By Laas Nury"What does AI cost?" is the question behind almost every AI conversation, and the answers online range from "a few euros per user" to "a seven-figure programme". Both are true for something. This article breaks the budget into its parts, so you can see what you would actually be paying for, what usually gets forgotten, and how to check whether it pays off.
The short answer
There is no single price, because "AI implementation" covers very different things:
- Assistants for your people (Copilot, ChatGPT Business): roughly EUR 20 to 30 per user per month at 2026 list prices.
- A custom build for a process: a simple build can cost as little as around EUR 750; a set of solutions across several systems in a larger organisation can run to EUR 100,000 or more. Where yours lands depends on size, complexity and the factors below.
- On top of a build: running costs (model use, hosting, support), and the time of your own team.
Most companies start with one process and grow from there.
The budget, line by line
| Cost | What it covers | One-off or recurring | What drives it |
|---|---|---|---|
| Scoping and advice | Choosing the process, agreeing rules and exceptions, defining success | One-off | How clear the process already is |
| Build | Connections to your systems, the AI logic, exception handling, logging, guardrails | One-off | Number of systems, rules, exceptions, how much it may do on its own |
| Data work | Cleaning or connecting the data the process needs | Mostly one-off | The state of the data today |
| Your team's time | Describing the process, testing, approving, giving feedback | Mostly one-off | Availability of the process owner and the team |
| Training and roll-out | Explaining how it works and where it goes wrong | One-off, plus new starters | Number of users |
| Model use | The AI model processing each case | Recurring | Volume |
| Hosting | Where it runs and where its logs are stored | Recurring | EU-only, own environment or standard cloud |
| Support and changes | Keeping it running, fixing what breaks, adapting to process changes | Recurring | How often the process changes |
The build is usually the line people focus on. In practice, the lines that decide whether a project stays on budget are the data work and your own team's time.
What moves a build from simple to large
- Systems. One system is simple. Every extra system adds a connection, a data format and failure points.
- Access. A system with an API is quick to connect. One without needs a workaround.
- Rules and exceptions. Five rules are quick. Customer-specific agreements across hundreds of customers are not.
- Autonomy. Suggesting is simpler than acting. Acting needs limits, approvals, a full log and a way back.
- Volume and speed. A few dozen cases a day is different from thousands, or answers within seconds.
- Location. Standard cloud is simplest. EU-only, your own environment or strict security add work.
The costs that get forgotten
- Your own time. Someone who knows the process has to describe it, answer questions and test. Plan a few hours a week from them during the build.
- Data clean-up. If the customer master data is out of date, the AI will be too.
- The exceptions. The last 20% of cases usually takes more thought than the first 80%.
- Change. People need to know how it works, where it goes wrong and when to overrule it.
- Maintenance. Processes change, systems get updated. Budget for it every month, or the quality will slowly drop.
Ask any supplier to name these up front. A quote that only shows the build is half a quote.
Checking whether it pays off
A simple calculation, with illustrative numbers:
- What does the process cost today? Say 400 orders a week at 4 minutes each: about 27 hours a week.
- What remains after AI? Checking and exceptions, say 1 minute per order: about 7 hours a week. Saving: about 20 hours a week.
- What is that worth? At EUR 45 per hour including overheads: about EUR 900 a week, or roughly EUR 40,000 over a working year.
- What does it cost? The build, plus one year of running costs.
- What else changes? Fewer errors, faster handling, fewer late deliveries. Add these where you can put a number on them; mention them where you cannot.
If the saving in year one is larger than the build plus a year of running costs, the case is usually clear. If it is not, look for a process with more volume, or start with something simpler.
Buy or build?
| Buy a tool or licence | Have something built | |
|---|---|---|
| Best for | Personal productivity; standard processes the tool was made for | Processes that run through your own systems with your own rules |
| Cost profile | Per user per month | A build, then running costs by volume |
| Where it stops | What the tool supports; the rest stays manual | Work too varied to define |
| Watch out for | Paying for licences nobody uses; manual work around the tool's limits | Scope that grows without a check on value |
Many companies do both: assistants for everyone, a build for the process where the volume and the money are. Copilot, ChatGPT or a custom AI build helps you decide which is which.
Where to start
Pick the process with the most volume and the clearest rules, check it with the AI readiness assessment, and ask for a quote that names every line in the table above. If you want a straight estimate for your case, tell us the process. The first conversation is free.
