How much does an AI agent cost a company in 2026: real prices, cost breakdown and how to compute ROI

"How much does an AI agent cost?" is the first question in almost every intro call, and the honest answer is not "it depends": it is a range, with the variables that move it spelled out. In 2026, for a small or mid-sized company, a well-built single automation starts at 1,200€, a complete vertical AI agent starts at 2,500€, and an enterprise solution with multiple integrations, granular permissions and SLAs typically lands between 30,000€ and 50,000€. This article explains what sits inside each of those numbers, which costs recur every month, and how to estimate the payback period. No magic price list: just the way I actually build quotes.

Start with the difference between the two extremes, because it is the main source of confusion. An automation is a deterministic flow: when X happens, run Y and Z. Extract data from PDF invoices arriving by email and write it into the ERP, sync e-commerce orders with the warehouse, generate a weekly report from three different sources. There is no decision, just a chain of steps a person used to do by hand. An AI agent receives ambiguous context and decides: it reads a request written in natural language, works out what the sender wants, picks a tool, executes, and knows when to stop and hand over to a human. The second costs more because it requires prompt engineering, guardrails, error handling and a pilot phase. If your problem is deterministic, do not pay for an agent: pay for an automation and save half the budget.

A 1,200–2,000€ quote for an automation typically contains: half a day analysing the current process with the people who run it today, building the workflow on self-hosted n8n, integrating two or three systems that already expose APIs (Gmail or Outlook, the CRM, the ERP or a Google Sheet), error handling with a Slack or Telegram alert when something breaks, a short operating manual and two weeks of post-launch support. That is usually three to five working days. What pushes the number up is always the same: an ERP without APIs that has to be queried through the database or scraped, dirty data that needs normalising, and processes nobody can describe precisely because "everyone does it their own way". That last one is not a technical cost, it is an analysis cost, and it is the one everybody underestimates.

A 2,500–8,000€ vertical AI agent adds the decision layer. Inside: defining the use case with measurable success criteria, the system prompt and its iterations, access to the company knowledge base through semantic search (documents, price lists, procedures, ticket history), the tool set the agent can use with its permissions, guardrails separating what it may do alone from what needs human approval, structured logging of every action for audit and debugging, and a supervised pilot of two to four weeks where the agent works but a human reviews. The pilot is not an optional extra: it is the part that turns a prototype that demos well into a system that survives real traffic from real customers and their badly worded requests.

The 30,000–50,000€ band is not the same thing, only bigger. It changes in nature: multiple agents collaborating, integration with legacy systems that have no modern APIs, roles and permissions for dozens of users, documented compliance requirements (EU AI Act, GDPR, traceability of automated decisions), separate staging and production environments, automated tests and a maintenance contract with guaranteed response times. The cost here is not the AI model, it is the engineering around it: when a 200-person company puts an agent on its order process, one unhandled failure can burn the entire project budget in a week.

Then there are the recurring costs, the part almost nobody quotes clearly

There are three. Model consumption: an agent handling one hundred to five hundred conversations a month on Claude or GPT realistically costs 20–150€ per month, rising when every request carries a lot of context. Infrastructure: a VPS for self-hosted n8n costs 5–20€ per month, a managed database is often free below certain thresholds and 25€ per month once you grow. Maintenance: APIs change, models get deprecated, business processes evolve. A light maintenance contract (monitoring, updates, small changes) runs 150–400€ per month; without one, budget a few days a year on a time-and-materials basis. All in, a well-sized agent for an SMB costs between 50€ and 500€ per month to keep alive.

ROI is simpler to compute than it looks, and it should be computed before signing. Count the person-hours the process absorbs in a month, multiply by the fully loaded hourly cost (not net salary: the employer cost, often 25–45€ per hour for office roles in Italy), and you get the current monthly cost. Then estimate the realistically automatable share: for a first project, 50–70% is prudent, not 100%. If a process absorbs 60 hours a month at 30€ per hour it costs 1,800€ monthly; automating 60% of it saves roughly 1,080€ a month, so a 4,000€ agent with 200€ of running costs pays back in a little over four months. If the maths does not work within twelve months, the project should not happen — and I will say so on the call, rather than let you find out at year end.

Two sources of value sit outside that calculation but weigh on the decision

First, response time: going from two hours to twenty minutes on first reply saves no person-hours, but it changes the conversion rate on inbound leads, and the return there can dwarf the operational saving. Second, error reduction: a wrong order, a mishandled return or a lost lead all carry a cost nobody usually tracks, yet the people running the process can estimate it surprisingly well. It is worth asking them before you start.

Three mistakes that inflate cost, and that I keep seeing

One: starting from the tool instead of the process. "We want an agent on WhatsApp" is not a requirement, it is a channel; the requirement is what must happen when a message arrives. Two: automating a broken process. If the manual flow is chaotic, the agent amplifies the chaos faster — fix the process first, then automate it. Three: skipping the pilot to go live immediately. The pilot costs two weeks and saves you the angry-customer phone call that is the real cost of the project.

How to get to a trustworthy number for your case: three pieces of information. Which specific process, at what monthly volume. Which systems must be touched, and whether they have APIs. What the agent may decide alone and what it may not. With those three answers a serious quote takes an hour to write and does not drift mid-project. The free fifteen-minute call exists exactly for this: to check whether the use case holds up, tell you which price band you are in, and tell you when the ROI is not there.