What an AI agent actually costs in an SMB
€40 a month in model spend, €6,200 to stand it up. Both numbers are true, and the gap between them explains almost every budget surprise in AI projects.
The triage agent we built for an eighteen-person SMB burns about €40 a month in model spend. Standing it up cost €6,200. Both numbers are true, and the gap between them explains almost every budget surprise in AI projects.
The people selling agents quote the first number. The people running them live with the second.
The three line items
- The model — cents per run. It’s the most visible line item and almost always the smallest. For an agent at normal SMB volume it rarely exceeds €30 to €80 a month.
- The wiring — the work of making it talk to what already exists: CRM, invoicing, calendar, the spreadsheet nobody wants to replace. This is where most of the upfront cost lives, and it’s the line nobody can estimate honestly without seeing the systems from the inside.
- The supervision — the person who still approves, corrects, and works out why the agent decided what it decided. It doesn’t disappear. It shrinks.
What makes the number jump
Systems with no API. If the invoicing software only exports CSV by email, the wiring cost triples — and so does the fragility.
Processes nobody wrote down. Automating a process that only exists in two people’s heads means discovering it first. That work is consulting, not engineering, and it gets billed all the same.
Real exceptions. An agent that handles 80% of cases is cheap. Taking it from 80% to 95% usually costs more than the first 80% did.
The calculation that matters
It isn’t "what does the agent cost". It’s what the work it will do costs you today, measured before you start. If manual triage eats twelve hours a week from someone at €18 an hour, that’s roughly €11,000 a year. A €6,200 project running at €40 a month pays for itself in seven months — and from there the comparison is honest.
Without that starting number, any quote looks expensive or cheap depending on the mood of whoever reads it.
When it doesn’t pay off
Processes that happen fewer than a few dozen times a week rarely justify the wiring. The fixed cost of integrating doesn’t drop because volume is low — it just takes far longer to amortise. In those cases the right answer is usually a better checklist, not an agent.
The practical rule: if you can’t say how many hours a week the process eats today, you aren’t ready to ask for a quote. Measure first.