Vendors selling AI agents to small and mid-sized businesses report one-time costs of 1,500 to 25,000 euros plus 200 to 600 euros a month for maintenance. Those are vendor claims, not independent numbers. The real math is bigger: add data preparation, ongoing ownership, and a realistic failure rate before you compare anything to an FTE.
If you only look at the price on the quote, you’re comparing apples to a fruit basket. Here’s the full equation, the same one Renforza runs in an intake, with a named source next to every number, because in this market a number without an author is worthless.
What do vendors themselves report?
The ranges quoted by companies that sell agents, which means you should read them as sales figures:
- Build cost: 1,500 to 25,000 euros one-time for an SMB-grade agent, depending on complexity and how many systems it connects to.
- Maintenance: 200 to 600 euros a month.
- Comparison to staff: vendor SlimIQ claims an administrative FTE costs 35,000 to 45,000 euros a year versus 1,800 to 6,000 euros a year for an agent. The same vendor claims an AI-handled customer service conversation costs 0.10 to 0.50 euros versus 5 to 15 euros for a human one.
- Time savings: for invoice and document processing, claims of 70 to 90 percent less processing time circulate, among others from ctlabs, again in a vendor context.
There are bigger names with more concrete pilots too. Microsoft reports that Sandvik Coromant saved over 120 hours and $19,000 in the first three weeks of using its Sales Qualification Agent. Impressive, but note who’s reporting it: the company selling the agent.
That 1,500-to-25,000 range is itself a warning sign, by the way. A sixteenfold spread means “an agent” isn’t a product. It’s a project category, like “a renovation.”
What’s missing from the quote?
Three line items almost never appear, and they’re exactly where projects run aground.
Data preparation. An agent is only as good as the data and systems it runs on. If your customer records live in three systems that contradict each other, you pay for cleanup first. A field report from Project NANDA at MIT names data quality as a recurring implementation barrier. This preparation can cost more than the agent itself.
Workflow integration. An agent that sits next to your process instead of inside it doesn’t get used. MIT names that as a core failure cause too, along with tools that don’t learn from feedback. Integration is human work: deciding who hands off what, who reviews, and what happens when the agent isn’t sure. Budget for internal hours that never show up on an invoice but absolutely exist.
Maintenance as a permanent line. That 200 to 600 euros a month sounds like a subscription, but it comes with ownership attached. Someone has to spot-check output, adjust prompts and rules when your process changes, and step in when a provider updates its model. Without that owner, an agent quietly degrades.
How do you price in the failure rate?
This is the line no quote mentions and the one that hits the business case hardest. The NANDA report said 95 percent of the organisations studied saw no rapid GenAI P&L impact. That is not a universal failure probability, but it is reason to include a serious chance of zero return in your expected value.
Practically, that means two things. One: start with a project small enough that it is allowed to fail. Two: agree up front what success looks like, in euros or hours, not in enthusiasm. External partnerships outperformed internal builds in the implementations studied. That is not causal proof, but it is reason to treat relevant experience as a potentially valuable cost.
When is an FTE still the better buy?
The comparison to an FTE only holds if the agent truly does the same work. It rarely does, completely. An admin doesn’t just process invoices. She notices a supplier suddenly double-billing, picks up the phone, and trains the new hire. Replace the role with an agent and that residual work has to land somewhere, and the landing has a price too.
The reverse is also true: in today’s labor market, the real comparison is often not agent versus FTE but agent versus an open role that’s been sitting unfilled for eight months. According to CBS, the Dutch national statistics office, automation or AI is the primary measure in 2026 for nearly half of companies facing staffing shortages. An agent that catches sixty percent of the work beats an empty chair by a mile. How to make that call systematically is covered in our piece on whether a human is really the best candidate, and more broadly on our page for employers.
The whole thing in one equation
Take the vendor price, add data prep and internal integration hours, put a permanent maintenance line next to it, and multiply the expected benefit by an honest success rate. If the total still clears the human alternative comfortably, you’ve got a real business case instead of a brochure.
Renforza deliberately doesn’t publish a price list for agents, because that sixteenfold spread shows a price without an intake is a guess. After an intake, we quote a fixed price for a scoped agent, with the success metric attached. More on our agents page. Rather run the numbers yourself first? This article is the same math we’d do.
