How much does it cost to integrate AI into a company? Real prices and what drives them
TL;DR: Integrating AI into an SME has three cost items: the setup, the recurring cost and your own people's time. At VisibleSoft, a diagnosis costs €1,500 (1–2 weeks) and a production pilot starts at €5,500, usually between €5,500 and €9,000 and 4–6 weeks, plus €250–400/month of maintenance. An off-the-shelf product such as Ana, the AI receptionist, starts at €990 setup and €149/month. All prices exclude VAT. Model usage is usually the small item in text use cases; what really makes a project expensive is the integration with your systems.
Search for “how much does it cost to integrate AI into a company” and you’ll find plenty of articles explaining what the price depends on without giving you a number. I’m going to give you mine, because they’re published on my website, and explain where each one comes from.
If you run an SME and someone has suggested “adding AI”, by the end you’ll know which items make up the budget, which are big and which are small, and how to do a quick calculation yourself to see whether your case deserves a pilot.
The short answer, with numbers
These are VisibleSoft’s prices for integrating AI into the systems you already use, excluding VAT:
| Item | Price | Timeline |
|---|---|---|
| Diagnosis | €1,500 | 1–2 weeks |
| Production pilot | from €5,500 (usually €5,500–9,000) | usually 4–6 weeks |
| Maintenance | €250–400/month | recurring |
| AI receptionist (off-the-shelf product) | from €990 + €149/month, with 300 minutes | 2–3 weeks (Essential plan) |
The diagnosis maps the candidate processes, estimates the savings and closes the scope of the pilot. The pilot is an agent running in production, connected to your real system, with a demo every two weeks and basic documentation with the AI Act in mind. It isn’t a mock-up to show off in a meeting.
The gap between €5,500 and €9,000 isn’t arbitrary. It’s almost always explained by the same thing: what needs connecting.
The three cost items of any AI project
It doesn’t matter who you hire. An AI project has three costs, and you won’t always see all three in the quote.
The setup. Diagnosis and building the pilot. It’s the big item and the only one you pay once.
The recurring cost. There are two different things here that are worth keeping apart. One is maintenance: someone watches what the agent does, adjusts its instructions when your prices or services change and reviews the errors. The other is model usage, which the provider (OpenAI, Mistral, Anthropic…) charges per use. Always ask whether it’s included in the fee or billed separately.
Your time. Nobody invoices it, but it’s there. Someone in your company has to explain the process, test the agent with real cases and say “we wouldn’t do it like that”. If that person has no time, the pilot drags on, and not because of the technology.
Model usage is usually the small item
It’s what worries people most and, in text use cases, what weighs least.
Here’s a real case. In Cerrajería Carlos Rodríguez’s management software there’s a button that polishes the wording of quotes and invoices. The first version ran a small model on their own server: zero euros per use, but it permanently reserved 4 GB of RAM and 3.5 CPUs and each proposal took several seconds. In August I moved it to the Mistral API. In the worst case the application allows, a thousand calls a month cost less than one dollar.
In other words: the “free” model wasn’t free. The server was paying for it.
In an SME’s text use cases (sorting emails, extracting data from invoices, drafting replies), usage tends to range from a few cents to a few tens of euros a month. It goes up with volume, with very long documents and, above all, with voice: a real-time phone conversation consumes far more than a paragraph. That’s why keeping usage under control is part of maintenance from day one, not something you look at when a surprise bill arrives.
What really makes a project expensive
If the model is cheap, where does the money go? Into connecting the AI to your company. These are the things that push a pilot towards €9,000 or beyond:
- Systems without an API. If your ERP doesn’t offer a documented way to read and write data, you have to build one or work with the database or with exports. It can be done, but it takes more hours and more testing.
- Writing, not just reading. An agent that looks up orders is much simpler than one that creates them. Writing to your system requires validations, permissions and a plan for when it gets something wrong.
- Several integrations at once. WhatsApp plus CRM plus ERP isn’t three times one integration: you also pay for the connections between them.
- Scattered or dirty data. If the information is spread across three spreadsheets and someone’s head, it has to be sorted out first. No model fixes that.
- Strict data requirements. If the data can’t leave your infrastructure, the model has to run on your servers. It’s possible, but it’s a separate project with its own budget.
Notice that none of these has anything to do with “which AI we use”. Choosing the model is the cheapest decision in the whole project.
What makes it cheaper
Starting with a single process. The one that repeats most, has clear rules and takes hours away from a specific person today. Better to solve supplier invoice entry properly than to attempt “an assistant that does everything”.
Starting by reading. Let the AI propose and a person approve. In the locksmith’s case, in testing after the model change, a description with no price came back with an invented one: €186, copied from an example in the instructions. It was fixed in the instructions, but the real safeguard is something else: nothing is saved until a person has reviewed it. And that design, besides being safer, is cheaper to build.
And measuring before expanding. A pilot that proves savings with numbers justifies the next one. One that “works fine” justifies nothing.
Off-the-shelf product or custom development
There’s an alternative to all of the above: buying something ready-made. Ana, our AI receptionist, has a fixed price because the problem is always the same (answering calls, collecting the details and sending a summary by email) and what changes from one company to another is the configuration. If your case fits a product like that, it’s cheaper and faster.
If it doesn’t, which is the norm when the AI has to talk to your ERP, the price is indicative until the diagnosis. And if you don’t even need new software, I’ll tell you that too: I wrote about when you don’t need custom software.
Before asking for a quote, do this calculation
Take the process you have in mind and write down three numbers: how many times it happens a month, how many minutes it takes each time and how much an hour of the person doing it costs.
A made-up example to show how it works: 400 supplier invoices a month at 4 minutes each is almost 27 hours a month. If the AI takes over most of that work, you already have a figure to compare with the pilot and the monthly fee. If you get 20 invoices a month, the honest conclusion is that it isn’t worth it, and that’s fine.
If the numbers work and you want to check them, that’s what the diagnosis is for: in 1–2 weeks we measure the process properly and I tell you whether there’s a case. If there isn’t, I’ll tell you that too.
Resources
- Applied AI for companies: use cases, process and prices.
- How to add AI to your ERP without switching systems: the ways to connect AI to what you already have.
- AI agent or chatbot: which one you need: what changes in price and risk.
- Case study: AI in Cerrajería Carlos Rodríguez’s management software
Frequently asked questions
How much does it cost to integrate AI into an SME?
It depends on how many systems need connecting and on whether the AI only reads data or also writes it. As a reference, at VisibleSoft a production pilot starts at €5,500, usually between €5,500 and €9,000, preceded by a €1,500 diagnosis and followed by €250–400/month of maintenance. Prices exclude VAT.
Is there a monthly fee?
Yes, and be wary of anyone who doesn't mention it. An AI agent in production needs supervision: reviewing answers, adjusting instructions when the business changes and keeping model usage under control. At VisibleSoft that maintenance costs between €250 and €400 a month.
How much does using the AI model cost?
For text use cases, very little. In the assistant that polishes quotes inside Cerrajería Carlos Rodríguez's management software, a thousand calls a month to the Mistral API cost less than one dollar in the worst case. Voice and high volumes cost more, which is why usage is measured from day one.
What if the diagnosis concludes it isn't worth it?
Then the diagnosis has done its job and saved you the pilot. That's exactly what it's for: measuring the process, estimating the real savings and deciding with numbers whether there's a case or not.