Who should own AI agent costs?
Natalia PletnevaCo-Founder and Commercial Manager, Inlight ITAn AI agent can turn a twenty-minute task into a few minutes of review. Its costs, though, can sit across licences, credits, model calls and cloud services, while the time saved shows up in another team.
Who should own the overall cost? Usually the business owner of the process the agent supports, with IT and finance helping to bring the figures together. Start with the cost of each completed task, then compare it with the time saved and the quality of the result.
- Microsoft 365 Copilot licenceFixed
- Copilot Credits4.0
- Model calls (Microsoft Foundry)3.0
- Power Automate runs2.0
- Retrieval and cloud services1.5
- Retries and re-runs0.5
Runs when needed, on a light path, with a limit. Consumption is a small part of the value created.
Which charges apply depends on how the agent is built, licensed and used. Some usage is included, and other services are billed separately. See Microsoft’s Copilot Studio billing guide.
An AI agent’s costs can sit with five different people
Licences, credits, model calls, automation runs and cloud services can each land with a different person. Each sees an accurate slice. None of them sees what a finished piece of work costs, so the question of whether the agent is worth it has no clear owner.
The figures usually exist already. They sit in the Microsoft 365 admin centre, the Power Platform admin centre, the Azure portal and the finance system. What’s missing is one person who joins them up for each agent.
Build decisions shape the AI bill
When an agent runs, which model it uses, how much it reads and how it handles a failure are build decisions, and they drive its variable cost. Here’s one service request built by two different teams on the same licences and platform.
- Runs when a case arrives, and only then
- Uses a lighter model for routine cases and keeps the premium model for exceptions
- Reads the customer record, the latest emails and the one document that matters
- Retries a failed action twice, then stops and flags it
- The person twenty minutes of handling becomes three minutes of review
- Runs every few minutes, whether or not there’s work waiting
- Sends each case to the most expensive model
- Rereads the full document set on each pass
- Keeps retrying failed actions until something gives
- The person staff still spend fifteen minutes checking the output
Both builds work, but they don’t perform the same. In this worked comparison, the four design choices raise variable consumption about 47 times per completed task, and checking each result takes staff fifteen minutes, up from three. The extra review time is visible to staff. The consumption cost can stay hidden until someone joins up the bills. The figures are relative units built to show how the choices compound. They aren’t Microsoft prices or a client benchmark, and licence costs and staff time sit outside the 47×.
How to calculate the cost per completed task
Add up what the agent consumed over a period and divide it by the tasks it completed to the agreed standard. Then put that figure next to the effort each task used to take. The four levels below show where it sits relative to what the platform reports.
- Count failed attempts and retries in the cost.
- Define “completed” before you start: an invoice accepted after review counts, a draft that still needs rework doesn’t.
- Agree how shared licences are split across agents, and keep the split the same each month.
- Track human review and rework time next to the agent cost.
For exampleA$600 of agent costs in a month ÷ 300 accepted tasks = A$2 per completed task, before review time.
What the task is worth
The handling time, delay or outside cost the agent removes, and the quality of what’s left. Count the minutes someone still spends checking the output, because they decide whether the saving is real. An agent that costs more this month may be doing far more useful work. A cheap one can still be poor value if it creates more review than the process it replaced. Time saved frees capacity, and it only becomes a cash saving when the business uses that capacity differently.
Cost per completed task
The full cost of one finished piece of work: a resolved request, a processed invoice, an onboarding pack. This is the figure you can put next to a salary or a supplier invoice.
Cost per run
What one run of the agent costs once model calls, actions and supporting services are included. Run frequency, document size and retries multiply here, which is why an agent behaves differently at a few hundred runs a month than it did in the pilot.
Tokens and credits
The raw units of consumption. Useful when you’re choosing models and platforms, and the place most reporting starts, but only the first step towards the figure above.
Four things the owner settles before an agent goes into daily use
Start with one agent and a small set of measures. These are much easier to settle while the agent does one job for one team than after a whole department relies on it.
When should it run?
Where possible, trigger the agent when work arrives. An agent that checks every few minutes creates consumption whether or not anything has changed.
Which model does the routine work?
Test whether a lighter model handles routine cases reliably, and keep the premium model and the wider document search for the exceptions that need them.
Where does it stop?
Agree a consumption limit and clear rules for failures and retries. Alerts and stop conditions protect the budget as much as the service, and retries are the cost line that rarely makes it into a forecast.
What counts as a good result?
Pick a unit the business already understands: requests resolved, documents processed, hours of handling removed. A rising bill with no output figure beside it will look like a problem even when the agent is doing well.
Start with one agent that’s live or nearly live, and leave the rest of your AI work for later.
- Name the ownerUsually the person who runs the process the agent serves.
- Write down how the process runs todayThe handling time, delays and rework involved now.
- List what the agent draws onLicences, credits, model calls, flows, cloud services and any third-party APIs.
- Run it for a month and measureThe calculation above, next to the time or effort removed.
- Scale it or rebuild itScale the agents with strong economics. Rebuild the ones where consumption is growing faster than the value.
Once someone owns the figure, a bigger AI bill is easy to read
AI consumption will rise as agents take on more work, and that’s the point of it. A rising bill only causes alarm when the work behind it can’t be seen. Give each agent an owner and a cost per task, and the same bill tells a clear story: which agents are paying their way, which ones to scale and which ones to rebuild.
One owner for each agent
One measure, agreed up front
Limits set before it scales
Natalia brings more than 20 years of finance, planning and operating experience to the commercial side of technology decisions: what the options are, what they really cost and where the investment is worth it.
Natalia Pletneva on LinkedInFind out what each of your AI agents really costs
Bring one agent that’s live or close to it, and we’ll map its consumption, owner and limits with you.
Discuss an AI workload