AI running costs sit on four lines, and only three of them ever reach an invoice. The licence is the part you agreed to, so it usually sits still. Usage charges and add-ons move with how much work you put through the tool. The fourth line, the hours your people spend checking output, never appears on a bill at all.

The four lines behind your AI running costs

Those four lines are worth writing out separately, because one number on an invoice tells you nothing you can act on.

The first line is the per-seat licence. You pay a set amount per person per month, so it moves only when you add or remove people. Of the four, it is the easiest to predict and the one most owners already know by heart.

The second line is usage. Plenty of tools charge for what you consume on top of the seat: messages, credits, minutes of transcription, pages processed. Usage tracks how hard your team leans on the tool. It can climb sharply in a busy month, even though nobody changed a setting.

The third line is add-ons and connectors. A connector joins the tool to your email, your file store or your accounts package. Suppliers often price each one separately from the seat. Connectors also arrive quietly, because the person setting the tool up adds one to make a job work and nobody logs it anywhere.

The cost that never reaches an invoice

The fourth line is the time your people spend checking output, and no supplier will ever bill you for it. That makes it the line to price yourself.

Ask the person doing the checking for the figure rather than guessing it. Then work it out the way you would price any other task. If checking takes 30 minutes a day, that is 2.5 hours a week, or about 120 hours a year. At £25 an hour, checking costs you roughly £3,000 a year, which can be more than the licence and the usage put together.

Then run the same sum for how long the work took before the tool arrived. Having both figures side by side tells you something the invoice cannot. It shows whether checking shrinks as people learn the job, or holds steady. Steady checking after a few months is worth a proper look, and measuring a soft win covers how to hold that comparison honestly.

What to put on the table each month

Now that all four lines exist on paper, the review itself is a short meeting with a fixed set of papers.

Bring last month’s invoice, split into the lines above. Bring the supplier’s usage report, which most tools produce on request. Bring a count of the work the tool finished: emails drafted, invoices coded, calls summarised, whatever the job actually is. Finally, bring a note of any change you made since the last review.

That is four documents and 30 minutes in the diary. The meeting stays short because the papers do the work. Keep them in the same place each month so you can lay this month beside last month without hunting for anything.

Who owns the monthly review

Papers alone change nothing, so name the person who brings them and keep that name the same each month.

The owner should be whoever runs the work the tool sits inside. Finance sees the invoice but not the output, and your IT partner sees the tool but not whether the output is any good. Only the person close to the work can tell you both. If nobody holds that role yet, who owns the AI job sets out what the job involves.

That person also needs authority to act. A review that can only report upwards turns into a monthly complaint about the bill.

Reading supplier usage against finished work

With an owner and the papers in place, you can make the comparison an invoice can never make on its own.

Divide the month’s total AI running costs by the volume of work the tool completed. If the tool cost £180 and handled 400 supplier invoices, each one cost about 45p to process. That figure is your unit cost, and its direction matters far more than its size.

Cost per item falling while volume rises is a tool settling in. Cost per item rising while volume stays flat means something changed in how people use it. Ask the supplier for the usage report if the tool does not show one. Many break it down by seat, which tells you who leans on the tool and who has quietly stopped. Check seats against people too, since a seat nobody opened last month is money you can stop paying today.

When is a rise acceptable?

The unit cost tells you which way the bill moved, but not whether the move is a problem. So agree in advance what makes a rise acceptable.

A rise is fine when the work it bought grew at least as fast. Paying 20 per cent more to get through 40 per cent more work is a good trade, and the unit cost proves it.

A rise with flat output is a signal, and usually one of three things sits behind it. People may be retrying the same request because the first answer disappoints them. The job may have crept into work it was never chosen for, which is worth catching early. Or the supplier may have changed the plan or its prices, which the invoice itself will show once you hold it beside last month’s.

None of those three needs a new tool. Each one needs a decision, and that is what the review exists to produce.

The four outcomes a review can produce

Each of those signals points somewhere, so end every review with one of four decisions, said out loud and written down.

Keep it as it is, when the numbers hold and nobody is complaining. That decision still counts, so record it. Change the process, when the tool is fine but the input, the instructions or the timing of the check is causing rework. Change the commercials, by dropping unused seats, moving plan or asking the supplier for a usage cap.

The fourth decision is to stop the job. Stopping is a normal outcome rather than a failure, and it protects the credibility of the next thing you try. Before you commit to anything bigger, read what a first AI job costs. Most surprises later trace back to a line nobody counted at the start.

You control AI running costs by knowing every month what you paid, what you got, and which of the four lines moved. So put the next review in the diary now, name the person who brings the four papers, and give it 30 minutes.