What an AI project really costs
Same brief, three quotes, a factor of three between them. Usually none of them is dishonest. They just do not contain the same things. Here is what moves the price, and where every cost sits with us.
- Cost of an AI project
- The cost of an AI project is the sum of data preparation, connection to existing systems, accuracy testing and handover into operation. Model usage is only a small part of it, and the number of features decides almost nothing.
Between the quotes lies a range, not a price.
Hand the same brief to three vendors and three amounts come back, two or three times apart. The cheap one is rarely a bargain, the expensive one rarely a rip-off.
The cheap quote is almost always missing something: the connection to your ERP, the handling of exception data, the handover manual, maintenance. You will need all of it later, and then it sits on an extra invoice.
That does not make the expensive quote right. The numbers only become comparable once all three list the same items side by side. How to get there is in section 06. What moves the amounts in the first place is right below.
Five things move the price. The number of features is not one of them.
What actually eats hours is rarely visible in the brief. It shows in the data and in the systems around it.
State of the data
Clean documents in one format, or mixed stock with paper, old versions and three templates for the same document? This is the biggest lever. A quote can change without a single requirement changing, just because we have seen the data.
Number of connections
ERP, document store, groupware, CRM. One interface or five is a different project. Authentication, permissions and rate limits have to be seen, otherwise the estimate is a guess. Most miscalculations live here.
Required accuracy
Some workflows are fine at 90 percent, others are only usable from 99 percent. For the second kind the testing effort grows by an order of magnitude. How much is needed is a business decision, not a technical one.
Who runs the system
After handover somebody touches it. If you run it yourselves, you need a manual and training. If we run it, a maintenance fee comes on top. Which is cheaper depends on usage and on your IT.
GDPR and the EU AI Act
As soon as personal data is involved, the record of processing activities is part of the job. Since August 2026 deployers also owe technical documentation under the EU AI Act. Both are work that belongs in the quote.
What barely moves the price
The number of screens and the number of features. Both are mostly rearranging things that already exist. Comparing quotes by feature list compares the wrong thing.
Where our projects sit
Ranges, not a price list. The fixed price for your case comes after the concept, and it holds.
| Project type | Price range | What is included | Time |
|---|---|---|---|
| Assessment | 0 EUR | Process recording, data review, benefit calculation, a recommendation, even when it says leave it | 2 weeks |
| Prototype for one narrow case | from 5,000 EUR | One process, one data source, real data, evaluated with your people | A few weeks |
| AI agent in operation | Five figures, fixed after the concept | Connection to your systems, approvals, audit trail, manual, handover | Weeks, then operation |
| RAG knowledge system | Five to six figures, depending on the number of sources | Connecting sources, index, search with citations, self test | Source by source, first one usable after weeks |
| Ongoing operation | Monthly, stated in the first quote | Hosting, model usage per request, monitoring, a named contact | From go-live, cancellable |
Not in the development price: server costs, usage based fees of external services, working time on your side. All three appear in the quote as separate lines.
A quote gets sharper in four stages
Anyone who names a firm figure in the first call has either seen your data already or is guessing. With us the path looks like this.
You can step off after any stage. Some people do after the second, with a number in hand and no invoice.
Rough estimate
In the first call, free of charge. The range is wide, roughly a factor of three between low and high. You still see whether the order of magnitude fits.
Assessment
After we have seen real data and the documentation of the interfaces. The range narrows to about a factor of two.
Fixed price
About two weeks after the requirements are settled. The amount is fixed. If we need longer, that is our problem, not yours.
Changes
If a requirement changes along the way, you get a quote for the difference. You decide whether the change goes in.
How we set the price
Five questions for every vendor, us included
Ask all of them the same way and the quotes become comparable. Whoever dodges one of these questions usually has a later invoice hidden there.
AI project costs: frequently asked questions
Related topics
AI consulting Munich
A free assessment: where AI pays off for you and where it does not, in writing within two weeks.
AI agency Munich
AI agents that run in production rather than in a demo. Seven steps, fixed price, on site meetings.
Process automation
Recurring work handled automatically, with approvals and an audit trail, from 5,000 EUR.
Tell us what should be built. We name the order of magnitude right away.
Thirty minutes are enough for a rough estimate. If we think the build is not worth it, we say so and why.