It’s the first question everyone asks and the one almost nobody answers with a number: how much does it cost to implement AI in a company. You ask for three quotes and get three beautiful PDFs that end in "let’s book a call to talk about your case". Nobody lies, but nobody gives you the figure. We’re going to do the opposite: first the ranges that do exist, with a source; then where the money goes; and finally why the industry keeps the price in its pocket.
How much does it cost to implement AI in a company: the numbers that do exist
Even if nobody says it to your face, the market ranges are published. An AI project for an SMB in the Spanish/European market costs, according to 2026 market analysis, between €15,000 and €60,000 to roll out, plus €500 to €2,000 a month in maintenance and licenses (Hiberus, 2026). But that wide range hides the fact that there isn’t "an" AI project: there are several project sizes, and each has its own price.
| Type of project | Ballpark range | When it makes sense |
|---|---|---|
| Off-the-shelf SaaS tool (subscription) | €50–500/mo | A standard pain a market product already solves |
| Low-code automation on top of what you have | €2,000–8,000 | Connect apps and remove one concrete copy-paste |
| Custom development on your operation | €8,000–30,000 | A process of your own no product covers |
| Full SMB project (several processes) | €15,000–60,000 + €500–2,000/mo | Redesign a real part of the operation, not a pilot |
The SaaS, low-code and custom-build tiers come from the same kind of market analysis (Utilia, 2026). They’re not laws: they’re orders of magnitude. But they already tell you something no vendor will on the first call: if a €200/mo subscription solves your problem, paying €30,000 for "a custom AI project" is throwing money away — and the other way around.
Where the money goes (the breakdown nobody shows you)
An honest quote can be opened into line items, and when you open it you see why the range is so wide. The typical split of a rollout project is roughly this:
- Consulting and diagnosis: 10–20%. Understanding the process and deciding what to automate. This is where many set up camp (more on that below).
- Development and integration: 40–55%. The expensive part and the one that actually does the work: wiring your systems, writing the logic, leaving it running. If this line is small, be suspicious: they’re selling you diagnosis, not implementation.
- Cloud infrastructure: 10–15%. Where the system runs.
- Model and platform licenses: 5–10%. The usage cost of the AI itself, which is almost never the big line.
- Team training: 5–10%. What separates a living tool from a dead license.
- Ongoing maintenance: 15–25% a year of the initial cost. AI in production isn’t furniture: it moves, degrades and gets tuned.
That breakdown also comes from the cited market analysis. And it has an uncomfortable reading: over half of the real budget is building and integrating — the part you can check and that almost nobody wants to sign for at a fixed price. When you decide what to automate and what not, that’s where the 40–55% is decided: we develop it in the guide to automating processes with agents, because picking the wrong process multiplies the cost without multiplying the result.
Why the industry hides the price behind "request a quote"
There’s a legitimate reason and a comfortable reason, and it’s worth telling them apart. The legitimate one: the price really does depend on your case —how many processes, which integrations, what state your data is in— and giving a figure without looking would be lying in the other direction. That’s true. The problem is that this truth gets used as an alibi for the comfortable reason.
The comfortable reason is that "request a quote" does two things that don’t benefit you: it anchors the price upward once you’ve already gotten excited, and it sells you first of all a three-month discovery phase billed separately that often just confirms what you already knew. The industry has turned the non-price into a product. The blurrier the entry price, the more room for the project to grow without you being able to compare.
What actually moves the price
If you want to estimate your own case before asking for a quote, these are the four dials that really move the figure, in order of impact:
- How many processes you touch. One well-chosen process is a bounded project; five at once is another animal. Start with the one that eats the most hours.
- How many integrations it needs. Connecting one system is cheap; connecting your old ERP, your CRM and three sheets nobody documented is where the budget goes.
- What state your data is in. If it’s tidy, you start; if it lives in unstructured PDFs and emails, there’s prep work somebody has to pay for.
- How much autonomy you want. An assistant that proposes and a person approves is cheaper and safer than a system that decides alone. Autonomy isn’t a switch, it’s a dial you turn up when accuracy earns it.
And a common-sense shortcut before signing anything: if the pain is that your team spends the day moving data between the same apps, that rarely needs a €60,000 project. Often it’s operations automation on top of the tools you already use —setting up the internal use of AI well is half the work, and we explain it in using ChatGPT in your company—.
How to ask for a price and actually get one
The way to break the "request a quote" game is to ask for the price differently. Three sentences that change the conversation: "give me a range before the diagnosis"; "I want a fixed price per phase, not per hour"; "what happens if you go over the estimated time?". The vendor who implements answers all three without getting nervous, because they live off their work working. The one who does theatre dodges them, because they live off you not looking too closely.
And a note that changes the final number in Spain: a good share of these rollouts are fundable with public grants like Kit Digital and Kit Consulting, which cover exactly diagnosis and AI rollout in SMBs. It won’t turn a bad project into a good one, but it does change how much comes out of your own pocket.
The ROI, without the slide deck
Cost only makes sense against what it returns. Market analysis puts the typical payback of a well-executed project at 4 to 12 months (Utilia, 2026). But watch out for slide-deck ROI: a promised return isn’t a measured return. The difference between the two —and how to tell the figure that holds up from the one that just decorates a proposal— we develop in real ROI vs. PowerPoint ROI. If your quote doesn’t say what will be measured each week and from what baseline, you don’t have a ROI: you have an illusion with decimals.