What does AI cost? Costs of AI consulting and implementation

In SMEs, an AI solution for a single process typically costs a low five-figure sum to implement. After that, there are monthly costs for model usage, hosting and maintenance. How expensive the project becomes depends mainly on connecting legacy systems and on the number of special cases. Below you will find the individual cost items, our fixed prices for prototypes and a calculation you can replicate with your own figures.

The five cost items

1. Licences for ready-made AI tools

If your team is to work with an assistant such as Microsoft 365 Copilot, ChatGPT or Claude, you pay per person per month. For up to 300 licences, Microsoft 365 Copilot costs €18.20 net per person per month as "Copilot Business" with annual payment. Until the end of 2026, a promotional price of €15.60 applies in the first year. The Enterprise version for larger organisations costs €26.00 (as of October 2026). A team of 25 people therefore pays €5,460 net a year for Copilot Business at the regular price.

Copilot Chat is included at no extra cost in eligible Microsoft 365 plans. Agents that access shared company data such as SharePoint there are billed by usage and are switched off by default.

A licence gives each person a tool for writing, summarising and searching. That alone does not implement a process. Whether the team makes good use of the tools in day-to-day work depends on training and clear rules.

2. Implementing an AI solution

An AI solution takes over a specific task: it reads order emails and creates a draft in the ERP system, checks invoices against purchase orders or prepares quotes. An ERP system manages business operations such as purchase orders and customer orders. For the AI solution, someone has to connect the model to your data and systems, build in checkpoints and test the workflows with real cases. This is the largest single item.

Our fixed prices for a working prototype in two weeks, each net:

The price depends mainly on how many systems are connected and how varied the incoming data and documents are. A prototype is not yet a solution for ongoing operation. It shows on your data whether the approach works before you invest more. The details are on the industry pages.

3. Model costs

When used through an interface, language models are billed by the volume processed. The unit for this is tokens, that is, parts of words. For typical office processes such as sorting emails or reading receipts, the amounts are small: with a few hundred tasks a month they are usually in the tens of euros. Long documents, images and top-tier models increase the costs. Ask your provider for a projection based on your actual volumes.

4. Hosting and operation

The solution needs a server, logs, backups and monitoring. If the server and the language model run in a data centre in Germany or on your own servers, costs rise somewhat, but your data does not leave the country. A German server alone is not enough for this if the model is called somewhere else. Then there is maintenance: models receive updates, interfaces change and someone has to respond when something fails. Ask for operation and maintenance to be offered separately from the project, with agreed response times.

5. Training and internal time

Also plan for your own team's time: for workshops to map the process, tests with real cases and the rollout. For a prototype, allow a few hours a week for a subject-matter contact person. Training for the team comes on top. Under Art. 4 of the EU AI Act, companies that use AI must take measures to ensure the AI literacy of their staff in any case. What is sufficient for this is covered in our article on the AI training record.

A calculation to replicate

Take a sales back office that transfers orders from emails into the ERP by hand.

• 600 orders a month
• 6 minutes per order today
• 2 minutes per order with an AI draft and review by the team
• €55 internal hourly rate

The team saves 4 minutes per order. With 600 orders, that is 40 hours a month, or €2,200 valued at the internal hourly rate. Over a year this comes to €26,400. Against this are €12,500 for the prototype, the cost of expanding it for ongoing operation, and ongoing costs. Whether the investment pays off depends on these total costs and on whether the team reaches the 2 minutes per order in a test with real emails. That is what the prototype is meant to show.

These figures are an example calculation, not a commitment. You can work out your own values with the time calculator.

When AI is not worthwhile

• The task occurs fewer than a few dozen times a month.
• Every case is different and requires a lot of experience.
• The necessary data exists only on paper or in people's heads.
• An error would be expensive and a human review takes as long as the work itself.

In these cases, a better template or a simple automation without AI is often the cheaper solution.

Reducing costs before you start

Check funding programmes before you award a contract. Many programmes require the application to be submitted before the contract. There is an overview in our article on AI funding. Start with the process where volume, time spent and the state of the data fit together best. Our scoring framework helps with this.

If you would like a projection for your own process, write to us via the contact form. The initial consultation takes 30 minutes and is free of charge.