Pre-mortem Example: The Failure a Premortem Cannot Catch

A pre-mortem is a prediction exercise. You sit down before the work starts, assume it has already failed, and generate the reasons. Which means the honest way to evaluate one is against a known outcome — and almost nobody does that, because the two halves are rarely both public.

Here they are both public. In February 2026 Rachel Draelos, MD PhD published "Why I Shut Down My Bootstrapped Health AI Startup After 7 Years" — an itemised, first-person, nine-cause account of why Cydoc did not work.

That is a more rigorous self-assessment than most companies ever produce, and it makes a fair test possible: run a competent premortem on the April 2018 starting conditions, and see how many of her nine it would have found.

The answer is six. What matters is which three it missed, and why one of them was never findable.

The situation being analysed

Cydoc launched in April 2018 (incorporated April 2019) and shut down on 22 August 2025. It sold automated patient intake forms and AI-generated clinical notes to small medical practices. It was not a non-starter: it reached paying customers, was granted two US patents, and demonstrated a measured 10-minute saving per visit.

Draelos is a physician with a computer science PhD from Duke, who worked on health AI during her medical training. On paper, close to an ideal founder for the problem.

What was true at the start
Founder background	MD + CS PhD, health AI research
Team	Solo founder
Funding	Bootstrapped
Target customer	Small medical practices
Founder's stated pricing philosophy	"I'll price the software at what it costs to run!"
Founder's stated customer assumption	"I was representative of my target customer"
Distribution plan	Effectively none — "build it and they will come"

Her nine causes, sorted by whether a premortem finds them

The site's premortem procedure is a 60–90 minute session: assume total failure, write reasons silently, round-robin, cluster, vote, assign mitigations. Run that in April 2018 with a couple of outside skeptics in the room, and here is the realistic yield.

#	Her stated cause	Would a day-one premortem generate it?
1	Waited over a year before talking to real potential customers	Yes — a top-three answer in almost any startup premortem
2	Misunderstood the MVP; built complexity, omitted essentials	Yes — follows directly from #1
4	No sales or marketing owner; "build it and they will come"	Yes — the single most-generated failure narrative there is
5	Solo founder wearing CEO, CTO and CRO hats	Yes — and it was a known fact, not a forecast
6	No EHR integration → workflow friction → churn	Yes, plausibly — she is a physician; the friction is domain-obvious
7	Weak technological moat; competitors replicated the core	Maybe — generic enough to appear, hard to make actionable
3	Inefficient early software practices, wasted rewrites	Unlikely — real, but not the shape of a failure story
8	A bad custom-development contract for promised revenue that never came	No — a specific contingent event years out
9	Broken business model, never addressed	No — and not for lack of imagination

Six clear hits. That is not a bad showing, and it is consistent with what the technique's own evidence base claims: a premortem reliably gets risks onto the table.

Then there is number nine.

The one it could not have found

Draelos describes the pricing decision as a deliberate act, made on principle:

"I thought to myself, I'm going to be the first person to run a business where I sell things for a fair price. I'll price the software at what it costs to run!"

And the arithmetic that eventually followed from it:

"Doctors in the small practices we had been targeting wanted to pay fewer than $100 per month for our product. Now that we'd added the AI scribe, the hosting and AI costs for our platform had gone up to about $70 per doctor."

"With revenue only $30 above hosting costs, and $4,000 to $6,000 per EHR integration, it would take 11 years per practice just to break even."

That last figure is not rhetorical. $4,000 ÷ $30 per month = 133 months ≈ 11.1 years, and that is the low end. The integration was also cause #6 — the fix for the churn problem was priced out of reach by the pricing philosophy.

Her own diagnosis of the philosophy is the sharpest line in the postmortem:

"You have to factor in all kinds of other costs, like software engineers, tech support, salespeople, marketers, legal fees, accounting fees, and so on."

Now try to generate that in a premortem. The prompt is "it is 2025 and Cydoc has failed — why?" You are asking a room to imagine a future disaster. But this is not a future disaster. It is a ratio that is false today, held in place by something that feels like a virtue. There is no failure narrative to tell about it, because nothing goes wrong — the plan works exactly as designed, and the design cannot support a company.

The Decision Replay

What the framework says	What she actually did	What actually happened
A day-one premortem generates ~6 of the 9 eventual causes — customer discovery, MVP scope, no sales owner, solo-founder risk, workflow friction, weak moat. Per its own procedure, the team then votes and mitigates the top 5–7.	Built for over a year without customer discovery; priced at cost of hosting on principle; stayed solo; shipped standalone without EHR integration.	Reached paying customers, two US patents, a measured 10-minute saving per visit — and shut down on 22 August 2025 after seven years, with the business model listed as the ninth and final cause.

The uncomfortable part is that the premortem's six hits would probably not have saved it. Fix customer discovery, fix the MVP, hire a sales lead, find a cofounder — and you arrive faster at a company that still earns $30 a month per doctor against a $4,000 integration bill.

The premortem would have made Cydoc better at a business that could not work.

The Day-One Test

The site's own premortem page already carries the technique's central caveat, and it is worth restating precisely because this case extends it. The 1989 Mitchell, Russo and Pennington study behind prospective hindsight measured how many reasons people generated, not whether the reasons were any good. Getting risks onto the table is what the evidence supports; ranking them is the load-bearing step.

This case adds something the triage step cannot fix. Triage can only sort what was generated, and there is a class of failure the generation step never produces.

Look at what the 1989 study actually found: describing an event as certain rather than possible produced longer explanations containing more episodic reasons. Episodic means story-shaped — events, sequences, things that happen. That is what the technique is built to recruit, and it is very good at it.

A broken unit economic is not episodic. Nothing happens. It is simply already true.

The Day-One Test — before running a premortem, sort your plan's load-bearing assumptions into two piles: the ones that need imagination to evaluate, and the ones that need a calculator. Compute the second pile now, because those quantities are already either true or false and no failure narrative will change them. Then run the premortem on the first pile, which is what prospective hindsight is genuinely good at.

The calculator pile is short and nearly always the same shape:

• Unit cost against unit price. What does one customer cost to serve, and what do they pay? Cydoc: ~$70 versus under $100.
• Payback period on the thing you'll be forced to build. Cydoc: $4,000–$6,000 EHR integration against a $30 monthly margin — 11 years.
• Customers needed to cover fixed costs. At a $30 margin, one salary is thousands of doctors.
• Runway against sales-cycle length. How many full sales cycles fit before the money runs out?

None of these requires foresight. All of them were computable in April 2018, and each would have produced a number that no amount of imagined disaster could have produced.

Where this misleads

This is not a criticism of Draelos. She built a product that worked, earned patents and paying customers, and then published a nine-cause postmortem more honest and more specific than most companies manage. The finding here is about the technique, and it only exists because she did the unusual thing of publishing both halves. The counterfactual premortem in this piece is generous to her and hard on the framework, which is the right way round.

Six out of nine is a good result, not a bad one. It would be easy to read this as "premortems don't work." They do the thing they claim: surface risks a team would otherwise not voice. The claim here is narrower — that there is a category they structurally cannot reach, and that category is cheap to cover by other means.

The Day-One Test is not a substitute for a premortem, and its output can be wrong too. Unit economics change: costs fall, prices rise, a segment shifts. Cydoc's got worse mid-life, when adding the AI scribe pushed hosting to ~$70. So the calculator pile needs recomputing at every major product change, not just at founding — which is the same discipline as the premortem's own step 8, rereading the list at the midpoint.

One case is not evidence about the technique in general. This is a single well-documented failure, chosen precisely because both halves are public. It is a demonstration of a mechanism — narrative generation misses non-narrative facts — not a measurement of how often that mechanism bites.

Now run yours

1. Write the two piles first, before booking the premortem. Anything with a unit, a rate, or a period goes in the calculator pile.
2. Compute the calculator pile. If a number comes back structurally impossible, stop — you have found something a premortem would not have told you, and no mitigation list fixes an inverted ratio.
3. Then run the premortem on what's left, and remember that voting and triage is the load-bearing step, not the brainstorm.

The free Pre-mortem worksheet runs step 3 in the browser with no signup — assume the failure, list the causes, then rank them.

Want to run a premortem on your phone? Framework for iPhone & iPad ships a pre-mortem worksheet with AI assistance alongside 100+ other frameworks.

Want to go deeper

• Pre-mortem framework — the procedure, the provenance chain, and The 30% Gap
• How to run a premortem — the full facilitation guide
• Pre-mortem vs post-mortem — the two halves of the same instinct, and when each is worth the meeting
• Ikigai Career Example: The Four Circles Pointed at Google — the other Decision Replay, on a personal decision with a known outcome
• Pros and Cons Example: 9 Pros, 5 Cons, and the One That Decided It — a real published pros/cons list, graded seven years later
• Ansoff Matrix Example: The Growth Move That Isn't on the Grid — the first business-decision replay, where the winning move is not on the grid

Cover photo: Patrick Perkins on Unsplash.