The monthly running costs of AI automation, and how to cap them

Every AI automation proposal leads with the build price. Almost none lead with the second number, the one you pay every month for as long as the system runs. The monthly running costs of AI automation are usually small, but "usually small" is not a budget line, and a system whose bill surprises the owner is a system that eventually gets switched off. Below is the full list, with 2026 numbers.

Four meters

• Automation platform. $20 to $80 a month, whether the workflows run on a hosted plan or a self-hosted instance. Volume tiers mean a busy system costs more than an idle one, but for most small and mid-sized businesses it stays under $80, and which end of the range you land on is mostly a Zapier versus Make versus n8n decision.
• Model usage is the variable one: $10 to $150 a month. A lead qualification flow handling a few hundred leads sits near $10. A document-heavy system that summarizes everything crossing an inbox can reach $150.
• Voice minutes, $0.10 to $0.20 each. Telephony plus real-time inference, metered per minute if the system answers or places calls. 500 minutes is $50 to $100, and this line only grows when the phone is ringing, which is the right direction for a cost to grow. It is the same meter that sits underneath what an AI voice agent costs to build and run.
• Hosting and storage. Databases, files, a small server for custom code, monitoring. $10 to $50, and individually tiny.

A single-workflow deployment lands at $40 to $150 a month all in. A multi-system deployment with voice can reach $300 to $500. Neither should be a surprise, because all four meters sit in dashboards you can open yourself. The running-cost calculator does this arithmetic for your own volumes, on the vendors' list prices. The document generator we run for a hydrogeology client has produced 168 documents and holds 349 boreholes on file as of this month, and the owner can read every one of its meters without asking us. That client came back for a second build, which is what tends to happen when the monthly bill stays boring.

One caveat. If your volume is large, say tens of thousands of model calls a day, the per-line figures above stop applying and model usage becomes the whole budget. None of the six systems we run is at that scale, so we will not pretend to have numbers for it.

Why we pass these through at cost

Every subscription and API account in our builds is opened in the client's name, billed to the client's card, at the provider's list price. Plenty of vendors do the opposite, so the reasoning deserves spelling out.

A vendor who marks up infrastructure earns more when your system consumes more. Nobody on that side is then motivated to trim a prompt that sends far more context than it needs; the waste is a profit line. Bundling also means the accounts sit in the vendor's name, which turns every renewal into a hostage negotiation. Leave, and your phone number, your run history and your data go dark.

Pass-through removes both problems. The only things we sell are the build and the care plan, so our interest is in keeping your meters low, because a client with a small monthly bill renews more happily than one with a large one. And since you own every account, firing us is always possible, which is the pressure a vendor should work under. We think a markup on infrastructure is the clearest warning sign in a whole proposal.

What makes the bill spike

In a well-built system the running costs barely move from month to month. When they jump, the cause is nearly always one of these.

1. Chatty prompts. A workflow that pastes the whole conversation history, an entire document or an unfiltered database dump into every model call pays for tokens the model never uses. It is the first thing we look at in a rescue.
2. Retry loops. An automation hits an error, retries, fails, retries again, and a loop that should have stopped after a few attempts runs all night. A sane retry policy is the first thing we set when an n8n workflow keeps failing in production. One bad retry policy is the fastest way we know to turn a small month into a large one.
3. The wrong model tier. Providers price their tiers ten times apart or more. Sorting an email into one of five folders does not need the flagship model, but systems built in a hurry default everything to the tier that worked first in testing.
4. Stale data. A support chatbot we took over was quoting a 2024 price list, which is the usual reason a chatbot starts making things up: an index nobody rebuilt. That is a running cost too, one that arrives as refunds and apologies and never appears on the API invoice.

How to cap them

Each of those has a cheap, boring control.

• A hard spending limit on every API account, set at about twice the expected monthly spend. Model and telephony providers all support caps and alert thresholds, and a capped system fails loudly, which beats billing quietly.
• Retry budgets. A few attempts, backoff between them, then stop and alert a human in a channel someone reads. The lead-routing n8n flow in one of our rescues ran about 200 times a day and failed about once a day, and nobody knew, because its alerts posted to a Slack channel that had been archived. Ten minutes of configuration would have caught it.
• Model tiering as a design step, with a test case per step proving the cheaper model passes.
• A cost review once a month. Five minutes, four dashboards. Trend up with no volume increase means something is wrong.

Ask any vendor two questions: what this costs per month at your volume, and whose name is on the accounts the costs run through. A vague answer to either is the warning sign. Spending caps and the monthly review are part of every one of our care plans, which start at $1,000 a month, and that is where the meters get watched.