What an enterprise AI assistant costs in 2026

It's one of the first questions any company asks me when we talk about internal AI: «how much would something like this cost me?». And the short answer is uncomfortable: it depends less on the technology than on how you buy it. The same outcome — an assistant that answers with your company's knowledge — can cost you a monthly subscription or the price of a car, depending on the route you choose.

Let's put some numbers on the table and, above all, understand what sits behind them.

What you're actually buying

«Connect ChatGPT to my documents» sounds like a Friday afternoon, but a serious system has far more pieces than a demo shows: ingesting your sources (Drive, email, chats, PDFs), chunking and indexing the documents, the semantic search, permission control per person and department, the chat interface, the integrations with your tools and — the part almost nobody budgets for — continuously evaluating that the answers are correct. Every one of those pieces is engineering work. And every one needs maintenance when the models and the tools change, which is constantly.

The bespoke route: what it costs in 2026

If you hire a consultancy to build it from scratch, general market ranges in 2026 look like this: a basic project (one data source, a narrow use case, no serious permission requirements) usually starts in the several-thousand-euro range. A complete system, with multiple sources, integrations, serious access control and production deployment, easily reaches several tens of thousands. These are indicative ranges, not any particular consultancy's quote: every project is its own world. But that's the order of magnitude.

To make «several thousand» and «several tens of thousands» mean something, these are the brackets published in 2026 market guides. They are not my prices and not anyone's in particular: they're the range the circulating quotes move in, and they help you tell whether what lands in your inbox is inside or outside the normal.

Size and scope	Rollout	Maintenance
10-50 employees, one narrow use case	€3,000 – 25,000	€600 – 2,500/month
SME with several sources and integrations	up to €35,000
Large bespoke project, multi-department	€15,000 – 150,000

Ranges collected from market guides published in 2026. They are not an audited study: they're what the people selling these projects say, so read them as an order of magnitude and not as a price reference.

There's rough consensus on how those amounts break down internally too: between 10 and 20% goes on consulting and analysis, between 40 and 55% on development, and between 10 and 15% on infrastructure. The rest is management and contingency. And on return, what gets repeated is a band of 6 to 18 months, which is worth reading sceptically: it depends entirely on what the problem costs you today, which is exactly what the calculator tries to put in euros for you.

And the part that rarely comes up in the first meeting is missing: ongoing monthly maintenance, paid separately. AI models update every few months, connectors break when tools change their APIs, answer quality has to be watched. An abandoned AI system degrades; it isn't a cabinet you assemble once and forget.

And building it in-house?

There's a third route: having your own technical team do it, if you have one. The visible cost is those people's salaries for however many months it takes; the real cost includes what they stop doing meanwhile and the learning curve in a field where mistakes are expensive, especially in security and permissions. In the end it's a bespoke project by another name: you take on the full development and the maintenance forever.

A closed project versus a product that improves on its own

Here is the structural difference between routes, more important than any figure. A consultancy project is a photograph: it's delivered, invoiced and finished. The system they handed you in March is exactly the same one in November, unless you pay for another round of development. Every improvement, every new model, every security refinement is an additional quote that comes entirely out of your pocket.

A subscription product works the other way round: it's a film. It improves every week for all its clients at once, because the vendor lives on the product continuing to be good, not on billing you hours. When a better model comes out, you have it. When security is strengthened, it reaches you. No new projects and no extra budgets.

Why the subscription comes out cheaper (it isn't magic, it's sharing)

The reason is simple arithmetic. Building the hard parts properly (real multi-tenancy, per-department isolation, traceability for every answer, quality evaluation) costs the same whether it's used at one company or a hundred. In the consultancy model, each company pays for that development in full from scratch, as though it were the first in the world to need it. In the product model, that cost is borne once and shared between all the clients. You pay a fraction of the development, not the total.

When bespoke does make sense

Honestly: there are cases where the bespoke project is the right option. Very specific regulatory requirements, integrations with internal systems nobody else uses, or simply the size and the technical team to maintain your own build for years. If that's your case and the full sum works out, a good consultancy will do a good job.

But if you're a company of between 20 and 200 people, doing the whole sum (development, plus maintenance, plus every future improvement) usually tips the balance very clearly towards the product.

Savia exists precisely because of that sum. I wanted a mid-sized company to have the same level of technical sophistication as an expensive bespoke project (real multi-tenancy, isolated spaces per department, every answer with its source cited) while paying for a bounded rollout and a maintenance fee, not an entire development that comes wholly out of your pocket. In the end, the right question isn't «how much does an AI assistant cost?». It's «what share of that cost am I paying on my own?».