AI cost management: know the cost per call or fly blind
Your AI provider's invoice won't tell you which workflow is spending the money. How to track AI costs per call, budget each app and weigh cost against value.
If you don't know your AI costs per call, you're flying blind. A lot of leadership teams only ever see the provider's monthly invoice, and then argue about the subscription, the seat count or a cap for the whole company. That total won't tell you which workflow is spending the money or whether it's worth it. At SUMAX every AI call goes through our own layer between our applications and the models and gets booked there with its cost. By now that's a five-figure number of calls every month. My position as COO: AI costs belong to the business unit that owns the workflow.
The monthly invoice hides where the money goes.
AI providers bill by usage, in tokens, the chunks of text a model reads and writes. At the end of the month you get one number, and it behaves like a phone bill with no itemized calls: it goes up or down and nobody knows why. So our layer tags every booked call to the application that made it. Broken down per call, the AI bill shows which workflow uses the money.
A call with no cost data isn't free. It stays visible as a gap, kept apart from known values and estimates, because mixing them fakes a certainty you don't have. Book every call with its cost and its application, and report gaps as gaps.
Budgets belong to the application.
A company-wide cap slows down the application that pays off just as much as the one wasting money. A budget per application makes spend predictable, and an overrun shows up before the invoice does. When an application keeps hitting its limit, either it's used more than planned and paying off, or it repeats steps it doesn't need to. Only per-call data tells you which. Give every application its own budget, and take a close look when it blows through it.
Broken down per call, the AI bill shows which workflow uses the money.
The cheap call is often the expensive one.
An expensive call that saves an hour of work costs less than a cheap one whose output someone has to rewrite from scratch. Pass on an AI result unchecked and you become a pass-through, and the cost comes back later as rework. Review, corrections and training all belong in the math.
Measure the benefit in hours per deliverable. If you count AI's value in headcount, you're measuring the wrong thing, because AI is there to take routine work off people's plates, not to cut jobs. So next to cost tracking you need time tracking, which shows how many hours a deliverable took before and after. Our editorial team now turns out a multiple of the expert articles per hour it managed in 2024. That number only supports a decision once the cost of the same work sits next to it. Put cost and hours per deliverable side by side before you decide to scale something up or shut it down.
Defaulting to the strongest model means paying for convenience.
Narrow tasks need different model capabilities than complex analysis, so picking a model is a cost decision. Running the strongest model on everything is convenient and expensive. A cheap model whose output the team keeps fixing is just as wrong. Because our layer keeps the model swappable, that mapping can change without anyone rebuilding an application. How much it saves depends on the workflow, and anyone who promises you a flat savings rate hasn't looked at yours. Give each task the smallest model that delivers the quality it needs.
The business side decides, IT measures.
High AI costs usually start in the workflow, and the business side knows the workflow best.
A hypothetical example: a content job goes through research, a draft and a review, and after feedback another draft follows. The last call shows only a small slice of the cost. Tie every call to the job and you see the full amount, and whether the feedback loop or the research is eating the money. Now the business side can act on it, for example with a tighter brief.
So the business side owns whether an application is worth the money. The technical team makes sure tracking and budgets are right, and finance sets the frame and sees the total and the breakdown. For the CFO, AI becomes a cost block like any other, traceable and budgeted. Put the analysis where the workflow is owned.
Five questions for your AI bill.
If you can't answer one of them, that's where you're flying blind.
- Can every AI call be traced to an application and a workflow?
- How many calls have no cost data, and do they show up as a gap in your reports?
- Does every application have its own budget that flags an overrun before the invoice arrives?
- Next to the cost of a deliverable, can you see the hours before and after, including review and rework?
- Is any task running on the strongest model when a smaller one would do, and can you switch models without rebuilding the application?
Start with the question you have the weakest answer to.
A handful of licenses needs no per-call tracking.
Tracking every call takes effort of its own. If a few people use AI tools now and then on flat-rate licenses, a monthly look at the invoices is enough. Per-call tracking pays off once several workflows run in production. At SUMAX most employees work with our own AI tools every month, so for us the question is settled. And tracking only measures what a call costs. Whether the output is any good is still for the business side to judge.
Without per-call costs, every AI decision is a guess. Pick the workflow you believe costs the most and tag every AI call it makes for one month. How to measure the value side is in Measuring AI ROI.