By Laas Nury"What does an AI agent cost?" gets answered online with everything from "a few hundred euros a month" to "a six-figure project". Both can be true, because an agent that sorts one inbox and an agent set that runs the order process of a wholesaler across three systems are different things with the same name.
This article gives you an honest indication of what a build costs, explains what moves the price, how long a build takes, and which questions to ask any builder, us included, before you sign.
What a build costs
There is no fixed price list, because no two processes are the same. As an indication: a simple build can cost as little as around EUR 750, and a large one can run to EUR 100,000 or more. Where yours lands depends on the size of the company, the complexity of the process and the other factors below.
- At the simple end sits a single agent for one clear task on one system: sorting a shared inbox, reading one type of document, sending reminders. Clear rules, one place where it reads and writes, a limited number of exceptions.
- At the complex end sits a set of agents that together run a process across several systems in a larger organisation: reading orders, checking them against contracts and stock, preparing them in the ERP, chasing missing information, with approvals, limits and reporting for several teams.
Most companies start at the simple end with one process, and add from there once the first agent has proven itself.
These figures are for the build only. Running the agent costs extra, and it is worth knowing how before you start.
What moves the price
Six things decide what a build costs:
- The number of systems it works in. Reading and writing in one system is simple. Every extra system means another connection, other data formats and more ways for things to go wrong.
- Whether those systems have a proper way in. A modern system with an API is quick to connect. An old system without one needs a workaround, which takes longer and needs more maintenance.
- The number of rules and exceptions. "Every invoice goes to finance" is one rule. Customer-specific prices, approval limits per department and what to do with the fifteen ways a supplier can say "delayed" are many.
- How much it may do on its own. An agent that only suggests is simpler than one that acts within limits, because acting requires limits, approvals, a full log and a way to undo.
- The volume and the speed required. A few dozen cases a day is different from thousands, or from answers needed within seconds.
- Where it has to run. Standard cloud hosting is the simplest. EU-only hosting, your own environment or strict security requirements add work.
The model itself is rarely the expensive part. The work is in the connections, the rules and the guardrails.
The running costs
On top of the build, an agent has running costs. Three parts:
- Model usage. Every case the agent handles costs a small amount in AI model use. It grows with volume, so ask for an estimate per case and per month at your expected volume.
- Hosting. Where the agent runs, and the data storage for its logs. EU-only or on your own infrastructure changes this figure.
- Support and changes. Keeping it running, fixing what breaks, and small changes when your process changes. At NRY this is part of one monthly fee.
Ask for these as one monthly figure before you sign. An agent with a cheap build and an unknown monthly bill is not cheap.
How long it takes
| Phase | What happens | Typical duration |
|---|---|---|
| Scoping | Agree the process, the rules, the exceptions and what "right" means | Days, for one process |
| Build | Connect the systems, build the agent, build the guardrails | Days for a simple agent; weeks for a set of agents |
| Shadow run | The agent runs alongside your team without acting, and the results are compared | Two to four weeks |
| Go-live | The agent starts acting, first with approvals, then within limits | Stepwise |
The shadow run is not optional. It is the cheapest way to find out whether the agent is ready, and the only way to show your team that it is.
The phase that usually takes longest is not the build. It is agreeing the rules with the people who own the process. That time is not lost: most processes improve the moment someone writes the exceptions down.
What you bring to the table
A builder can move fast only if the company can answer a few things quickly:
- One owner for the process who can decide on rules and exceptions.
- Access to the systems involved, with a test environment if you have one.
- Examples: a few hundred past cases where you know the right outcome, to test against.
- Clear limits: what the agent may do on its own, what needs approval, and what it may never touch.
Seven questions to ask any builder before you sign
- What exactly is included in the build price, and what is not?
- What will it cost per month to run at our volume, split into model use, hosting and support?
- Where does it run, and where does our data go? Is EU-only possible?
- Who owns the solution and the rules, and can we change or stop it without you?
- How is it tested before it acts, and what accuracy counts as good enough?
- What does it log, and can we see every decision it made?
- What happens when our process changes, or when a connected system is updated?
A builder who answers these clearly and in writing has usually done it before. A builder who talks mostly about the model has usually not.
Where to start
Pick the process where your team spends the most time copying, checking or chasing, and where a mistake is visible and reversible. That is usually where a simple first agent pays for itself fastest. If you are not sure whether your process needs an agent at all, RPA, AI or AI agents walks through the choice, and AI agents for business explains the guardrails in more detail.
And if you want a straight estimate for your own case, tell us the process. The first conversation is free.
