A software licence has a ceiling. It is in the contract, it is in the budget, and if somebody needs more they have to get a signature. Your whole apparatus is built for that: requisition, approval, purchase order, invoice. Spend on artificial intelligence arises differently. It arises per transaction, in fractions of a cent, without a signature, and it grows with use. Nobody sees it until the month closes.
In most companies this consumption currently runs through an existing cloud framework agreement or a credit card held by IT. Neither is wrong. Both mean that a growing spend position bypasses exactly the function whose job it is to order spend. This is not an IT problem with procurement involvement. It is a procurement problem currently parked in IT.
Why this lands on your desk
Because the position ends up in your cost centre, and because you are the one who gets asked once it becomes uncomfortable. A head of procurement who can name three figures at the first query from finance — what was consumed, for what, and what a single transaction costs — is in a different position from one who needs a week to understand the bill.
It is also one of the few areas in which procurement does not have to be the latecomer. The patterns are familiar; they just have new names: buying outside the contract, no bundling, no chargeback to the department causing it. That is your trade. What is new is only the speed — what used to build up over months through individual orders now builds up over a weekend.
How unplanned cost actually arises
Not through expensive requests. Through many of them. Five patterns recur, and none of them is an operator error.
The test run nobody stopped. An automated routine is set up to try something out, keeps running hourly, and from Friday evening processes nobody but keeps processing everything. On Monday there is a number.
The long context. In a multi-turn conversation the history so far is sent again with every question. The twentieth exchange therefore costs a multiple of the first, without the benefit rising accordingly.
Re-indexing the whole corpus. A search system over your own documents has to be rebuilt when the structure or the model changes. That is a one-off cost, but a substantial one — and it happens more often than the plan assumes.
The jump from pilot to operation. A trial with five people and two hundred transactions a month is extended to eighty people. Consumption does not rise by a factor of sixteen but by more, because live operation also brings the cases nobody tried in the pilot.
The agent that calls itself. A system that hands sub-tasks to itself has no natural end without an explicit stopping condition. This is rare, and when it happens it is expensive.
Seven points that put a lid on it
None of this is elaborate. Together they cost a day and take effect immediately.
One: a spend limit before the first transaction runs. Every major platform supports hard ceilings per project or account plus alerts at fifty, eighty and a hundred per cent. That limit is set before anyone tries anything, not afterwards. It is the cheapest piece of advice in this entire text.
Two: a cost estimate per use case. Transactions per month times volume per transaction times two. The factor of two because input and output both count and because in practice more goes back and forth than planned. As a conversion, this is enough: roughly a hundred tokens correspond to about seventy German words, so an A4 page is broadly five to six hundred. The calculation takes ten minutes and eliminates half the ideas.
Three: its own spend category. Consumption-based AI needs a name in the system and an owner. As long as it rides along under „cloud“, nobody can steer it and nobody is responsible for it.
Four: chargeback to the department causing it. The cost belongs where the benefit is. Nothing disciplines the use of a tool as reliably as the experience of seeing it appear on your own cost centre.
Five: the contract side, before you bundle. Four points to check: from what volume does which tier apply, and is the tier assessed per account or per group? At what notice may the provider change prices? Do prepaid credits expire, and when? And what happens to a minimum commitment if you switch provider?
Six: minimum commitments only after three stable months. Committed-use discounts are attractive and, in the first quarter, the wrong decision, because nobody yet knows their consumption. Commit too early and you pay for capacity you do not draw, or you exceed it and pay twice.
Seven: no automated routine without a stopping condition. That line belongs in your approval rules, and it costs the vendor nothing. A vendor unwilling to commit to it has just told you something about the product.
Three figures for the monthly review
Spend per month, number of transactions processed, cost per transaction. No more is needed, and less is not enough. The third is the decisive one: it is the only value that stays comparable over time, and it shows immediately whether a use case is running out of hand or has simply grown.
If spend rises while cost per transaction holds, the company is doing more work — that is the good case. If cost per transaction rises, something in the construction has changed, and you want to know that before finance asks.
What is in it for you
You know a spend position that is currently growing unsupervised in many companies, and you know it first. You have a conversation with IT about numbers rather than about responsibility. You walk into the next negotiation with a documented consumption profile instead of an estimate, and that is the difference between a list price and a negotiated one. And when somebody asks in the autumn what this AI business actually costs, you have an answer with three figures rather than one with an excuse.
The rest is craft you already command. Somebody just has to start, and in this case you are the obvious person.
Where I come in
I set this structure up with you: the cost estimate per use case, the limits in the accounts, the checklist for the contracts, the chargeback and the monthly sheet with the three figures. In a mid-sized company that takes days, not weeks. I sell no licences and take no commission from any vendor, which occasionally leads to me recommending that you buy nothing at all.