Executive summary
A polished return forecast can give an early AI idea more certainty than the evidence supports. The article argues for bounded learning budgets, explicit tests, review dates and stop conditions while outcomes remain unknown. Conventional valuation becomes useful when real adoption and measured costs support credible cash-flow forecasts.
The confidence of 6:52
The other day my car (the GPS in my car to be precise!) told me I would arrive at 6:52. Not "around seven." Not "give or take." 6:52, in that calm confident voice, as if the next stretch of traffic lights and school zones and the one truck that always double-parks on the narrow lane had already been negotiated on my behalf. I smiled at it, because of course I rolled in at 7:11, and the car did not apologise. It simply recalculated and offered me a new, equally precise number it had no business being so sure about.
I have been thinking about that little "6:52" a lot lately, because I keep running into it again in a far more expensive setting.
When a guess becomes a decision
It shows up, dressed in a suit, whenever an AI project goes to the people who hold the money. A team has an idea. It might be a genuinely good one. They get asked the oldest question in business, the one finance has every right to ask: what is the return? And back comes a spreadsheet, with a net present value, a discount rate, a five-year horizon, an internal rate of return that clears the hurdle by a comfortable margin. Numbers to two decimal places. 6:52.
Here is the uncomfortable part, and I want to say it plainly even though I am not a finance person and this is finance's own tool I am poking at. The number is not wrong because anyone lied. It is wrong because it is answering a question that cannot yet be answered. Nobody in that room actually knows whether the thing works. The cash flows on that confident spreadsheet are not forecasts. They are, to borrow the sharpest phrase from the book I have been writing, hopes with decimal places.
And the trouble with a confident wrong number is that it travels. An honest "we don't know yet" stops a conversation and makes everyone in the room think. A number with a decimal point ends the conversation. It gets copied into a board deck, folded into a forecast, defended in next quarter's review by people who were nowhere near the moment it was invented. A blank makes you stop and look. A decimal point launders a guess into a decision, and then the decision wears the guess's clothes for years.
Precision without visibility
The strange comfort is that the data now says this out loud. In one survey of UK chief executives this autumn, only 14 percent said they had comprehensive, real-time visibility of what their AI actually costs. Not what it returns, mind you, just what it costs, the easier of the two to see. A separate report of finance leaders found sixty percent expecting AI costs and complexity to climb further through 2027. Read those together and a quiet picture forms. We are spending more, we can barely see the bill, and we are reaching for a tool that asks us to forecast the profit to the nearest dollar. The precision is loudest exactly where the knowledge is thinnest. (Why not, right? A clean number is so much easier to put on a slide than an honest shrug.)
The reflex, when someone points all this out, is to ask for more rigour. Tighter model. Firmer ROI. No dollar moves until the business case is bulletproof. I understand the instinct, it feels responsible, and in most of finance it is exactly right. But pointed at an early AI idea it quietly does the opposite of what you want, because it rejects the future to protect a spreadsheet. The projects that can hand you a bulletproof five-year return are, almost by definition, the ones with the least left to discover. Those are rarely the ones worth exploring. Demand certainty before the first experiment and you will fund only the things that were never really uncertain, and call it discipline.
Fund the question you can answer
So here is what I would offer instead, and it is less a formula than a habit. Match the confidence of the number to the confidence of what you actually know. That sounds almost too obvious to say until you notice how rarely it is actually done.
When a project is still a question, when you genuinely cannot yet say whether it works, stop asking it to forecast a profit it has no way of knowing. Ask it the question it can actually answer: what is the most we would be willing to pay to find out whether this works, and what result would tell us that it does? That is a question with a real answer. It has a number in it (a small, bounded one), it has a test in it, and it carries an honest admission folded inside it, that what you are buying here is an answer, not a product. You are funding the learning, not the build. A few tens of thousands of dollars spent to kill a shaky idea cleanly is not a loss on the spreadsheet. It is close to the cheapest thing you will buy all year.
Notice what that question does to the room. It is not softer than the NPV demand, it is harder, because it cannot be met with a confident fiction. You cannot answer "what would tell us this works" with a wish. You have to name a test, and a date, and a line past which you have agreed to walk away. The vagueness that hides comfortably behind a five-year projection has nowhere to stand in front of a single honest question.
When the spreadsheet fits
None of this means the spreadsheet is the enemy. I want to be careful here, because it would be easy to read all this as "ignore the finance," and the opposite is true. Net present value is a magnificent instrument. It has earned its half-century. But it runs on predictable cash flows the way an engine runs on fuel, and you have to actually have the fuel in the tank. Once an AI project has crossed over from "might work" to "works, and here is the measured cost per task and the real adoption," the uncertainty has drained out of it and it has become an ordinary investment. Then, by all means, bring the full weight of it, the hurdle rate, the payback period, all of it. The sophistication was never a fancier formula. It was knowing which ground you were standing on before you reached for one. Early on, the honest number is a range and a date. Later, it is a forecast. Using the second where the first belongs is the whole mistake, and it is a surprisingly expensive one.
Which brings me back to the car. The real problem with "6:52" was never that it was wrong. "Around seven" would have been a perfectly good estimate, dressed up as a prophecy. The problem was the two digits of false confidence, because they talked me into planning around a minute that was never really there. A business case does the same thing, only the stakes are a few hundred thousand dollars and a team's year, not whether my dinner is warm. The most rigorous thing you can set in front of the people who hold the money is not a bigger number with more decimal places. Sometimes it is a smaller, braver one, with the uncertainty left honestly in view: here is what we would spend to find out, here is what would tell us we were right, and here is the date we promise to stop and look.
I am fairly sure I am only restating something finance has always known about young, uncertain investments... AI has just turned the temptation to fake the number all the way up, because the sums are so large and the fear of looking unrigorous is so real. So I would genuinely love to hear from those of you who sit in these rooms for a living. When an AI case lands on your desk polished to two decimal places, does it earn more of your trust, or less? And have you found a graceful way to ask for the honest blank without sounding like the person standing in the way?
Original sources
Sources:
IT Brief UK, 26 October 2026, reporting an EY-Parthenon survey of 100 UK CEOs (part of a global survey of 1,200 CEOs): "Only 14% of respondents said they had comprehensive, real-time visibility of AI costs, while 32% reported significant gaps in their ability to assess return on investment." https://itbrief.co.uk/story/uk-ceos-see-ai-productivity-gains-but-struggle-on-roi CFO Dive, 15 September 2026, reporting a Deloitte survey of 1,434 finance leaders at companies with at least $1 billion in revenue across 26 countries: "Sixty percent of finance leaders expect artificial intelligence costs and technical complexity to rise substantially through 2027." https://www.cfodive.com/news/cfos-brace-higher-ai-costs-fiscal-2027-deloitte/830379
#CFO #FinOps #EnterpriseAI #AIROI #AIStrategy
Editorial note: The original article wording and sequence are preserved. The executive summary, section headings, takeaways and cover illustration are editorial additions.
STRATAGEM LABS
