Clef · Try the live demo →
Heabsy

What a project costs and what for

Two stages with a stated lower bound, no hourly rate and no charge per user. Below are the four things that move the number, and why we decline smaller scopes rather than doing them badly.

What a project costs

Starter

First step into enterprise AI

from €5K/ per project
  • GenAI Readiness Audit
  • 1 Use Case Discovery Workshop
  • Proof of Concept (1 use case)
  • Recommendations & roadmap
  • 2 weeks post-PoC support
  • —Production deployment
  • —On-premise installation
  • —SLA support
Start with an Audit
Most Popular

Growth

Into production

from €25K/ per project
  • Everything in Starter
  • Production deployment with measured answer quality
  • Connection to your existing storage and permissions
  • On-premise or our own EEA hardware
  • Procedure for new and changed documents
  • Training and a handover record
  • Documentation an EU AI Act assessment needs
  • —A team assigned exclusively to you
  • —Service levels beyond business hours
Discuss Your Project

Enterprise

Ongoing arrangement

Custom/ annual contract
  • Everything in Growth
  • Named people assigned to your account
  • Service levels set in the contract for your setup
  • Reviews at agreed intervals
  • New models evaluated on your own questions first
  • Several use cases in one arrangement
  • Operations handled by us or handed to your IT
Contact Us

💡 Not sure which plan fits? Start with a free 30-minute consultation — we'll help you define the scope and budget for your project.

Book a Free Consultation

What moves the number

Four things, and the model is not one of them. This is also the order in which they matter.

The state of your documents, not their volume

A base where three versions of the same contract sit side by side with none marked current costs more than twice the page count. Sorting that out is work, and it is usually the first finding of Discovery.

How many systems have to be connected

One store is straightforward. SharePoint, a network share and a line-of-business system together mean three access methods, three permission models and three places where something changes without notice.

How strictly you must prove who saw what

A regulated institution needs to show who accessed which document and when. Logging and evidence are a separate part of the work, not a premium feature.

Who runs it afterwards

Handing operations to your IT is cheaper to run and more expensive to prepare: it has to be documented and taught. Keeping it with us is the opposite arrangement.

Why not an hourly rate?

Because of who carries the estimating risk. With an hourly rate it is the client: the longer the work takes, the more they pay, and they never know in advance whether it will be two hundred hours or four hundred.

With a price per stage the risk is ours, which is why it is worth arguing about scope before the project rather than during it. That argument is the useful part: it forces both sides to say what «done» means while it is still cheap to disagree.

It also explains the lower bound. Connecting storage, taking over permissions and training people is the same amount of work for ten users as for a hundred — so below roughly €25,000 an implementation does not work out for either side, and we would rather decline it than deliver it badly. What is available below that is Discovery, which is a smaller and honest piece of work with a written outcome.

Frequently Asked Questions

What is the minimum to get started?+
Discovery starts at €5,000: an audit of your documents, the choice of a first use case, a working trial on your own data and an effort estimate. It ends with numbers you decide on — including when those numbers argue against the project.
How much does an enterprise RAG project cost?+
Discovery with a document audit starts at €5,000 and ends with a measured figure: the share of your questions the material can actually answer. A production RAG deployment on existing infrastructure starts at €25,000. Inference afterwards runs on our own hardware, so the running cost does not scale with the number of people asking.
Why does an implementation project start at €25,000?+
Below that nothing works out for either side. Connecting storage, taking over permissions and training people is the same amount of work for ten users as for a hundred, so a smaller scope gets done badly or not at all. We would rather decline it than rush it.
Why not charge by the hour?+
With an hourly rate the client carries the estimating risk: the longer it takes, the more they pay. With a price per stage we carry it, which is why it is worth arguing about scope before we start rather than after.
Is the price per user?+
No. Inference runs on our own hardware, so the cost does not grow with headcount or with the number of questions asked. At a hundred people this is the difference between a fixed project cost and a subscription that increases with every hire.
What drives the price up?+
Four things, and the model is not one of them: the state of your documents rather than their volume, the number of systems to connect, how strictly you have to prove who saw what, and who runs the system afterwards. Several versions of the same document with none marked current costs more than twice the page count.
How long does it take?+
A test on a single document takes a day. A pilot two to four weeks. Production another four to eight, depending on how deep the integration goes. An on-premise deployment inside your own network adds two to three weeks.
Can it run entirely on our own servers?+
Yes. Open-weight models — Qwen, Mistral, DeepSeek, GLM and others — run on your hardware, and your documents do not leave your network. By default inference runs on our own hardware in the EEA instead, which removes the hardware investment without moving data to a US provider.
Do you work with companies outside the German-speaking region?+
Yes, including the United States. The arrangement does not change with geography: inference on our own hardware in the EEA, or on-premise in your own environment. We do not route documents through a US provider for one market and not another — that would make the data-location answer depend on who is asking.
Which systems do you integrate with?+
The ones your documents already live in: SharePoint, Confluence, Google Drive, network shares, document management and line-of-business systems such as DATEV or BMD. Where only a user interface exists and no API, we say so before the project rather than during it.
What happens if Discovery comes out negative?+
You get that in writing with the reasons. It is a legitimate outcome: better to pay for finding out that your document base does not support this than for a project that stalls on the same finding halfway through. In that case we do not propose the large project.
What is included in ongoing support?+
Monitoring of answer quality against a fixed set of questions, handling of new and changed documents, incident response and a review at agreed intervals. Service levels are set in the contract for your setup — we do not publish a blanket uptime figure, because one that is not tied to an architecture is not worth anything.
Do you guarantee EU AI Act compliance?+
No, and nobody can. What we can do is produce what an assessment needs: documented processing locations, logs of what the system answered and on what basis, the classification of the use case, and training records for the people working with it. Whether that satisfies a supervisor is a judgement your own people make and sign.

Ready to Discuss Your Project?

Tell us about your challenges — we'll propose a solution and estimate costs within 24 hours.

Request a Project Estimate

Try the API yourself

Sign up in the console, create a key and send a test request. OpenAI- and Anthropic-compatible, billed per token; the catalog shows for each model whether it runs on our own hardware in the EEA.

Sign up and try the API Models and prices