The first question almost every client asks is always the same: "how much does an AI agent cost?". And almost always the honest answer throws them, because it is not a number, it is a wide range that depends on things which are not yet on the table in the first meeting. We have seen projects close at 4,000 euros and projects of the same "type" — on paper — end up north of 120,000. The difference is not the model, it is not the trend of the moment: it is how messy the process the agent has to touch really is.
In this article we try to take apart the cost of artificial intelligence for companies line by line, with the ranges we actually see in the field, no promises. The goal is not to sell you an agent: sometimes the most honest quote for an SME is "you don't need an agent, you need a 2,000-euro automation". We say it often, and we will say it here too.
The six factors that make the price
The cost of an AI project is not a single line. There are at least six, and each carries its own weight. If a vendor quotes you only one, something is missing.
- ◆Discovery and audit. Understanding the process, mapping the data, deciding whether AI is really needed. It is the line everyone wants to skip, and the one that, once skipped, blows up everything else.
- ◆Model choice and token cost. Which model, in the cloud or on-premise, and how much each interaction costs multiplied by real volume.
- ◆Development. The software around the model: prompts, logic, interface, error handling. It is the visible part, and rarely the most expensive.
- ◆Integration with existing systems. Getting the agent to talk to the ERP, the CRM, the PMS, that endpoint written back in 2019. Half the unexpected bill hides here.
- ◆Infrastructure and hosting. Where the thing runs, who keeps it up, how it holds under load.
- ◆Maintenance and monitoring. The cost nobody budgets for and which, at 12 months, often exceeds the build cost.
You quote the first five lines during the sale. The sixth you forget, and you remember it when the model bumps a version and the agent stops working one Tuesday morning.
Discovery: the line that looks useless and isn't
Discovery is the phase where we sit down with the client and try to answer an uncomfortable question: does this problem call for an AI agent, or for three lines of automation? It lasts from a few days to a couple of weeks, and for a small-to-medium project we quote it between 1,500 and 6,000 euros.
It looks like a cost you can cut. You can't. A manufacturing SME in northern Italy asked us for an agent to "reply to supplier emails". Two days of audit showed that 80% of those emails were three recurring request types, always the same: a parser and a template were enough. We delivered an automation of around 2,500 euros instead of a 30,000-euro agent. The client came back a year later, for a real project, because they trusted us.
Skipping discovery means building the wrong thing well. It is the most expensive mistake of all, because you find out at the end of the project.
Model and tokens: the cost that scales with success
Here lies the most common misunderstanding about the price of AI automation: people think the model is the big line. It almost never is, at the start. A call to a good commercial model costs fractions of a cent for light operations and a few cents for heavy ones with lots of context.
The problem is not the unit cost, it is the multiplication. A customer service agent handling 200 conversations a day, each with five or six steps and some context pulled from documents, can comfortably sit between 150 and 600 euros a month in tokens alone. The same agent handling 5,000 a day is a different budget line.
Three things move this number a lot:
- ◆Context length. Re-stuffing the entire knowledge base into every call is expensive. Retrieving only the three right chunks costs a tenth of that.
- ◆The chosen model. For many tasks a mid-size model does the same job as a large one at a fifth of the price. The choice should follow the task, not the name.
- ◆Cloud versus on-premise. Running an open model on your own infrastructure zeroes the per-token cost but shifts everything onto hardware and keeping it alive. It only pays off at high volume or under strict privacy constraints.
On CORA, our operation management system for hospitality, we spent weeks reducing the context sent on every call: same answer quality, token bill more than halved. That kind of work never shows up in the demo, but it is the difference between a healthy margin and a project that loses money at steady state.
Integration: where half the bill lives
Building the agent — prompts, logic, interface — is the part the client pictures when thinking about cost. It is also the most predictable. For a small project we are talking a few weeks of work, for a medium one a month or two.
Integration with existing systems is another planet. Getting the agent to read the right data from the client's ERP is easy if that ERP has clean, documented APIs. Almost nobody has them. On Smartlet CRM we have the AI assistant built in precisely because we control the data: we know where it is, in what format, with what permissions. When instead the agent has to hook into a closed PMS or a legacy system — think of the integration between CORA and the Passepartout PMS — the job is not writing prompts, it is real integration engineering, with its own timelines and surprises.
The rule we use: if the system to integrate has modern APIs, integration is 20-30% of the project. If it doesn't, it can become 50% and beyond. It is the line that wrecks quotes made without looking at the client's systems.
What it really costs: ranges by size
Let us give some numbers, in orders of magnitude, valid for the Italian SME and mid-market segment. They are project budgets, not price lists, and they assume discovery is included.
- ◆Small project (5,000 - 20,000 euros). An agent on a narrow scope: an assistant answering questions about company documents, a ticket classifier, an automation with a pinch of AI. Few, accessible systems. A few weeks of work.
- ◆Medium project (20,000 - 70,000 euros). An agent touching two or three systems, with business logic, controls and a real interface. Serious customer service, internal sales support, document generation with verification. One to three months.
- ◆Enterprise project (70,000 euros and up). Multiple orchestrated agents, integrations with legacy systems, security and compliance requirements, high volumes. Here the number depends too much on context to give just one, and annual maintenance becomes a structural line.
On top of these come the recurring costs: tokens, hosting and maintenance. A rough rule we use for the annual running budget is 15-25% of the build cost, every year. On a complex enterprise it can climb higher.
Build vs buy: the question that saves you the most
Before building anything, the right question is: does a product already exist that does 80% of this? Often yes. For many standard cases — a chatbot on the website, a writing assistant, a meeting summariser — buying an existing tool at a few tens of euros a month beats any custom build, and beats it badly.
Custom makes sense when the process is your competitive advantage, when the data cannot leave your perimeter, or when integration with your systems is the real value. CORA, Smartlet and SMACE are products because behind each one sits a specific domain — hospitality, a CRM with contract signing via BoldSign, corporate offsites with matching — where a generic tool falls short. But if your need is generic, do your budget a favour and buy.
The middle path, which we often recommend to SMEs, is to buy the model and build only the thin layer of integration around it: little code, little risk, contained cost.
Mistakes that inflate the bill
These are the most frequent ways an AI budget doubles against the initial quote. We have seen them all, and some we made ourselves.
- ◆Skipping discovery. Building before understanding. You pay twice: the wrong thing, then the right one.
- ◆Underestimated dirty data. The agent is only as good as the data it reads. If the customer database is a mess, half the project becomes data cleanup nobody had budgeted.
- ◆Scope creep. You start with "answer the FAQs" and end with "also handle orders, returns and complaints". Every expansion is another project disguised as an innocent request.
- ◆Zero maintenance in the quote. Models bump versions, system APIs change, prompts degrade. Without a maintenance budget, the agent that worked in June is broken by December.
- ◆No monitoring. Without structured logs of what the agent does, the first production error is impossible to debug and turns into a multi-day ghost hunt.
How to estimate ROI without fooling yourself
An AI project is justified if it frees up time worth more than it costs to keep it running. It sounds obvious, but the honest calculation is rare. Our minimum formula: how many person-hours a week the process eats, what you pay for them, what share you realistically automate. Never 100%. A good agent removes 40-70% of the repetitive work and leaves the hard cases to humans.
Then subtract the annual running costs — tokens, hosting, maintenance — and see how many months it takes to break even. If you break even in six to twelve months, it is a good project. If it takes three years, you are probably solving the wrong problem, or solving it too big.
And here comes the advice we give everyone: start with a narrow scope. The first act in the ring, not the whole show. An agent that does one thing well gives you real data on costs and benefits in a few weeks, with an investment you can afford to throw away if you realise you got it wrong. From there you expand on what works, with real numbers in hand instead of promises. It is slower to pitch and much harder to get wrong.
The answer to the opening question, then, is this: an AI agent in a company costs as much as the process it touches is messy, plus however much you forget to budget to keep it alive. Start small, measure, and make whoever sells you the solution tell you when you don't even need it.