In January your software bill goes up. The increase is AI you never approved.

If your company renews software licences in January, the per user price is going up. On 3 September the most widely used sales software provider in the world reordered its catalogue into three editions costing $195, $395 and $550 per user per month. They are 11% to 13% more expensive than the ones they replace, and what explains the increase is that they now come with AI agents inside.

For a company with a hundred licences of the top edition, that jump is about $6,300 more per month, close to $76,000 a year. Nobody is going to ask you to approve that spend, because it does not arrive as an AI project: it arrives inside the usual renewal.

Because whoever builds the data centres where AI runs stopped paying for them out of their own pocket, and is looking for somebody else to pay. In September, one data centre was financed with a loan at 8.25% interest. That is a high rate: the kind charged when the lender sees risk. Five days later a similar one was put together.

That same month, a large cloud provider reported it had invested $28.5 billion in a single quarter and closed that quarter spending more cash than came in. It also explained how it plans to sustain that: charging up front, or building with hardware the customer buys. That money does not disappear. It changes pockets. And the pocket it ends up in is the bill for the software you already use every day.

What got automated this year was preparing: the report, the reconciliation, the first draft of the contract. Reviewing did not. So there is more to review and the same people reviewing it, with less time for each one. Along the way an old rule broke, one that has nothing to do with AI: whoever prepares a document should not be the one who approves it. When the system prepares, the person who used to review now signs, and signs faster.

The function that should be reviewing this says itself that it cannot keep up. In this year's European internal audit survey, AI is the second risk on the list and sits in the top five for 52% of teams. Worldwide, only one in nine says it has it properly covered. And time does not help. In a survey of 160 finance directors, the simple tasks already automated take nine to ten months to show a saving. That is longer than the year in which somebody promised it.

In a single month, three different places asked for the same thing. On 30 September California passed a law: if a dismissal rests mainly on an automated system, a person has to review it and there has to be written notice. A New York court vacated a student sanction because it rested only on what an AI detector said. And in Peru the deadline to explain how systems decide expired on 10 September.

What the three of them ask is not that the system be right. It is that somebody can account for it afterwards, in front of a third party: on what data, which version was running, who authorised it and who reviewed it before it affected anyone. That cannot be bought ready made. No vendor is going to certify the evidence of its own system, for the same reason nobody audits their own books.

This season's way of selling is to lend engineers. The vendor sends people to work inside your company, builds custom agents for you on its own platform, and closes the sale with that. When those people leave, what they built is left with nobody who understands it. An analyst firm predicted on 29 September that by 2028 most companies will abandon those builds. We report it as direction and not as data, because it is a prediction with no survey behind it and our rule does not allow a percentage next to it.

The reason, on the other hand, explains itself: whoever builds to close a sale is not building for somebody else to operate. In July there was already a concrete case. Some agents escaped the test environment they were in and the company affected had to rebuild close to a third of its infrastructure. The lawsuit was filed on 30 September.

Four figures are going around with nothing behind them, and all four sound good on a slide. That 70% of companies abandon their agents, quoted as if somebody had measured it, when it is a prediction. The first European fines of 47 million euros, which appear in no official document.

The 10,000 security incidents involving agents. And the 40% market share one vendor claims for itself. If any of them reaches your committee, ask for the source before the conclusion. None of the four is used in this report.