Four in Ten Agent Projects Will Be Cancelled. Gartner Already Told You Why.
· 6 min read · Nic Keating
Gartner names three causes for agent-project cancellations. Two are the same failure: nobody built the number that would prove the programme worked.
Four in ten agent projects will be cancelled by 2027.
That figure is Gartner's, published in June 2025, and it has been quoted so often it has stopped meaning anything. What almost nobody quotes is the rest of the sentence, which is where the useful part lives. Gartner names three causes: escalating costs, unclear business value, and inadequate risk controls.
Read those three again. Two of them are the same failure.
The three causes are really two, and one of them is a measurement problem
Escalating cost and unclear value are not independent variables. A cost becomes "escalating" at the moment nobody can put a number on the other side of the ledger. An organisation will carry an expensive thing that visibly works, and will keep carrying it through a hard budget round. What it will not carry is an expensive thing that nobody in the room can defend, and an agent programme with no outcome measure is exactly that, no matter how well the agents perform.
So the honest reading of Gartner's list is shorter than the list. Agent programmes die because they cannot prove they worked, or because someone senior decided that letting them work unsupervised was the bigger risk. Those two failures have completely different fixes, and most programmes are treating both as a technology problem. Neither one is.
Nobody in the programme is paid to produce the number that kills it
Not the vendor. Not the executive who sponsored it and stood up at the all-hands to announce it. Not the delivery team, who are measured on shipping the thing rather than on whether the thing changed anything.
So the measure never gets built. Usage dashboards get built instead, because usage is easy to instrument and always goes up and to the right. Seats, sessions, prompts, engagement. None of it is evidence. It is attendance.
And the reason the stop rule is missing goes deeper than nobody getting round to it. Writing down what would make you cancel means naming, in advance, the conditions under which a specific named person will have to be publicly wrong. That is not an administrative oversight. That is a document nobody wants to sign, in a room where everybody understands why.
Which leaves one person able to end it, and it is the person furthest from the evidence: the finance lead reading a renewal invoice larger than last year's. I cannot prove that these cancellations cluster at renewal rather than at review, and I have not seen anyone publish the timing. But Gartner's own word is *escalating*, and costs do not escalate in month three. The decision is being made when the bill arrives rather than when the information does.
The strongest argument against all of this
The serious counter-case is that a 40% cancellation rate is not a crisis. It is a healthy failure rate for an emerging technology. Portfolios are supposed to prune. Gartner's own hype cycle predicts a trough, most of these were proofs of concept anyway, and an organisation that cancels nothing is an organisation that never experimented.
That argument is right about pruning and wrong about these cancellations, for one reason: timing.
Pruning is only cheap when it is early and deliberate. A programme stopped at month three, against a criterion someone wrote down before the build started, is portfolio management working as designed. The same programme stopped under duress at month eighteen, after the integration work and the licence commitment and the internal communications, is a write-off with a press release attached. The word *escalating* tells you which of the two is being described.
Factories bought electric motors for forty years before they got anything for it
Paul David set this out in 1990, in a paper called *The Dynamo and the Computer*.
Electricity was commercially available from the 1880s. The productivity acceleration did not arrive until the 1920s. Not because early electric motors were poor, and not because factory owners were slow to buy them. They bought them enthusiastically. What they did was swap one central steam engine for one central electric motor and leave the floor exactly as it was, still laid out around the driveshafts and belts that the steam engine had required.
The gain arrived when factories moved to unit drive, putting a small motor on each machine, and then rearranged the floor around the work instead of around the power source. The technology had been sitting in the building for four decades. The building was the constraint.
Buying agents and keeping the process shape is the central-motor swap. It is the most expensive available way to get no result, and it looks like progress the entire time it is happening.
One narrower failure sits underneath the big one, and it is worth separating rather than conflating. Gartner uses the term "agent washing" for vendors rebranding assistants, chatbots and existing RPA as agentic, and estimates that of the thousands of vendors claiming agentic capability, only around 130 are real. So a share of the 40% will die for a much simpler reason than organisational shape: there was never an agent in the building. That is a procurement failure, and it is not the argument here. It is easier to live with than the argument here, which is why it is worth naming and setting aside.
Two things are true in New Zealand at once
The intent is real. Deloitte's research puts around 74% of organisations expecting at least moderate use of AI agents by 2027, against roughly 21% with a mature governance model for them. The adoption curve and the control curve are about three years apart, and that gap is the whole of Gartner's third cause.
The local settings do not close it. The Public Service AI Framework exists, and agencies are encouraged to align with it. *Encouraged* is carrying a great deal of weight in that sentence.
The contrast worth drawing is not with the United States. It is with the European Union, where high-risk obligations under the AI Act now fall due in December 2027 and August 2028. Those dates have already moved once, which is the point rather than the objection: a date that slips is still a date, and a date is a planning object. It gets a line in a roadmap, an owner, and a budget. Encouragement gets none of those things, because there is nothing to be late for.
For a board, that is not a compliance question. It is a capital question. You are being asked to fund something with a well-documented failure mode, three publicly named causes, and no external event that will tell you when you have got it wrong.
The question to ask before your next agent review
Do not ask for the success criteria. Everyone has success criteria. They were in the business case, and they are mostly a list of things that would be pleasant.
Ask two questions instead. What specifically would make us stop this, and who is allowed to say it? And what number are we looking at that is not a usage figure?
If both have written answers, you have a programme. If they do not, you have an experiment with a budget and no exit, and the person who ends it will be the one reading the invoice.
That conversation takes twenty minutes. The other version of it happens in month eighteen.