How much does it cost to implement AI in a company? The August 2026 numbers
Every few months I review what the market guides publish about the cost of implementing AI in a company, because it's the question everyone arrives with and because the numbers move fast. This is the August 2026 picture: what a custom build costs, what companies in Spain actually pay, where the money really goes (spoiler: not the AI model) and how to read a quote without surprises in month six.
A warning before we start: none of these ranges is an audited study. They are the brackets published by the people who sell these projects, so read them as orders of magnitude, not as a price list. That's what they're useful for: knowing whether the quote on your table is inside or outside the normal range.
What a custom build costs in 2026
If you hire the from-scratch development of a system that answers with your company's knowledge (what's technically called RAG: the AI searches your documents and answers with what it finds), this year's international guides agree with each other to a surprising degree:
| Type of project | Published range |
|---|---|
| Simple system, one data source, no serious permission requirements | $15,000 – $50,000 |
| Production system: several sources, hybrid search, serious access control | $40,000 – $80,000 |
| Enterprise, on-premise or agentic project | $80,000 – $150,000+ |
| Monthly operation of an on-premise system | $3,000 – $10,000/month |
Collected from pricing guides published in 2026 (linked at the end). In euros the order of magnitude is the same.
One more figure worth knowing even if it isn't your case: for a complete enterprise rollout, from proof of concept to a hardened system with compliance and SSO, the first-year budgets in circulation run from $60,000 to over a million. Not because the AI is expensive, but because of everything around it. And that's the key to this whole article.
And what companies pay in Spain
The Spanish guides tell a similar story with numbers closer to the SME:
| Type of project | Published range |
|---|---|
| Implementing AI in a company (serious project, with integration) | €15,000 – €150,000 |
| Typical SME project (AI chatbot or process automation) | €3,000 – €10,000 for the initial phase |
| Template chatbot | from €1,200 |
| Custom chatbot with its own flows | around €6,000 |
| SaaS subscription with AI | €80 – €250/month; custom, €1,500 – €5,000/month |
Same warning: indicative ranges from the people who sell these projects, useful as a compass and not as a tariff.
Notice the jump: between the €1,200 chatbot and the €40,000 project there is no difference in the quality of the AI. The model doing the answering can be literally the same. The difference is in what surrounds it, and that takes us to the most important figure of all.
Where the money really goes (it's not the model)
This is the point where every serious guide agrees and the one almost nobody mentions in the first meeting: the biggest cost of an enterprise AI project is not the AI. It's the access control layer, the integrations and regulatory compliance. The data preparation work eats between 30 and 50 percent of the total budget.
The typical breakdown published by the Spanish guides confirms it: between 10 and 20 percent on consulting and diagnosis, between 40 and 55 on development and integration, 10-15 on infrastructure, 5-10 on model and platform licences, and another 5-10 on training. The call to the AI model, the part you see in demos, is a small fraction of the bill.
And it makes sense if you think about it from the builder's side: making the AI answer is easy; making it answer only with what each person has permission to see is the hard part. A semantic search engine finds the most relevant thing, not what each person should be allowed to read, and that's where the information leaks between departments I covered in another article come from. Permissions, isolation and traceability aren't an extra on the project: they are the project. When you compare quotes, ask exactly what each one includes under that heading, because that's where the real difference between a €15,000 quote and a €60,000 one hides.
The cost that doesn't show up in the quote
An AI system isn't a wardrobe you assemble once. Models update every few months, connectors break when tools change their APIs, and the quality of the answers has to be watched continuously. The guides talk about maintenance between €600 and €2,500 a month for custom projects, and $3,000 to $10,000 a month for on-premise deployments. If the quote you've been given says nothing about maintenance, it's not that it doesn't exist: it will arrive on a different invoice.
On returns, the bracket that keeps repeating is 6 to 18 months. Read it with scepticism, because it depends entirely on what the problem you want to solve costs you today. Nobody can give you that figure from outside: it comes from your interruptions, your searches and your staff turnover, and it's exactly what the knowledge-loss calculator tries to put in euros with your own numbers.
How to read an AI quote without getting played
With all of the above, reading a quote comes down to five questions:
1. Where are the permissions? If per-person and per-department access control doesn't appear as a line item, or appears as «phase 2», the price you've been given is not the price of the system you need.
2. What happens when the model changes? The models of 2026 won't be the models of 2027. Is the upgrade included or is it another project?
3. How much is monthly maintenance and what does it cover? Get it in writing and with scope. «Support» without a definition covers nothing.
4. Who checks that the answers are correct? Continuous quality evaluation is the item most often omitted and the one that decides whether the system gets used or abandoned after three months.
5. What happens to your data if you leave? Export and verifiable deletion, before signing. I keep a full checklist of questions for vendors if you want to go deeper.
The final sum: pay for the whole build, or a fraction of it
The ranges above are what it costs to build it for you alone. The alternative is an already-built product, where the development of the hard parts (isolation, permissions, traceability, evaluation) was paid for once and is spread across all its customers. I worked through that arithmetic in another article dedicated to the cost of an AI assistant, with the full comparison between a closed project and a product that improves on its own.
There are cases where going custom is the right call: very specific regulatory requirements, integrations nobody else uses, or the size and the technical team to maintain your own build for years. But if you're a company of 20 to 200 people, the complete sum (development, plus maintenance, plus every future improvement paid in full by you) usually tips the scales clearly. Savia exists because of that sum: the technical level of an expensive custom project, with a three-week implementation and a monthly fee, instead of an entire build coming out of your pocket.
Sources consulted (August 2026): SoluLab, ScalaCode, Ortem Tech, Ment, Hiberus, IA Consultora, Soul IA and Potenzzia.