Short answer: AI running costs are what a system bills you every month once it is live, and hardly anyone quotes them. The build price is a single figure with a signature under it. The running number arrives in pieces, from five or six vendors, and it moves with use.
This is the other half of the cost question. The buying side is covered in how much business automation costs and what an AI consultant costs. This piece is about the bill that keeps arriving after everyone has gone home.
Why the build price is the wrong number to negotiate
A quote for an AI build reads like a quote for a kitchen. One figure, one delivery date, one signature. The comparison breaks the morning it goes live, because a kitchen does not invoice you every time someone cooks in it.
An AI system does. Every call it answers costs money. Every document it reads costs money. Every automation run costs money. Each unit is small enough to ignore during a demo and big enough to matter at volume, which is the exact shape of cost that ambushes a finance team a year later.
There is a second reason to care, and it is political rather than financial. A system whose running cost nobody tracks is a system nobody can defend at budget time. When the question comes, the team that can say what the thing costs per booked call keeps it. The team that shrugs loses it, whether or not it was working.
The six lines on an AI bill
Take a live system apart and the monthly spend lands in the same six places. Not every system carries all six. Every system carries at least three.
1. Model usage, priced per token
Text going in and text coming out, metered. OpenAI and Anthropic both publish current rates on a public page, and both price input and output separately, with output the dearer of the two. Read the number on the day you build the budget rather than trusting one you saw last quarter. These prices move, usually downward, and a stale figure in a spreadsheet is worse than no figure at all.
The trap here is not the rate. It is the volume. A support assistant that reads a full order history before answering spends far more per reply than one handed the three fields it needs. Same model, same question, a bill that differs by an order of magnitude.
2. Voice minutes
A phone agent stacks three meters on top of each other. There is the telephony leg, which carries the call itself and is priced per connected minute by a carrier such as Twilio. There is speech, where a voice vendor like ElevenLabs charges for what the agent says and hears. And there is the orchestration layer, Vapi or a competitor, which charges for holding the conversation together.
Three meters means three invoices for one phone call, and the cost per minute is the sum. We compared the vendors themselves in the voice stack breakdown. For budgeting, the only figure that matters is the blended cost of one finished call, and you will not find it on any vendor page. You get it by running a hundred real calls and dividing.
3. Automation platform operations
Make and n8n both meter work rather than time. The unit is an operation or an execution, and a workflow that looks like one step in the editor often costs several once it runs. Self hosting an open source engine moves the money from a licence to a server plus somebody's attention, which is cheaper on paper and not always cheaper in practice. That trade is the whole subject of the Make, n8n and Zapier comparison.
4. Hosting, storage and the database
Usually the smallest line and the easiest to forecast. An app on Vercel or an equivalent platform, a managed database, an object store for files. The part that grows quietly is anything holding embeddings, because a knowledge base that gets fed every new document forever will keep growing forever unless somebody decides what to delete.
5. Monitoring, logging and evaluation
The line teams cut first and regret second. Without call recordings, run logs and a stored sample of outputs, you cannot answer the two questions that decide the system's future: is it still doing the job, and what does one good outcome cost? A system you cannot see is a system you cannot defend, and it is also the reason so many pilots die quietly. We wrote about that pattern in why AI pilots fail to reach production.
6. The humans who handle what the system does not
The biggest line on most bills, and the one that never appears on a vendor invoice. Somebody reads the exceptions. Somebody checks the calls that ended badly. Somebody updates the answers when the price list changes.
Put a number on it, because it is measurable. Statistics Estonia puts Tallinn's average monthly gross wage at 2,538 euros in Q1 2026. The Estonian Tax and Customs Board adds 33% social tax and 0.8% unemployment insurance on top, so the employer pays roughly 3,396 euros a month for one person, about 40,750 euros a year. An hour a day of exception handling is not a rounding error. It is a real fraction of a real salary, and it belongs in the running cost next to the token spend.
Three ways the bill surprises people
Retries
A failed call that reconnects, a workflow that reruns after a timeout, a model request that gets sent twice because the first response looked wrong. Every retry is billed like an original. Systems with weak error handling can spend a quarter of their budget doing the same work more than once, and none of it shows up as a separate line anywhere.
Context growth
Month one, the assistant reads a short prompt and a customer record. Month six, somebody has added a policy document, a product catalogue and the last twenty messages of history, because each addition made one answer better. The per-request cost has quietly tripled and no single change caused it.
Success
The happiest surprise and still a surprise. Usage based pricing means a system that works costs more than a system nobody uses. That is fine when the unit economics hold, which is why the number to watch is cost per booked call or cost per resolved ticket rather than the monthly total. A rising total with a falling unit cost is a business working correctly.
Four questions to ask before you sign
Ask them in writing, and ask them before the build rather than after. A supplier who cannot answer these has not run a system at your volume.
- What will this cost per month at our current volume, and at triple it? Two numbers, not one. The gap between them is the shape of the pricing.
- Which accounts are in our name? Model keys, telephony, automation platform, hosting. Accounts in the supplier's name mean you cannot see the real cost and cannot leave without a rebuild.
- What happens to the bill when something breaks at 2am? Retry behaviour, spend caps, and who gets alerted.
- How many hours a week of our people does this need? If the answer is zero, the supplier has not run one of these before.
How to instrument it in the first month
You do not need a cost dashboard. You need four numbers, counted the same way every week, starting the week you go live.
Count the volume the system handled, the total it spent across every vendor, the human minutes spent cleaning up after it, and the outcomes it produced. Divide the second and third by the fourth. That gives you cost per outcome, which is the only figure worth taking into a budget meeting.
Set a spend alert on every metered account on day one, not after the first bad invoice. Most vendors offer them and most teams enable them late, usually the week after they needed them.
Then leave it alone for a month. Changing three things at once and reading the bill afterwards tells you nothing about which change moved it.
When the running cost says stop
Sometimes the honest answer is that the system should not exist. If a workflow runs 40 times a month and the automation costs more per run than the person it replaced, the maths does not care how good the demo was. Kill it and spend the money where the volume is.
This is rarer than it sounds. Eurostat's isoc_eb_ai dataset put AI use among EU enterprises with ten or more employees at 20.0% in 2025, up from 13.5% the year before. Most of the failures inside that jump are not cost failures. They are systems built on a workflow nobody had fixed first, which is a different problem with a cheaper solution.
Where we put this number
At kratt the running cost sits in the audit, before anything gets built and before anyone is asked for money. The audit is free, it names where the business is losing money today, and it is yours whether or not you hire us. When we do build, we host and run the system on accounts in the client's name, so the monthly figure is visible to the person paying it. You can see what that looks like in our case studies, or read how we work on the AI consultancy page.
What are AI running costs?
AI running costs are the recurring monthly spend of an AI system after launch: model usage priced per token, voice minutes across telephony and speech vendors, automation platform operations, hosting and storage, monitoring, and the human time spent handling exceptions. Build quotes almost never include them.
How much does it cost to run an AI agent per month?
It depends on volume, not on the agent. The honest way to answer is to price one unit of work, one call answered or one ticket resolved, then multiply by your real monthly volume. Any supplier quoting a flat monthly figure without knowing your volume is guessing.
Which line is usually the biggest?
People, on most systems. Model and platform fees are visible and often modest. The time your staff spend reviewing exceptions, correcting outputs and updating answers is invisible and rarely counted, and at Tallinn wage levels an hour a day is a meaningful share of a salary.
Do running costs go down over time?
Model prices have fallen steadily, which helps. Your own usage usually rises faster, because working systems get used more and get given more context. Expect the per unit cost to fall and the total to rise, and track the per unit number.
Should the supplier own the vendor accounts?
No. Keep model keys, telephony, automation and hosting in your own name and give the supplier access. It costs nothing to set up, makes the real spend visible, and means changing suppliers is a handover rather than a rebuild.
If you want the running number for your own workflows before anyone quotes you a build, that is what the free audit produces. Tell us what the workflow is and we will price the work it does, not the software that does it.