Almost nobody's is, and the reason has very little to do with the technology. A story about a question nobody could answer, and the three things that would have made it answerable.
Around 95% of enterprise AI pilots return nothing measurable, and the reason is almost never the technology. It is that the company was not built to absorb it.
Full argument below · about 13 minutes
Eighteen months into the program, somebody finally asks the right question.
It is usually the CFO, and it usually comes near the end of the meeting, after the slides are done and the room has started thinking about lunch. The question is not hostile. She wants to know what capability the company now has that it did not have when this started.
The program lead has an answer ready, because he has been expecting something like it for two quarters. Three hundred and forty employees have access to the assistant. Fourteen pilots ran, nine of which met their stated success criteria. Adoption is up eleven points since March. He has a slide for each of these, and the slides are good.
She listens to all of it and then says that none of those is a capability. Access is a licence count. A pilot is an experiment, and nine successful experiments are nine successful experiments. Adoption measures how many people opened something. A capability is a thing this company can do for a customer, at scale, that it could not do eighteen months ago, and the test for whether it exists is whether anybody outside this room would notice if it stopped.
So where does it let us compete that we could not compete before?
The program lead has no answer to this, and it is not his fault. Nobody chartered him to produce one. He was asked to deploy a technology to a population of employees, and he has done that, on time and close to budget.
And where it does let us compete, what stops the two companies we lose deals to from doing the same thing by Christmas?
This is the question that empties the room, because the honest answer is nothing. Everything on the slide can be bought. The licences can be bought, the assistant can be bought, and the vendor's implementation partner will happily run the same fourteen pilots for a competitor next quarter, using the deck that was built here.
Nobody in that room is being unreasonable, and the program delivered what it was chartered to deliver. The technology works, genuinely and impressively. The awkwardness comes from somewhere much further back, at the point where the charter was written, and what was written into it was a deployment rather than a choice about where this company intends to win and what will make that hard to copy. Eighteen months is a long time to find that out.
•••
This meeting is happening all over the place right now. Around $2.5 trillion has gone into enterprise AI, and researchers at MIT put the share of pilots with zero measurable impact on profit and loss at roughly ninety five percent. Fewer than one in five pilots ever crosses into full production. Gartner expects that close to a third of the generative AI projects started in 2024 will simply be abandoned.
The comfortable explanation is that the technology is early and the returns are coming. That was true for a while. It is much harder to argue now, because the models are good, and because they are available to everyone. Your competitor can license the same capability you did, on the same terms, before the end of the week. Whatever the advantage is, it cannot be the model.
There is a more useful number buried in the same research. Companies that did see real returns were about twice as likely to have redesigned their end to end workflows before they picked a model. They did the difficult organizational work first, and brought the technology in afterward, into something ready to receive it.
Which turns the question inside out. The thing worth asking is not which AI to buy. It is whether AI can be sustainable, predictable and profitable inside a company that has not built the structure to hold it. The evidence is fairly blunt on this. It cannot, and a company can spend an entire budget proving the point.
Strategy is not a plan, and it is certainly not a budget. It is a short set of choices that constrain each other: what winning would actually look like, which ground you intend to fight on, what makes the position hard to copy, what you have to be capable of to hold it, and what runs underneath to keep all of that honest once real work starts flowing through it.
An AI program touches every one of those, which is why it so rarely behaves like a technology decision. Buying a model answers exactly one of them, and it is the only one that comes on a price list. The rest get decided in-house, in an order, by people who will still be there afterwards.
Hold a typical program against that set and the shape of the failure appears almost immediately. Nearly all of them can name a vendor and a budget. Very few can say which processes they deliberately left alone, what the company would end up with that a competitor on the same subscription would not, or which capability had to be in place before any of it could pay.
Choosing the ground, and choosing what makes you hard to copy on it, is what defines a position. Capability is what decides whether you get to keep it. That one is unglamorous and slow and never demos well, which is exactly why it gets skipped.
Choosing where to play is an act of exclusion before it is anything else. It means naming the specific processes and decisions where automation earns its cost, and then, more painfully, naming the ones where it does not and leaving them alone.
The reliable ground has a recognizable shape. High volume, shallow judgment, data you already own. Triage. Reconciliation. Routing. First pass drafting. Monitoring. The reporting layer that quietly consumes a quarter of somebody's week, every week, without anyone deciding that it should. Work of that kind is repetitive enough to model and frequent enough to matter, and agents put into it tend to pay back steadily rather than spectacularly.
Notice where that ground actually sits. It is inside IT, HR, Finance and Operations. Those four functions are not the support cast for the strategy. They are the terrain the whole thing gets fought on, and a company that files them under overhead has quietly decided not to play where the returns are.
How you win has an uncomfortable answer in a market where capability arrives through an API. A subscription is not a moat. If the win has to come from somewhere, it comes from the things the model touches rather than from the model.
Your operational history, your customer record, the accumulated judgment sitting in people's heads and in fifteen years of decisions: none of that can be licensed by anyone else. Then there is the workflow itself, which is the single strongest predictor anyone has found of whether the investment pays, and which no vendor will redesign for you. And there is what your senior people do with the hours that automation gives back, which is the part that compounds.
All three are foundation rather than technology. Each of them also has a prerequisite, and this is where the story gets specific.
Capability is where the ninety five percent is decided. Three things determine whether a company can absorb what it just bought. They are independent of each other, they are measurable if you are honest, and they bind together in a way that is easy to underestimate.
Plot a company on all three and the useful property shows up straight away. The return does not average out across the axes. It is capped by the shortest one.
A culture that embraces change does not mean a workforce that enjoys upheaval. It means change has stopped being an event with a start date and a finish line. That distinction matters more than it used to, because the technology now re-bases every few months. A company organized around an eighteen month program is executing a plan written against a model generation that has already been replaced twice.
There is a simple test for this, and it is unkind. Can every activity in the company be traced back to a stated goal? Not a departmental mandate, not a habit, not the fact that somebody has always run that report on a Tuesday. A goal, named, with a number attached and an owner who can say what happens when the number moves.
Two things fall out of running that test honestly. Some work turns out to have no goal behind it at all, and stopping that work is usually what funds everything else. And some goals turn out to have no work behind them, which is a different problem wearing the same clothes. Those are the ones people call priorities in meetings and nobody has actually staffed.
If nobody can name the goal an activity serves, you have not found a gap in the strategy. You have found something to stop doing, and certainly something not to automate.
Tying work to goals rather than to plans is also what makes automation survivable over time. When conditions shift, and they will, you re-point a team instead of re-running a program.
Trust gets talked about as a feeling and budgeted as a perk. It is neither. Trust is what decides how quickly a decision can travel through a company without being re-argued at every level it passes, and the moment you introduce systems that act on their own, every one of those relationships starts carrying more weight than it was built for.
It runs in four directions and all four are load bearing. From the company to its employees, which means giving people real numbers, real constraints and genuine authority to act inside them. Put agents into a team that suspects the actual goal is their own headcount and you will get quiet, thorough, entirely rational resistance, and you will deserve it.
From employees back to the company, which is earned rather than requested. That one is decided the first time somebody says the automated output is wrong and the dashboard says otherwise, and everyone watches to see who gets believed.
From customers to the company, which comes down to saying what you will do and then doing it, including being straight about where a machine is involved in their outcome. In our experience customers forgive a missed date far more readily than a missed disclosure.
And from the company out to its partners, who do better work when they are treated as a bench you are fielding rather than a cost to be squeezed. A partner working without context will build exactly what was specified, which is rarely what was needed.
Where trust is missing, companies substitute process for it. Extra approvals, sign off chains, status reporting, hedged commitments, the meeting before the meeting. Every one of those is a tax paid in speed, and the bill arrives weekly whether or not anyone books it.
This is also the axis where AI programs tend to die without anyone noticing the cause of death. Ungoverned data. Unclear ownership. A workforce that was never told the truth about why the system showed up. None of that appears as a technology failure in the post mortem, but that is what it was.
The last one is appetite. Not the poster in the break room, but a genuine intolerance for work that produces nothing anybody can point at afterward.
Tangible has a working definition here. Somebody outside your function can name what changed. A number moved. A cycle got shorter. There is a revenue line this year that did not exist last year. "Improved alignment" does not survive that test. "Onboarding went from ninety days to nine, and first year attrition came down with it" does.
Which brings us to the cost center, and to the quiet damage the label does. Call a function a cost center and you have told everyone working in it that the best available outcome is being slightly cheaper next year. It is a ceiling, self imposed, and once it is written into the budget structure almost nobody questions it again. It also guarantees that any AI deployed there gets justified on headcount reduction, which is the business case least likely to survive contact with a board.
| Function | Managed as overhead | Run as a value engine |
|---|---|---|
| IT | Keep the lights on. Hit the budget. Do not be the reason something broke. | Ships what the company sells. Turns internal platforms into customer facing capability, and operational data into a product. |
| HR | Process the hiring. Stay compliant. Run the annual cycle. | Builds the capability the growth plan actually requires, and compounds every strong hire retained and every poor one avoided. |
| Finance | Close the books. Police the spend. Report the variance. | Prices the offers, models the scenarios, finds the margin nobody went looking for because nobody was asked to look. |
| Ops | Fulfill the orders. Do not break anything. Absorb the volume. | Turns operational excellence into something other companies will pay for. The capability you built for yourself, packaged and sold. |
That last row is not a thought experiment. A surprising number of companies with genuinely strong internal capability are sitting on a sellable service and calling it a department.
Underneath the other choices sits the plainest one: what you run week to week so they survive contact with real work. For automation, most of that reduces to respecting a sequence everyone knows and almost nobody follows.
Name the goal first, because automating a process that serves no goal only makes the waste efficient. Redesign the process second, since most of its steps exist because of a problem that was solved three years ago and nobody removed the scaffolding. Automate the deterministic parts third, with deterministic tooling, which needs no model and generates no model bill. Then, and only then, put agents where judgment is shallow and volume is high. Keep a human on the calls that carry consequences, because agents extend a team's reach without ever holding its accountability, and no governance structure should pretend otherwise.
What all of this buys is not a smaller payroll. It is attention, which is the genuinely scarce resource in almost every company we work with. The constraint was never labor. It is the number of hours senior people get to spend on the work that only they can do. Automation moves hours out of the first pile and into the second, and that transfer can be measured in the same units as everything else in the business.
You are not automating to employ fewer people. You are automating so the people you employ spend their week where they make a difference.
There is a better question than whether a decision is correct, and it is worth reaching for whenever a room has been arguing for a while. Ask what would have to be true for this to work, and then go and find out whether it is. It converts an argument into an errand, and errands can be assigned.
Pointed at an AI investment, the list comes out short and awkwardly checkable.
Seven statements. If most of them are already true where you work, the technology will pay, and probably faster than the business case assumed. If most of them are not, there is no model on the market that will make them true, and the spending will land in the ninety five percent along with everybody else's.
•••
Back to the meeting. The uncomfortable thing about that room is that the CFO's questions were answerable, and had been for eighteen months.
Somebody could have said: the capability is that we now take a whole category of customer request from arrival to resolution in hours rather than days, at a quality our own auditors sign off on, and we can do that because we redesigned the process before we bought anything and because it runs on fifteen years of our own operational history. It lets us compete for the segment we have never been able to serve at an acceptable margin. And what stops the two companies we lose deals to from following us by Christmas is not the model, which they will have by Christmas. It is the operational history, the redesigned process, and the fact that our own people now know which decisions they can delegate and which they cannot, which cost us a year and two bad quarters to find out.
None of that requires better technology than the company already had. It requires the work to have been done in an order that almost nobody follows, for the entirely human reason that the first steps are slow, unglamorous and difficult to put on a slide.
AI is an amplifier, and amplifiers are indifferent to what they amplify. Pointed at a company with a long change axis, trust running in every direction and a real appetite for measurable value, it compounds. Pointed at a company missing one of the three, it produces the same result faster and with more conviction, which is worse rather than better.
The work that makes the technology pay is the work that was worth doing anyway. That is the whole argument, and it is not a new one.
Talk to an advisor →