Your Spend Control Fires Too Late

  1. A SPEND CONTROL THAT FIRES TOO LATE 1 2 3 Order Spend Invoice Prior approval, found at invoice. After the money is committed. Taskeen taskeen.leapora.tech 1 / 4 (c) Leapora AI, Inc. All rights reserved. Siva Kowsika.
  2. WHERE THE SPEND CONTROL EXPECTED TO FIRE It assumed a person raising a requisition AS WRITTEN requisition, then approval, then spend CONSUMPTION TOOLING usage accrues against an account THE RESULT no requisition, so no threshold, so no control Taskeen taskeen.leapora.tech 2 / 4 (c) Leapora AI, Inc. All rights reserved. Siva Kowsika.
  3. WHERE A RULE HAS TO BE ENFORCED A rule enforced where the evidence arrives, rather than where the act happens, does not control anything. It reports. Taskeen taskeen.leapora.tech 3 / 4 (c) Leapora AI, Inc. All rights reserved. Siva Kowsika.
  4. WHY DELEGATION LIMITS STOPPED BITING Your delegation limits are intact and increasingly beside the point. Agentic tooling generates acts that never pass the desk your control is sitting on. Taskeen taskeen.leapora.tech 4 / 4 (c) Leapora AI, Inc. All rights reserved. Siva Kowsika.

Your policy says purchases above a threshold need prior approval. Finance discovers they did not get it at invoice stage, which is to say after the money is committed and the control has already failed.

The policy is fine. The enforcement point is in the wrong place, and it has been in the wrong place for so long that the quarterly exception list reads as normal.

The assumption nobody wrote down

The threshold rule rests on something that was never stated because it never needed to be: that a purchase begins with a person deciding to buy something and raising a requisition.

Everything downstream inherits it. The approval matrix assumes a requisition to approve. The delegation limits assume an amount known in advance. The exception report assumes a comparison between what was approved and what was spent, which assumes both numbers exist at the same moment.

None of that is naive. It describes procurement accurately for about forty years. A rule written against a stable reality does not announce the day the reality moves.

Why consumption spend never crosses the threshold

Consumption-based AI tooling does not begin with a decision to buy. It begins as usage.

It accrues against an account, and it becomes a purchase in the only place purchases get noticed, which is the invoice. No requisition is ever raised, so no threshold is ever crossed, so the control that was supposed to fire never gets the chance.

The amount is not knowable in advance either, which breaks the delegation limit in a second way. A person approving a commitment of up to fifty thousand is making a judgement about a number. A person switching on a per-token service is making a judgement about a rate, and the two are not the same act however similar the approval screen looks.

Your delegation limits are intact and increasingly beside the point. An approval matrix only governs spend that announces itself.

The pattern generalises past procurement

A rule enforced where the evidence arrives, rather than where the act happens, does not control anything. It reports.

That distinction is worth holding on to, because once you have it you start seeing the same shape elsewhere. Access reviews that run quarterly control nothing about the eleven weeks in between; they report on them. A data classification check at export time does not govern what was collected. A model usage policy enforced at the monthly bill does not govern what the model was asked to do.

In each case the control was placed where somebody could see the consequence, because that was the cheapest place to put a check. And in each case agentic tooling is now generating acts that never pass the desk the control is sitting on.

The conversation in March

Finance brings a number to a quarterly review. Spend on a tooling line is four times what was modelled, and nobody in the room approved any of it.

What follows is familiar. Somebody is asked who owns it, and the honest answer is that four teams use it. Somebody asks when it was approved, and the honest answer is that it was never a purchase, it was a key. Somebody asks whether it can be turned off, and the honest answer is that two of the four teams have something in production behind it.

Nobody in that room did anything wrong. The spend happened through a sequence of individually reasonable decisions, each one below any threshold that existed, by people who had no way of knowing what the others were doing.

The outcome is almost always the same: a new committee, and a request that teams declare their tooling. Which is a control placed where the evidence arrives, again, one level up.

Moving the control to the act

A commitment to spend is accepted by a named person at the point the consumption starts, and anything that begins without one surfaces then, while it is still a question rather than an invoice.

The objection is obvious and fair: that sounds like friction on exactly the experimentation an organisation wants. It would be, if the acceptance were an approval queue. It is not. It is a named person taking the commitment, which can be instant, and the record of their having taken it, which is the part that matters in March when somebody asks.

The quarterly exception list stops being a list of things that already happened and becomes a list of things that were accepted, by someone, on purpose. Those are very different documents to take into an audit.

Questions

Why do AI tool costs bypass procurement approval?

Because the approval threshold assumes a purchase begins with a requisition, and consumption-based tooling begins as usage. It accrues against an account and becomes a purchase at the invoice. No requisition is raised, so no threshold is crossed, and the control never fires.

Why does a delegation limit not cover consumption spend?

A delegation limit is a judgement about an amount known in advance. Switching on a per-token service is a judgement about a rate, with the total unknown until it has been incurred. The approval screen looks the same; the act being authorised is not.

Where should a spend control be enforced?

At the act, not at the evidence. A control placed where the consequence becomes visible does not govern anything, it reports on it. Accept the commitment at the point consumption starts, and anything beginning without an acceptance surfaces while it is still a question.

Does moving controls to the act add friction?

Only if the acceptance is an approval queue. A named person taking the commitment can be instant. What persists is the record that they took it, which is the part that matters when the quarterly exception list is read in an audit.

Taskeen records who was told, who acknowledged, what they did, and what proves it. See how it works.