MIXEDHow much does an AI project cost (and what makes it fail)
It's the question everyone asks and few answer with numbers. Here are the price ranges we see in SME projects, what pushes them up, what it costs after go-live, and the five reasons an AI project ends up in a drawer.
Price ranges (ballpark figures)
The figures below are ballpark estimates for companies of 10 to 100 people whose data is already digital. They're not a price list: every project has its own variables, listed in the next section.
| Project type | Initial investment | Time | Recurring / month |
|---|---|---|---|
| Assistant for incoming requests (emails, quotes) with draft replies | A few thousand € | 2-4 weeks | Tens of € |
| Data extraction from documents (invoices, delivery notes, orders) into the management system | From a few thousand to ~€15,000 | 3-6 weeks | Tens-hundreds of € |
| Assistant on company documents (procedures, contracts, history) | ~€8,000-20,000 | 4-8 weeks | Hundreds of € |
| Forecasting (demand, purchasing, cash) on historical data | ~€15,000-40,000 | 2-4 months | Hosting + maintenance |
| Computer vision (quality, safety, counting) | ~€20,000-50,000 | 2-4 months | Hardware + maintenance |
| Data platform + natural-language reports | Depends on the data foundation | 1-3 months | Tens-hundreds of € |
The pattern is clear: projects that work on text and documents are cheap because the models are ready-made and the work is integrating them; those that need clean historical data or training on your own cases cost more because most of the effort goes into the data, not the AI.
What pushes the price up
- Number of systems to connect. An assistant that only reads emails is one project; one that reads emails, searches the management system and writes to the CRM is three projects. Every integration takes time, and many business systems don't have decent APIs.
- Quality of the starting data. If the data is in misaligned Excel sheets, on paper or in free-text fields, it has to be put in order first. It's the most underestimated item: it can be half the budget. We cover it in Data strategy for SMEs.
- Required accuracy. A system that gets 1 in 10 wrong with a human checking costs half as much as one that must get only 1 in 100 wrong without checks. Deciding up front where you really need 99% saves a lot.
- Constraints on where data lives. Cloud models (cheap, fast) or models installed on-premise (more expensive, needed for certain sensitive data): the choice must be made at the start.
Costs after go-live
An AI project isn't a one-off purchase. Plan for: model usage on a pay-per-use basis (for an SME, from tens to a few hundred euros a month); hosting if there are installed components; maintenance, because models get updated, documents change and rules need recalibrating. A support fee of 15-20% a year of the initial investment is a prudent estimate.
The cheapest project is the one that starts from the right data. In the construction companies that replaced Excel sheets with V-Site, the first result wasn't "AI": it was the hours recovered every week by the admin office, site managers and workers, because attendance, deadlines and costs stopped being retyped by hand. The second, less visible result is that those companies now have reliable data — and a future forecasting or AI assistant project will cost them a fraction of what it would cost starting from spreadsheets.
The five causes of failure we see most often
- No specific problem. "We want to use AI" isn't a goal. "We want to reply to quote requests in 1 hour instead of 2 days" is.
- Data that doesn't exist or can't be used. Halfway through the project you discover the history is incomplete or the documents are illegible scans.
- No internal owner. The supplier delivers, nobody in the company adopts it. You need someone who uses it every day and flags what isn't working.
- No measurement. Without a "before", at the end nobody can say whether it worked — and the next budget never arrives.
- Expecting magic. AI makes mistakes. A project with no human checks in the first weeks gets abandoned at the first visible error.
How to protect the investment: a small pilot (a few thousand euros, one month), on a measured process, with an internal owner. If it works, the second project is funded by the hours saved by the first.
Frequently asked questions
How much does a first AI project cost for an SME?
A targeted pilot: a few thousand euros and a modest monthly cost. Projects with integrations, forecasting or vision: tens of thousands.
What drives the cost?
Systems to connect, quality of the starting data and the level of accuracy required.
What are the recurring costs?
Pay-per-use model usage, hosting and periodic maintenance.
Want a quote with numbers, not promises?
Describe the process: we'll reply with a cost range, timelines and what we'll need from you. Free 30-minute consultation.