Executive summary
A failing AI project can remain busy and persuasive long after its warning signs become visible. The article groups seven signs into four questions about stopping rules, real-world value, cost structure and dependence on one person. A warning is a reason to act, not proof of inevitable failure: the response can be to reset, restructure or stop.
The plant kept trying to tell me
I killed a money plant not too long ago. Slowly, and with great care.
It was the kind of plant people hand you precisely because it is supposed to be 'un-killable', and for a while it thrived. Then a leaf yellowed, and I told myself it was only settling into its new spot by the window. A second leaf went, and I decided I had been overwatering, so I eased off. When the stems began to soften I became convinced I had been under-watering all along, so I made up for it with enthusiasm. Every sign had an explanation, every explanation was reasonable, and each one bought me another fortnight of not admitting the obvious: the plant had been dying for well over a month, and I had refused to read what it kept trying to tell me.
I think about that plant more than I would like to, because the same quiet refusal turns up again and again in the AI conversations I am part of, and in the research behind the book I have been writing. The stakes there are higher than one sad pot of soil.
Busy is not the same as working
Here is the uncomfortable thing about a failing AI project. It rarely looks dead. It looks busy. The demos are still impressive, the team is still earnest, the roadmap still has next quarter's breakthrough marked in a hopeful shade of green. A project that has quietly stopped working does not announce it. It keeps its colour for a long time, the way a cut flower does, and the whole expense lives in the gap between the moment it became knowably finished and the moment somebody was finally willing to say so out loud. That gap is measured in quarters. Each quarter is costly. And closing it is, oddly enough, a finance job, for reasons I will come to.
The good news is that dying projects, like dying plants, give off signals. There are seven worth watching, and reciting them as seven separate things to memorize is the fastest way to never use them. So I find it easier to fold them into four plain questions you can ask about any project, in any review, without understanding a single line of the model underneath.
Did it ever agree to die?
Did it ever agree to die? This one is really about a thing that should already exist and often does not. When you ask "at what point would we stop this?" and what comes back is some warm version of "we'll know when we see it" or "we are committed to making it work," that is not optimism, however much it sounds like it. It is a project that has quietly declared itself immortal. Every setback will be reclassified as a learning, every delay as "almost there." You do not need the other six signs if you find this one, and acting on it is not the same as killing it. You go back and reset the project on honest terms, setting the stopping line that should have been fixed at funding, a real metric and a real date and a real cap, as the price of the next dollar you release. A project that cannot name the conditions under which it would end is not confident. It is unfalsifiable, which is a very expensive thing to be.
Does it work anywhere but the demo?
Does it work anywhere but the demo? Three of the seven signs live under this question, and they share one tell: the value is always just ahead of you. There is the pilot that performs beautifully on tidy test data and comes apart the moment it meets the real, messy, uncooperative world (a small gap here is normal, a gap that never closes is not). There is the project that has supposedly shipped but cannot tell you how many people actually use it, because nobody is measuring, and unmeasured adoption is almost always lower than everyone politely assumes. And there is the team that is forever three weeks from ready, endlessly polishing an edge case no real user will ever hit, because shipping would invite a verdict they would rather postpone. Different symptoms, same condition: something that costs money continuously and delivers value never, defended each time with the same lovely word, soon.
Is the money detonating quietly?
Is the money detonating quietly? Two signs here, and both are financial in the most literal way. When a budget is wrong by more than half, that is rarely a spreadsheet slip. It usually means the team did not really understand the cost structure of the thing they were building, and a team that misread the cost to stand it up will almost certainly misread the cost to run it. The sibling of this is cost that scales the wrong way, where doubling the usage somehow triples the bill and nobody can quite account for the extra. A project whose economics get worse as it grows is not a business waiting to happen. It is a subsidy that nobody has noticed yet. These two are the quietest of the lot, because the spend hides under ordinary headings like storage, and queries, and someone else's cloud bill, a problem I have spent a fair part of the book on.
Can it survive its own team?
Can it survive its own team? The last question looks like a personnel matter and is actually a balance-sheet one. If exactly one person understands how the thing truly works, usually the person who first built it, then you do not own an asset. You own a liability that happens to be functioning at the moment. The day that person takes another job or simply loses interest, you are left holding something you cannot operate, cannot extend, and cannot even safely switch off. In AI work, where so much of the real system hides in undocumented choices about data and tuning, this one is strikingly common, and it is almost never priced into the original case. A capability only one human can hold is not really a capability. It is a rumour with good availability.
A flag, not a funeral
None of these, on its own, means a project is doomed. One sign is a flag, not a funeral. The point is not to predict failure perfectly, which no one can, but to notice the pattern while noticing is still cheap, and to treat that as a reason to act rather than to wait for more certainty. Certainty, in a failing project, has cruel timing. It tends to arrive at precisely the moment intervention has become most expensive.
Which brings me to why this lands on the finance desk at all, and not simply with the people running the work. The reason is the shape of the cost over time: the price of stepping in does not sit still while you deliberate. It climbs, and it climbs faster than feels fair. A missing stopping rule caught in month two costs a slightly awkward conversation and a corrected agreement. The very same gap found in month twelve, after a team has built its headcount and its identity around the project, costs a political fight, a large sunk sum, and often a year of runway that could have funded three better bets. I will not give you a multiplier for this, because the honest answer is that the exact figure varies and anyone who quotes you a precise one is guessing. What does not vary is the slope. Waiting is never free, and it is rarely cheap.
Reset, restructure or stop
So what does acting actually mean? Not what people fear it means. The word intervention sounds like a funeral, and that association is precisely what freezes people in place, so it is worth saying plainly that stopping is only one of three moves, and usually not the first. You can reset: put the kill criteria, the date, and the cap in writing now, late but real, and let the work carry on with honest terms. You can restructure: take something that has drifted into an open-ended commitment and re-scope it back into a bounded bet with a defined size and a defined expiry. Or you can stop, and exit at the cost someone should have been made to quantify before the money ever moved. Three moves, not one, and two of them keep the project alive.
I think of it now as reading vital signs rather than pronouncing death, which is a gentler and a more accurate way to hold it. My money plant, to be fair to it, gave me every reading I needed. The yellow leaf, the soft stem, the slow retreat. The information was all there, on time, free of charge. I just preferred the version of the story in which everything was fine. AI projects offer the same courtesy, the same early readings, the same standing invitation to look away. The only real question is whether we read the gauges while the fix is still small, or whether we wait, as I did, for the certainty that only ever turns up once it is too late to use.
I would love to hear which of the four questions you find hardest to ask out loud. For me it is almost always the first one.
#CFO #FinOps #EnterpriseAI #AIGovernance #AICostManagement
Editorial note: The original article wording and sequence are preserved. The executive summary, section headings, takeaways and cover illustration are editorial additions.
STRATAGEM LABS
