What does AI actually cost you, per team?

You can probably find your total AI spend in about a minute. It is on an invoice.

Now try to answer any of the questions that follow from it. Which product line drove the increase last month. Whether the assistant your customers use costs more or less than the internal summarising job nobody talks about. Which team would notice if you halved their budget. Whether the growth is more usage or more expensive usage.

Most organisations cannot answer those, and the reason is structural rather than lazy. The invoice is per provider account. The work is per team, per product, per feature. Nothing in the middle attributes one to the other, so the only number anybody has is the one that arrives monthly, already spent.

Attribution belongs where the calls are

Every one of those calls already passes through something on its way out. That is the natural place to count them, because it is the only place that knows both the cost and the caller. Squidder counts them there, so the questions above become answerable without instrumenting a single service:

app                    team          model tier     30d      Δ
support-assistant      customer ops  large        $18,410   +6%
ticket-summariser      customer ops  small         $9,733  +240%
catalogue-enrich       commerce      small         $4,102   -2%
contract-review        legal         large         $2,890  +11%

The second row is the point. Nobody was thinking about the summariser, it is not customer-facing, and it is the fastest-growing line on the page.

Once you can attribute, you can cap: a ceiling on an application that is allowed to grow, a tighter one on an experiment that should not become a line item by accident, a number you can point at before it becomes an invoice rather than after.

The uses that actually come up

Chargeback that survives scrutiny. Telling a team what their AI usage costs only works if they believe the number. Attribution at the point of the call is a number they can check.

Catching the runaway early. Cost incidents are rarely a slow trend. They are a retry loop, a prompt that grew, a job someone scheduled hourly instead of daily. A ceiling turns that from a monthly surprise into an alert.

Seeing the cost of your own controls. Inspecting and classifying traffic is work, and work has a price. It should be visible rather than folded into a total, so you can decide whether a given check is worth what it costs.

Attribution is also what turns a general worry into a decision. "We should use a cheaper model" is unanswerable; "the summariser is a fifth of the bill and nobody would notice" is a workload with a number attached, and a way to prove the cheaper model holds up.

Where this stops

It covers inference. Model calls have a cost signal: the work is measurable and priced. A tool call, an interface call, or a page load does not. So the ceilings and the attribution here are about what your AI spends, and a claim that we govern the cost of every call your software makes would not be true. Anyone selling you that is describing something else.

A ceiling is a blunt instrument. Stopping work when a budget is exhausted is correct for an experiment and dangerous for something customer-facing. Set them where you mean them, and treat the alert as the product rather than the stop.

Attribution is only as good as your boundaries. If four teams share one application identity, the report will tell you about the application. Getting per team numbers means the identities have to reflect the teams, and that is organisational work no product does for you.

Where to start

Do not begin with budgets. Begin with a week of attributed spend and read it.

In most organisations the first read produces one genuine surprise: a workload nobody was thinking about that turns out to be a third of the bill, or a customer-facing feature that costs far less than the internal tooling around it.

Act on that surprise first. Set the ceilings afterwards, when you know where they belong.