Running an AI automation costs money every month, and most proposals do not tell you how much. There are four line items and only one of them is the model.
The four things you actually pay for
-
Model calls
What the AI itself costs per run. For document extraction and classification on business-API tiers this is usually pennies per document, and it scales linearly with volume. It is rarely the biggest number.
-
Hosting
Where the workflow runs. Self-hosted n8n on a modest UK or EU instance covers a lot of automations before it needs anything larger. This is a fixed monthly cost that does not move much with volume.
-
Monitoring and alerting
Somewhere collecting failures and telling a human. Cheap, and the line item people cut first and regret. A system nobody is watching fails quietly.
-
The hour a month someone spends
Reading the report, clearing the exception queue, noticing that a supplier changed their invoice layout. This is real work and it does not go away. If a proposal claims it does, the proposal is wrong.
Why the model is rarely the expensive part
People assume the AI is the cost. For most business automations it is not, because the volumes are small by AI standards. Four hundred invoices a month is a trivial workload for a model and a meaningful workload for a person. That asymmetry is the entire reason this work pays back.
Where model cost does bite is high-volume customer-facing agents answering thousands of conversations. There the arithmetic changes and it is worth choosing models per task rather than using the largest one for everything.
Costs that appear in year two
Suppliers change their document formats. APIs deprecate endpoints. A system you connected to gets replaced. None of this is dramatic, but it is why an automation needs an owner rather than just a launch. Budget for a few hours a year of maintenance per workflow and you will not be surprised.
What to ask before you sign
Ask for the monthly running cost in writing, separated into hosting, model usage and monitoring. Ask who gets the alert when it fails. Ask what happens to the workflows if you stop working with the supplier. If the answer to that last one is that they live inside a platform you cannot export, you are renting, not owning.
We run most builds on self-hosted n8n in a UK or EU region precisely so the workflows are exportable and documented. See how we use n8n.
Questions we get asked
Is running cost included in the build price?
It should be quoted separately and clearly. A build price with no stated running cost is an incomplete quote. Ask what year two looks like before you sign anything.
Can we host it ourselves?
Yes. Workflows built on n8n are exportable and documented, so you can take them in house or to another supplier. We would rather you were free to leave than locked in.
Does the cost go up as we grow?
Model usage scales with volume; hosting and monitoring largely do not until you are running a lot. For most SME workloads doubling the volume does not double the bill.
What is the cheapest way to run an automation?
Choosing the smallest model that does the job reliably, batching work where latency does not matter, and not paying for uptime guarantees you do not need. A queue for a human is usually cheaper and safer than a system engineered never to stop.