I get asked this regularly now. A procurement vacancy has been open for a few months, and then the question comes up whether we should still fill it at all. We have AI now, don’t we?
Some procurement work is indeed moving to AI. Spend analyses, contract summaries, flagging supplier risks — all of that is now done faster and more thoroughly than ever before. In itself, a perfectly good development.
But the sum behind that question doesn’t add up. And you only notice a year later, when the bill arrives and the role has long since been cut.
A salary is a fixed figure
A buyer costs what they cost. You know the amount before the year begins, you can budget for it, and it doesn’t change because it happens to be a busy quarter. If you want to know what that person will cost you next year, you look at one line in the budget.
An AI system works differently. You pay per user and often (partly) per query as well (per token). The bill grows with every query that needs extra tokens. And remember, AI doesn’t stop at five o’clock.
That difference sounds technical, but it isn’t. It’s the difference between a cost you can plan for and a cost you discover afterwards.
Why that bill keeps growing
With most software, you buy licences and that’s the end of it. AI is different, and it pays to understand why, because this is exactly the kind of cost structure procurement normally isn’t happy with.
With the newer systems that complete a task on their own, you also pay for the steps you never get to see. The intermediate steps, the attempts that lead nowhere, the rereading of the same documents in every round. You’re not buying a product with a price tag, you’re buying a meter that keeps running.
The costs that never appear on a quote
On top of that comes everything that never makes it into the business case.
Someone has to set up the system and maintain it. Someone has to check whether the outputs are correct, and in procurement that’s no formality: a wrongly summarised clause isn’t a typo, it’s a contract risk. Someone has to decide what the system may do on its own and what it may not, and record that so it can still be explained a year from now.
That work comes on top of the existing work. And it almost always lands with the same person.
What the research actually says
At this point it’s wise to get the figures out and read them carefully, because they’re often quoted more starkly than they really are.
MIT published The GenAI Divide: State of AI in Business, a study of companies that had adopted AI. The finding that stuck: 95 percent of the pilots delivered no measurable impact on the profit and loss account.
Note the wording, because it matters. It doesn’t say 95 percent failed. It says the impact didn’t show up in the figures. That’s something different, and for this discussion actually more relevant: the technology often did exactly what was promised, but the difference stayed invisible in the place where senior management looks.
Three figures, caveats included
Two other studies complete the picture, and I’ll add the caveats you rarely see in the LinkedIn version of these figures.
IDC found that 96 percent of organisations rolling out generative AI find the costs come in higher or much higher than expected. That study was paid for by a software vendor, so read it with that in mind — but the direction matches everything else.
Gartner expects at least half of generative AI projects to go over budget by 2028, because of poor architecture choices and too little experience of running them afterwards. For systems that operate on their own, Gartner predicts that more than forty percent of projects will be cancelled before the end of 2027.
And from the study I quoted earlier: of all the organisations that experimented in 2024, five percent got as far as a real implementation.
Why the pilot succeeds and the rollout stalls
That last figure explains most of it, especially alongside the fact that 82 percent of organisations want to be working with these systems within three years. That figure comes from Capgemini research among 1,100 executives at organisations with more than a billion dollars in revenue.
A pilot runs on selected data, with an enthusiastic colleague alongside it, and with nothing riding on it. A rollout runs on the data you actually have: supplier names spelt six different ways, contracts sitting in a folder as scanned PDFs, item numbers that differ from site to site, and a framework agreement that half the organisation buys outside of.
Three quarters of procurement managers say themselves that their data isn’t ready yet. That isn’t a software problem, it won’t solve itself, and it takes people to solve it.
The work that really is disappearing
I don’t want to claim the opposite here. Work is disappearing, and that’s fine.
Requesting quotes and comparing them side by side. Combing through spending patterns in search of a category worth the effort. Reducing a forty-page contract to the five points that matter. Spotting that a supplier has delivered late three quarters in a row.
That was the work you used to do as a junior. And to be honest: there are more enjoyable things to do.
The work that remains is more demanding
But then look at what remains, because that’s where the sum flips.
Negotiating with a supplier who won’t budge. Repairing a relationship after things have gone wrong. Translating a strategy into choices that cost money and that someone is going to say no to. And seeing what no dashboard shows: that a supplier is quietly in trouble, that a clause works out differently than intended, that the question the business is asking isn’t the question it means.
You don’t get that from data. It means the role you keep is more demanding than the role you cut, not less — and so you can’t fill it with a junior either.
The second bill: the knowledge that walked out too
There’s one more cost, and it isn’t in any budget.
When you let an experienced buyer go, knowledge leaves with them that was never recorded anywhere. Which supplier you call when things go wrong. Which clause always turns out worse than expected in this category. Why that one framework agreement is structured the way it is, and what arrangement it was built on eight years ago.
An AI system doesn’t take that over, because it never had it. It isn’t in any document you can load into it. We wrote earlier about the ageing of this profession; this is the same bill, just presented years earlier and by your own choice.
Why companies end up recruiting again
The pattern I’m seeing now is this: cut one year, recruited again the next. Often at a higher rate, because by then the role has become more demanding and the market tighter.
That’s the most expensive route to exactly the same place. You pay back the saving, plus the recruitment costs, plus half a year in which someone doesn’t yet know the organisation. And in the meantime you’ve lost the knowledge you can’t buy back.
This is not an argument against AI
I want to be clear about that. The organisations that get this right use AI extensively — more than the organisations that hesitate.
The difference is that they use it to let their people do more, not to have fewer of them. They work out the costs based on the usage they expect a year from now, not the usage in the demo. And they invest in two things at once: the technology and the people who have to work with it.
Four questions before you cut a role
What will this system cost at the usage we expect a year from now? Not at today’s usage, and not at the usage in the pilot set-up.
Who checks the outputs, how much time does that take, and is that time budgeted anywhere? If the answer is “we’ll just fit it in”, it isn’t budgeted.
What knowledge exists only in the head of the person we’re not replacing? And where do we record it before the answer no longer matters?
What will it cost if we have to reverse this in eighteen months? Including recruitment, including onboarding time, including the going rate at that point.
Conclusion
AI doesn’t make the buyer redundant. It makes routine work cheaper and the remaining work more demanding, and that’s something very different from a saving on the wage bill. Anyone who confuses the two is cutting a role based on a sum that hasn’t been worked out.
So don’t base the sum on the salary you save, but on the bill that takes its place. Do that up front, and you won’t need to reopen that vacancy a year from now.
