The Pre-Mortem Protocol: How to Kill Bad Plans Before They Fail (With Template)

TL;DR. The pre-mortem asks you to imagine your plan has already failed and write the story of why. It is widely recommended and the citation underneath it is misquoted. The study everyone points to, Mitchell, Russo and Pennington in the Journal of Behavioral Decision Making in 1989, is routinely summarised as showing that looking back from an imagined future increases the reasons you generate by 30%. Its own abstract says the opposite about the mechanism: temporal perspective showed little influence, while outcome uncertainty strongly affected the nature of explanations. Explanations for events framed as certain were longer, contained a higher proportion of concrete episodic reasons, and were expressed in the past tense. The active ingredient is not time travel. It is certainty. That correction is not academic housekeeping: it tells you that a pre-mortem run as “what might go wrong” is running without its engine, and a pre-mortem run as “it failed, write the obituary” is running with it. This piece gives you the protocol built on the correct mechanism, plus the base rates that tell you how wrong plans usually are, and a derived contingency table showing what buffer you would actually have needed.

This sits alongside our work on the personal decision stack and the decision journal. The journal audits decisions after the fact. The pre-mortem is the same discipline, run before.

First, the base rate

Before the technique, the reason for the technique.

The Oxford Olympics Study 2024, by Alexander Budzier and Bent Flyvbjerg of the University of Oxford, tracks the cost and cost overrun of every Olympic Games from 1960 to 2024. Twenty-three Games have usable overrun data. The headline finding is unambiguous: the Games are the only project type in the authors’ work that has never once delivered on budget.

Table 1. Real-term cost overruns, Olympic Games 1960 to 2024 (verified data, Budzier and Flyvbjerg 2024, Table 4)

Games	Real overrun	Games	Real overrun
Montreal 1976	720%	Albertville 1992	137%
Rio 2016	352%	Tokyo 2020	128%
Lake Placid 1980	324%	Sarajevo 1984	118%
Sochi 2014	289%	Paris 2024	115%
Lillehammer 1994	277%	Sydney 2000	90%
Barcelona 1992	266%	Torino 2006	80%
Grenoble 1968	181%	London 2012	76%
Atlanta 1996	151%	Calgary 1988	65%
Beijing 2022	149%	Nagano 1998	56%
		Athens 2004	49%
		Salt Lake City 2002	24%
		Vancouver 2010	13%
		Beijing 2008	2%
		PyeongChang 2018	2%

The summary statistics from the same source: mean real overrun 159%, median 118%. All 23 Games overran. Eighteen of 23, which is 78%, overran by more than 50%. Thirteen of 23, which is 57%, more than doubled.

Every one of those budgets was produced by professionals, reviewed by committees, and published as a serious forecast.

What buffer would have been enough

Here is a derivation we have not seen published, and it is the number a planner actually wants.

Take the 23 overruns above as an empirical distribution. Ask: if you had added a fixed contingency buffer to your budget, what buffer would have covered what share of the Games? This is a straightforward percentile calculation on the real data.

Table 2. The contingency buffer you would have needed (CEOtudent editorial framework, derived from the 23 verified overruns in Table 1)

If you wanted to be within budget this often	You needed a contingency of	Games actually covered
25% of the time	56%	6 of 23 (26%)
Half the time	118%	12 of 23 (52%)
75% of the time	266%	18 of 23 (78%)
90% of the time	324%	21 of 23 (91%)
Every time	720%	23 of 23 (100%)

Read the second row again. To have been right half the time, you needed to more than double your own budget at the outset. A contingency line of 10% or 20%, which is the range most plans use and most finance functions will accept, would have covered roughly one Games in five.

This is not a claim that your project will overrun like an Olympics. Megaprojects are an extreme. It is a claim about the shape of planning error: it is not symmetric, it is not small, and the tail is long. When you set a buffer by asking yourself what feels reasonable, you are sampling from your intuition, not from the record.

The Channel Tunnel makes the point in one sentence. At the initial public offering, Eurotunnel told investors that 10% would be, in its words, “a reasonable allowance for the possible impact of unforeseen circumstances.” Construction came in 80% over budget, and financing 140% over.

The citation everyone gets wrong

Now the technique, and the correction that changes how to run it.

The pre-mortem was introduced by Gary Klein in Harvard Business Review in September 2007, and reprinted the following year in IEEE Engineering Management Review. The instruction is simple: before a plan is finalised, tell the team to imagine that it has been implemented and has failed badly, then have each person independently write down the reasons why.

Nearly every article recommending this technique supports it with the same citation and the same number: a 1989 study by Deborah Mitchell, J. Edward Russo and Nancy Pennington, said to show that prospective hindsight, imagining an event has already happened, increases the ability to correctly identify reasons for future outcomes by 30%.

We went to the record. The paper is “Back to the future: temporal perspective in the explanation of events,” Journal of Behavioral Decision Making, 1989, volume 2, issue 1, pages 25 to 38. Its own abstract states the design and the result plainly. The researchers manipulated two factors: temporal perspective, meaning whether an event was set in the future or the past, and uncertainty, meaning whether the event’s occurrence was framed as certain or uncertain. On the first experiment, the abstract reports that temporal perspective had little influence, while outcome uncertainty strongly shaped the character of the explanations people produced. Explanations for events framed as certain ran longer, carried a higher proportion of episodic reasons, and came out in the past tense. The second experiment, per the same abstract, supports the reading that uncertainty governs not how long people spend explaining but which type of explanation they reach for.

No 30% figure appears in the abstract, and the manipulation that the popular retelling credits is the one the paper describes as having little influence.

Two things follow, and they point in opposite directions.

The number should not be repeated. We are not publishing it, because we could not locate it in the paper’s own summary of its findings, and a figure that has drifted through secondary sources for three decades needs a primary anchor before it earns a place in an argument. It may exist somewhere in the full text. Until someone shows the sentence, treat it as folklore.

The technique still has support, for a different reason. The paper found something real and useful: framing an outcome as certain produced longer explanations, with more concrete episodic content, in the past tense. That is a precise description of what a good pre-mortem output looks like. Klein’s instruction is not merely “look back from the future.” It is “the project has failed, and failed badly.” That is a certainty manipulation wearing a temporal costume, and certainty is the variable the study says did the work.

Why this changes the tool

If certainty is the mechanism, then the difference between a pre-mortem that works and one that produces a shrug is entirely in the framing sentence.

Weak framing, uncertain: “What are the risks here? What might go wrong?” This invites the abstract, hedged, category-level answer. Market conditions. Resourcing. Scope creep. These are risk-register nouns. They are not reasons.

Strong framing, certain: “It is nine months from now. The project failed. Everyone knows it failed. You are writing the account of why.” This forces the episodic, past-tense, concrete mode the 1989 study documented. You do not get “resourcing.” You get “the one person who understood the data pipeline left in March and nobody had written anything down.”

The second output is actionable because it names a mechanism, an actor and a moment. The first is not actionable because it names a category.

This is the CEO and student pairing in a single exercise. The CEO half is willing to say out loud that the plan failed, which requires enough security to treat your own proposal as a corpse. The student half is genuinely curious about the autopsy rather than defending the body. Most pre-mortems fail because the person running them is the person who wrote the plan, and they cannot get past the first half.

The CEOtudent Pre-Mortem Protocol

The template below is our own construction. The framing rules follow the certainty mechanism above. Everything else is chosen to keep the exercise short enough that you will actually do it.

Budget: 25 minutes solo, 40 with a group of three or more. Do it after the plan is written and before it is committed.

Table 3. The CEOtudent Pre-Mortem Protocol (editorial framework)

#	Step	Time	The exact words to use	Failure mode it prevents
1	Fix the horizon and the verdict	2 min	“It is [specific date]. The plan failed. It is not ambiguous, everyone agrees it failed.”	Hedged, category-level risk lists
2	Write the account, alone and silently	8 min	“Write what happened, in the past tense, as a story with people and dates.”	Groupthink, anchoring on the loudest voice
3	Read out without discussion	5 min	Each person reads their account. No responses, no defending.	Premature debate that kills the weaker causes
4	Cluster into distinct causes	5 min	Group the accounts. Count how many independent accounts named each cause.	Confusing one loud cause with a common one
5	For the top three, name the earliest visible signal	5 min	“If this were happening, what is the first thing we would see, and when?”	Causes you can name but not detect in time
6	Assign one owner and one tripwire per cause	5 min	“Who is watching for that signal, and what do we do the day we see it?”	A document that changes nothing
7	Set the buffer against the base rate, not the feeling	5 min	“What is the overrun record for work like this, and what buffer does it imply?”	Contingency set by optimism

Step 5 is the step that converts a pre-mortem from a discussion into a system. A cause you cannot detect early is a cause you will only meet on the day it lands. The question “what would we see first” is the one that turns a named risk into a monitorable one.

Step 7 is where Table 2 belongs. Do not ask what buffer feels prudent. Ask what the record for comparable work says, and start there.

The base rate for everything else

The Olympics are an outlier, deliberately chosen because they are the cleanest verified dataset. For calibration, here are the figures for ordinary large projects from Flyvbjerg’s overview in Project Management Journal, drawn from his own database work.

Table 4. Overrun and shortfall by project type (verified data, Flyvbjerg 2014)

Measure	Value	What it means
Rail, average cost overrun	44.7%	Costs nearly half again over forecast
Rail, average demand shortfall	51.4%	And roughly half the predicted users do not appear
Roads, average cost overrun	20.4%	With a coin-flip risk that demand is also wrong by over 20%
Dams, average delay	45%	A ten-year build takes about 14.5 years
Cost of one year of delay	+4.64% overrun	Delay and overrun are the same problem, compounding

And the framing figure the author calls the iron law of megaprojects: if roughly one in ten projects lands on budget, one in ten on schedule and one in ten on benefits, then roughly one in a thousand hits all three. Even if the true rate were double on every dimension, it would be eight in a thousand.

The lesson for an individual planner is not that everything fails. It is that on-target delivery is the outlier and should be planned for as one.

When the pre-mortem is the wrong tool

When the decision is reversible and cheap. A pre-mortem costs 25 to 40 minutes of focused attention. Spend it on decisions that are expensive to unwind. For everything else, decide and correct. Our piece on probabilistic decision-making covers the sorting.

When you are looking for permission, not information. If you have already committed and you run the exercise to demonstrate rigour, you will get a document and no change. The tell is whether step 6 produces owners with real authority to pull a tripwire.

When the group cannot say the plan failed. If the plan’s author outranks everyone in the room and is present, the certainty framing collapses into politeness and you get a risk register. Run it anonymously, or run it without them.

Frequently asked questions

Is the pre-mortem actually evidence-based?
Partly, and less than its popularity implies. The most-cited support for it is misreported, as shown above. What the primary source does support is that framing an outcome as certain changes the kind of explanation people generate, toward longer, more concrete, past-tense accounts. That is a mechanism, and it is the one the technique uses. It is not a demonstration that pre-mortems improve project outcomes, and we are not aware of a large controlled test that shows they do.

How is this different from a risk register?
A risk register lists categories and assigns probabilities. A pre-mortem produces narratives with actors and moments. The 1989 finding is that certainty framing shifts people toward episodic content, which is exactly the difference. Both have a use. Only one tends to name the person who left in March.

How far out should I set the horizon?
Far enough that the failure is plausible and near enough that you can picture the intervening months. For most individual plans that is the natural endpoint of the plan itself. The evidence gives no guidance on this, so treat it as a working rule.

Does doing this make me pessimistic?
It makes your forecast less optimistic, which is the point. Table 2 is the argument: the median Olympic budget needed to be more than doubled to be right. A planner who added 20% and felt cautious was not cautious, they were wrong by a factor of five. Adjusting toward the record is not pessimism, it is calibration.

Can I run this alone?
Yes, and it is worth doing, though you lose the main benefit of step 4, which is counting how many independent accounts converged on the same cause. Solo, compensate by writing the account in one sitting without editing, then leaving it for a day before clustering.

Should I use AI to generate the failure story?
As a supplement after your own pass, not before it. If you read the model’s account first you anchor on it, which defeats the independence that step 2 is protecting. Write yours, then ask for a second account, then compare what only one of you named. Our piece on which decisions to delegate to AI sets out the wider boundary.

Sources

Budzier, A., and Flyvbjerg, B. The Oxford Olympics Study 2024: are cost and cost overrun at the Games coming down? Said Business School working paper, University of Oxford, July 2024.

Flyvbjerg, B. What you should know about megaprojects and why: an overview. Project Management Journal, 2014, volume 45, issue 2, pages 6 to 19.

Mitchell, D. J., Russo, J. E., and Pennington, N. Back to the future: temporal perspective in the explanation of events. Journal of Behavioral Decision Making, 1989, volume 2, issue 1, pages 25 to 38.

Klein, G. Performing a project premortem. Harvard Business Review, September 2007. Reprinted in IEEE Engineering Management Review, 2008, volume 36, issue 2, pages 103 to 104.

Crossref metadata registry. Record queries run 4 September 2026.

Contingency percentiles in Table 2 were computed by us from the 23 published real-term overruns in Table 1 and independently verified against the source’s own reported counts of Games exceeding 50% and 100%.

---

This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.

This post is also available in: