What an AI automation actually costs per run

The model is rarely the expensive part. Platform operations, the APIs you call, the retries and the human who checks the output usually add up to more. A worked breakdown of one real invoice-processing flow.

Every business case we are shown for an AI automation has one number in it: the price per thousand tokens. It is almost never the number that decides whether the automation pays for itself.

Here is a real flow, anonymised, with every line in it. Supplier invoices arrive by email; the automation extracts the fields, matches them against a purchase order, posts the ones that match and queues the rest for a human. Around 4,000 invoices a month.

The full cost per invoice

Model calls: 0.004 €. Two calls — one extraction on a mid-size model with the PDF text, around 3,000 input and 300 output tokens, and one small classification. This is the number everyone quotes and it is under half a cent.

Document conversion: 0.006 €. The PDFs are scans, so they go through OCR before the model sees them. Managed OCR is priced per page and roughly matches the model cost. Teams forget this line because the model provider's pricing page does not have it.

Platform operations: 0.011 €. Twenty-two module executions per invoice on a paid Make plan. Not large individually; larger than the model in aggregate, which is the point.

The APIs being called: 0.000 € here, because the ERP is on-premise and licensed per seat. In a company on a per-API-call SaaS ERP this line can dwarf everything else, so check it before you design the fan-out.

Retries and reprocessing: 0.003 €. About 7 percent of runs fail once and succeed on retry, mostly OCR timeouts on large scans. You pay the failed attempt too.

Human review: 0.42 €. 18 percent of invoices land in the exception queue, and each takes a person about ninety seconds. At a fully-loaded 55 €/hour that is 1.38 € per reviewed invoice, or 0.25 € averaged across all of them. Add the 12 percent of auto-posted invoices that get spot-checked and you are at 0.42 €.

Total: about 0.45 € per invoice, of which the model is under one percent.

What that tells you to optimise

The exception rate is the whole economics. Going from 18 percent to 9 percent exceptions saves more per month than eliminating every model call, every OCR page and the entire platform subscription combined. That is why the work after go-live is not prompt tuning for its own sake — it is finding the three supplier formats that generate half the exceptions and handling them specifically.

The second lever is the platform operation count, not the platform choice. Twenty-two operations per invoice is a design, and in this flow eleven of them were an iterator that existed to reformat a list, removed in an afternoon. That is the operation counter as architecture.

The model matters when — and only when — you are running a chat-shaped agent with a tool loop, where a run is fifteen calls instead of two. Then it moves to the top of the list.

The number that justifies the project

Against 0.45 €, the manual process was a person entering invoices at about four minutes each: 3.67 € of labour per invoice, 176 hours a month. The automation costs roughly 1,800 € a month all in, including the reviewer's time, against 9,700 € of manual work, and the build was fifteen days.

What we did not put in the business case, and should have: three days a quarter of maintenance, because supplier formats change and an ERP upgrade broke the posting step once. Budget ten percent of the build per year and the case still holds comfortably.

How to measure yours

Write model, tokens, latency and outcome to a table on every run, and join it against the exception queue. Two weeks of that gives you a real cost per run and, more usefully, a cost per successful unattended run — which is the only number that compares honestly against the manual process. Most teams cannot produce it, which is why so many AI automation business cases are argued rather than calculated. Observability for automations is how you get there.

What to do this week

Take your most-run automation and fill in the six lines above for a single execution. If the human-review line is missing because nobody reviews the output, that is not a saving — that is the risk line, and it belongs in the same table.

ConsultorIA

Want this done on your cloud?

A ten-day read-only assessment is free, and Skyline lets you see your estate on a map before you write to us.

Related articles