Adding a person to the payroll is one of the most scrutinised decisions a business makes. There’s a business case. A budget line. Someone signs it off. There’s a probation period, a review cycle, and a fairly honest internal conversation if the role isn’t working. Payroll is the largest cost in most businesses and it’s treated accordingly.

Now consider how AI tools enter the same business. Someone signs up with a company card. A team adopts a second tool because the first one didn’t quite do it. A developer starts calling an API and the bill scales with usage. Someone leaves and their seat carries on renewing.

Within a year there are twenty subscriptions across four currencies, and no one person could name them all.

For some technology businesses it is already one of the largest software costs. It is being managed with a fraction of the discipline.

Why it slips through

It isn’t negligence. It’s three ordinary things happening at once.

The individual amounts look small. A £30 seat doesn’t trigger anyone’s attention, but forty of them is £14,400 a year.

It’s distributed by design. AI tools spread because they’re genuinely useful and easy to adopt - that’s the point of them. But nothing in that process routes back to a single owner.

And it doesn’t behave like a subscription. Usage-based costs move with customer activity and with product decisions. A change to how something is built can shift the bill materially, and it shows up after the fact rather than at the point the decision was made.

What rigour actually looks like

You don’t need a procurement function. You need the same four things you’d apply to a hire.

  • One list, one owner. Every AI tool, what it costs, who sponsors it, when it renews. If nobody can produce this in five minutes, that’s the first job.
  • A unit number, not just a total. Cost per employee, or per active user, or per customer served. A total tells you what you spent; a unit number tells you whether it’s working, and it belongs in your management accounts alongside gross margin.
  • A review cadence. Quarterly is enough. Which tools are being used, by how many people, and what would break if you cancelled. Most businesses find something they’re paying for that nobody has opened in months.
  • An off-boarding step. When someone leaves, their seats go with them. In the month-end reconciliations we run, it’s one of the most common sources of quiet waste.

Isn’t this just subscription management?

For the seat-based tools, partly - and a spreadsheet handles those. What a subscription tracker cannot see is the usage-based half. An API bill isn’t a seat; it’s a cost of delivery that moves when your product changes and when your customers do.

That half is why this sits with finance rather than with office admin. It behaves like cost of goods, not like software.

Where finance comes in

This is a finance question, not an IT one. The point of measuring AI spend isn’t tidiness. It feeds directly into the two things a growing business actually needs to know: what your runway is, and what your product genuinely costs to deliver.

If your cost of delivery is partly inference and you can’t see the inference, you don’t know your margin. You know an approximation of it.

A disclosure that is also the credential: Misia.AI, a cost analytics platform that puts a single number on a company’s AI spend, shares its directors with Ozria. We see this problem from both sides - as the people making the spending decisions, and as the people reconciling them at month-end.

Misia.AI answers what your AI spend is. We answer what it means for your runway. Most companies are currently asking neither question.

Ozria Consulting runs the finance function for SMEs and high-growth technology businesses. If your numbers have stopped keeping up with what you’re building, book a discovery call.