real: it surfaces the failure you would never have said aloud in the meeting. Deciding which ones actually matter is still your job.

Additional reading

• Performing a Project Premortem — Gary Klein, HBR (Sept 2007) — the original framing: assume the patient has died, and ask what killed it, because people won’t voice doubts about a plan up front.
• Back to the Future — Mitchell, Russo & Pennington (1989) — the study behind “prospective hindsight”: imagining a certain outcome yields more, and more specific, reasons.
• The premortem — Jason Collins — a careful read of that 1989 study, and why the widely-quoted “30% better at correctly identifying reasons” overstates what it measured (number of reasons, not accuracy).
• Cowork is now on web and mobile — Anthropic (July 2026) — the shift that makes this timely: agents that take a plan and run it unattended.
• The AI Pre-Mortem — Linas Beliūnas (June 2026) — one practitioner packaging the pre-mortem as a repeatable AI move.

Editor’s note

I love this one, and it’s a concept that has taken me longer than it should have to publish. The “Try it” example gives a future-set time period, which can be really useful for projects that have ongoing exposure to an audience (this website, for example), but is totally unnecessary for specifically-timed events. If your meeting is in two days and you’re working on a meeting strategy, the same rule will still apply — just assume that it’s 10 minutes after the strategy didn’t go to plan in the meeting. The key is to force reasons for failure. You can follow a cake recipe and still end up with batter if your oven is broken. This can help you turn your mind to the things that sit in the corners.