Decide What Would Make You Stop Before You Start
By Martin Bergmann | AI Project Lab | October 2026
Why a business case should end with a stop rule, and how an AI pre-mortem can write it.
Before you approve your next project, ask yourself one question: what would need to happen in 90 days for us to stop it?
I think most business cases cannot answer it. The numbers are rarely the weak part. The weak part is that nobody decided in advance what evidence would change the decision, so once the money is approved, stopping becomes an argument rather than routine.
Approval rewards the case that looks best.
Antonio Nieto-Rodriguez recently argued that a business case functions more like a bid than a forecast: it must be convincing enough to secure approval. Bent Flyvbjerg’s research on transport projects shows what can be selected for. He argues that the projects that look best on paper tend to be the ones with the largest cost underestimates and benefit overestimates, a pattern he calls “survival of the unfittest.” In his data, rail projects averaged a 44.7 percent cost overrun, and rail passenger traffic came in 51.4 percent below forecast. For IT projects, he and Alexander Budzier studied 1,471 of them in 2011. The average cost overrun was 27 percent, which sounds manageable, but about one in six was a "black swan" that overran its budget by 200 percent on average.
A business case answers “should we start?” once, and it is judged on how convincing it sounds that day. Nothing in it asks whether the project should continue.
Fund 90 days, then decide.
Venture investors do not hand over everything at once. Paul Gompers’s 1995 study of venture capital, based on data from 794 venture-backed firms, describes staged funding as a way for investors to gather information and maintain the option to stop.
Nieto-Rodriguez makes a similar move for corporate projects: fund in tranches tied to evidence. I like a plain version of it. Venture funds do not literally release money 90 days at a time, so the rhythm is my adaptation, not their practice. Fund the first 90 days. Then look at the evidence and choose one of three: stop, change direction, or fund the next 90.
This only works if you define the evidence before the first stage. Without that, the 90-day review just turns into another status meeting, and the decision goes back to whoever makes the best argument. My earlier post on zombie projects was about having permission to stop. This approach builds that permission in before the project even starts.
Where AI actually helps.
The way to define the evidence in advance is a pre-mortem. Gary Klein described it in Harvard Business Review in 2007: imagine the project has already failed, then explain why. He reports that imagining the outcome as already known improved people’s ability to identify reasons for future outcomes by 30 percent, citing a 1989 study.
AI can do the first pass. Give it the business case, the plan, and the risk log, tell it the project has failed, and ask for four things: three to five distinct ways it failed, the systemic root cause of each, what you would see early, and what you would do about it before approval. The early warning signs become the tests for each 90-day tranche. People then challenge the draft, which is where the real thinking happens.
Three cautions. Any likelihood the AI assigns is a structured guess, not a measurement. It can write a confident failure story that is not grounded in your documents, so ask it to name the line in the business case behind each scenario. And as my last post argued, put real project files only into the AI tool your company has approved.
A business case asks whether to start. A tranche asks, every 90 days, whether to continue. The pre-mortem writes down in advance what would make the answer no.
Martin Bergmann is the Director of an IT PMO and founder of the AI Project Lab. He writes about AI, project management, and the future of intelligent work.
Comments