An AI Gate for Business Cases: Challenge the Investment Before You Fund It

The business case is a story, and stories go unchallenged

After years of sitting on funding boards, I have come to a simple conclusion: every business case is a narrative. It is a set of assumptions about value, cost, risk, and timing, dressed in a spreadsheet. Most governance processes interrogate the spreadsheet and accept the narrative. The assumptions (the part most likely to be wrong) almost never get a serious, adversarial challenge, and there is a structural reason for that: the people signing the case are the same people who will own its delivery. The incentive to challenge hard is low; the incentive to get the case through is high.

I have watched this asymmetry produce the same failure repeatedly. A flagship case is presented. The board questions the numbers, which have been stress-tested. Nobody questions the assumptions, which have not. The case is funded. Eighteen months later the outcome is missing, and the post-mortem lands on the same finding every time: the value drivers were assumptions presented as facts, and nobody pressure-tested them.

This is the problem I set out to fix with an AI gate: before a business case is funded, it runs through an LLM-driven challenge panel that stress-tests the assumptions against the organisation's own data, the market, and the known failure modes of similar initiatives.

Why assumptions escape challenge

Amy Edmondson's work on psychological safety explains the human half of why assumptions go unchallenged: in most organisations, the cost of dissent in a funding forum is higher than the cost of acquiescence, so dissent gets priced out.¹ A junior analyst who knows the adoption assumption is heroic is not going to say so in front of the sponsor who wrote it. The structural half is that committees are time-boxed, and assumption-challenge is unbounded; you cannot, in a one-hour board, walk through forty assumptions adversarially.

An AI gate addresses both halves. It has no career cost for being the one who challenges. It is not time-boxed. And it applies the same adversarial standard to a flagship case and a back-office case, which no human committee ever does, because human committees ration their challenge according to politics.

How the gate works

I designed the gate in four steps, and it is the four steps that matter; the model behind it is secondary.

1. Extract assumptions. The case is decomposed into its explicit and implicit assumptions: the value drivers, the cost levers, the timing dependencies, the people dependencies. This step alone is valuable, because most cases never make their implicit assumptions visible, and a case that cannot list its assumptions is a case that cannot be challenged.
2. Challenge each one. The gate asks, for every assumption: What would have to be true for this to hold? What evidence supports it? What is the closest historical analogue, and what happened? Kersten's point in Project to Product is relevant here, most initiatives fail not at delivery but at the value-realisation step, because the value assumption was never tested against the flow the initiative actually changed.² The gate makes the value assumption explicit and tests it.
3. Score fitness. Each case gets a fitness score across value, risk, and assumption strength, plus a set of conditions that must be met before funding. The score is not a verdict; it is a flag. A low fitness score does not kill a case; it forces the sponsor to answer the conditions before the board sees it.
4. Recommend, don't decide. The gate never approves funding; it raises the bar for the conversation that does. This is the line I will not cross. Funding decisions belong to humans who are accountable for the money. The gate's job is to make sure those humans walk into the room already knowing where the weak assumptions are.

What I learned running it

The first thing I learned is that sponsors start writing better cases the moment they know the gate is in the loop. Within two cycles, the assumption sections of incoming cases got materially denser, because the gate embarrassed thin cases publicly enough that nobody wanted to submit one. That is the gate's biggest effect, and it is behavioural, not analytical.

The second thing I learned is that the gate finds the assumption everyone was quietly avoiding. In nearly every case I have run through it, there is one assumption that the sponsor and the board both knew was heroic and agreed not to discuss. The gate does not agree not to discuss it. That is the point.

The third thing (and the one I had to be careful about) is that the gate has to be calibrated to the organisation's own history, or it reverts to generic challenge. A gate that asks "is this assumption reasonable" is useless; every sponsor will say yes. A gate that asks "the last three initiatives with this adoption profile delivered 22% of their value case: what makes this one different" is doing the job. Marty Cagan's argument in Transformed (that product teams succeed when they own outcomes and are given real discovery capability) applies here in a different form: the gate is a discovery capability for the portfolio.³

Why an LLM, not another committee

A committee can be lobbied, deferenced, or worn down. An LLM gate is tireless, consistent, and applies the same adversarial standard to a flagship case and a back-office case. Its job is not to replace human judgement; it is to make the weak assumptions visible before the money commits, so human judgement has something real to judge.

I am aware of the risk that an AI gate becomes theatre of a different kind: a ritual that produces a score nobody reads. The defence is the same as for any governance artefact; the gate only works if the board is willing to let a low fitness score delay a case. If the board overrides the gate every time, the gate is decoration. I have seen both. The organisations where it works are the ones where the leadership team decided, once, that a case that fails the gate does not go to the board that cycle. That single decision is what makes the gate real.

The cheapest fix to a bad investment is to never fund it. The second cheapest is to fund it with eyes open.

I would rather lose a good case to an over-strict gate than fund a bad one through an under-strict committee. You can re-run a case. You cannot un-spend a year of capacity.

---

¹ Amy C. Edmondson, The Fearless Organization (Wiley, 2018). ² Mik Kersten, Project to Product (IT Revolution, 2018). ³ Marty Cagan, Transformed (Wiley, 2024).