Blog · Published · Updated
What a running AI agent actually costs
Short answer
A running AI agent has a build cost from €3,000 net and a monthly bill for its work. The model is often the smallest item: at 1,500 cases with 3 calls each on Claude Sonnet 5.5 it costs about €48 a month. People's time for checking and exceptions, servers, action logs and maintenance cost more.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- net per agent for one process, build paid once
- from €3,000
- from signing the contract to acceptance
- 6–16 weeks
- free prototype on your cases
- €0
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
AI running costs: statistics 2026
€48
a month for the model at 1,500 cases with 3 calls each on Claude Sonnet 5.5. People checking the output cost €785. Scenario.
50%
According to the European Commission (2026), half of EU SMEs and mid-caps name high investment costs as a barrier to AI.
Source: European Commission, Flash Eurobarometer on AI (2026) ↗
40%
According to Gartner (2025), over 40% of agentic AI projects will be cancelled by the end of 2027 because of costs, unclear value or weak risk controls.
14–56%
According to Anthropic (2025), AI cut task time by 14 to 56% in studies with a control group.
Source: Anthropic, Estimating AI productivity gains (2025) ↗
20.0%
According to Eurostat (2025), 20.0% of EU companies with at least 10 employees used AI.
Figures checked: 22 October 2026
Three amounts to keep apart
| Type of spending | What it covers | When it arises |
|---|---|---|
| Build | Process description, application, integrations, tests, launch | Once, from €3,000 |
| The system's work | Models, tools, servers, database, storage | Every month, by usage |
| Operating responsibility | Checking, corrections, data, security, maintenance | Ongoing, by process |
The model fee is not the price of the system. If the agent sorts email, pulls data from the CRM, prepares a draft and hands exceptions to an employee, each of these steps changes the bill.
A full month of one agent
- Today
- 1,500 cases × 6 min = 150 h (€5,235)
- With the system a person checks the output in 15% of the time
- €785
- Model, 3 calls per case, Claude Sonnet 5.5
- €48
- Cost of one case
- €3.49 today, €0.56 with the system
- Saving
- €4,402 a month
Payback of €3,000
0.7 months
The volume, the time and the checking share are the scenario's assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Model: 3,000 input tokens and 600 output tokens per call at Anthropic's rates, $54 a month, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
In this scenario the model is a small part of the cost with the system. The rest is people's time spent checking the output. That is why a better agent saves more money than a cheaper model.
The model and API calls
Vendors set separate rates for input and output, and charge separately for reading from and writing to the prompt cache. One case usually needs several calls. The Claude price list and the OpenAI price list give rates per million tokens.
3,000 input tokens and 600 output tokens per call, no prompt cache. Anthropic's rates in dollars, at the ECB rate of 1.1205 USD per euro.
Źródło: Anthropic, API pricing (2026) ↗
An employee's chat subscription is a separate product. An application that calls the model in the background uses API billing.
The application, servers and logs
An agent needs code that takes in a case, keeps its state and checks the output before writing it into the company's system. Reading a procedure and writing a draft carry less risk than changing customer data, sending a message or making a payment, which need extra safeguards. A small process runs on a single service. A larger one needs a database, a case queue, backups and a separate test environment. With n8n, queue mode needs separate processes and Redis.
After launch you must be able to replay a case: the input, the model version, the tool chosen, the write and the point where a person decided.
People's time
- Cases finished automatically
- Cases handed to a person
- Cases rejected for missing data
- Corrections before the output is sent
- Errors that need a rerun
If these items are missing from the spreadsheet, the cost looks smaller only because someone does the work without a line in the budget.
How to lower the monthly bill
Better quality gives the most: every case a person does not have to correct saves more than switching to a cheaper model. The second step is splitting work between models, with simple steps on a cheaper model and hard cases on a stronger one. The third is the prompt cache for repeated parts of instructions and documents. On Claude Sonnet 5.5 a cache read costs $0.10 per million input tokens. A normal input token costs $2 per million, twenty times more.
Who runs the environment
Your own infrastructure
servers and on-call
The company is responsible for backups, updates, secrets and security. For companies with a technical team.
Care from the contractor
by scope
Monitoring, fixes and changes priced separately from the build. The maintenance page describes it.
The scope of care is described on the AI system maintenance page, and the build prices in the guide how much an AI agent costs.
Worked example · Scenario
How we count the payback: a running system
In the free process review we count a full month: the model, the servers and people's time.
€5,817
a month for answering 2,000 customer emails
Payback: 0.6 mo.
5 minutes to read an email and write a reply and checking in 15% of the old time are our assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost: Claude Sonnet 5.5, about 3,000 + 600 tokens per message with a draft reply, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
See the full calculation →
