AI consulting for German companies: what it delivers

Why this decision is sitting on your desk right now: we observed the same pattern across every company that asked us about AI this year, and the number that describes it is not the adoption rate. Bitkom's representative survey of 604 German companies with 20 or more employees, published on 11 March 2026, found that 41 percent now use AI and another 48 percent are planning or discussing it. In the same survey, 52 percent of the companies already using AI report a measurable contribution to business results. Roughly half of the adopters cannot point at one. That gap, between running AI and being able to name what it returned, is the real bottleneck, and it is what a consultancy is actually bought for. The budget question, the process owner and the approval logic decide it long before the model does.

What does AI consulting actually deliver?

Good AI consulting turns your operations into a short list of concrete use cases, ranks them by effort and return, and stays for the implementation of the first one. The output is a decision, not a technology showcase: which process gets improved first, measured against which number, by when.

That sounds modest until you compare it with the alternative. Bitkom's March 2026 numbers show what happens without the ranking step. A third of the companies using AI, 33 percent, found that it drove costs significantly higher than they had expected. Nineteen percent say they cut jobs because of AI. Those are not the marks of a technology problem. They are the marks of decisions taken without a business case underneath them.

We deliberately do not start with a tool comparison. A tool comparison is answerable by anyone with a browser, and it answers a question you do not have yet. The question you have is which of your processes carries enough repeated, rule-bound work that a machine can take a defined slice of it, and who inside the company owns the result once it does.

When is it worth it for a mid-sized company?

It is worth it when the internal time or the internal knowledge to pick the right first use case is missing, which is the normal state rather than the exception. The evidence for how normal it is comes from KfW Research.

In its release of 11 February 2026, KfW reported that 20 percent of German Mittelstand companies used AI between 2022 and 2024, against 4 percent between 2016 and 2018. In absolute terms that is close to 780,000 companies. The spread by size is the part worth reading twice: 36 percent among firms with 50 or more employees, 19 percent among firms with fewer than five. By sector, knowledge based services lead at 28 percent and research intensive manufacturing follows at 23 percent.

Put the two datasets next to each other and the picture sharpens. Adoption is no longer the constraint. Direction is. A company that has already tried something and has nothing to show for it does not need another pilot; it needs someone to name the one process where the return is countable and to stay until the number moves.

What separates good consulting from expensive theatre?

Three things are visible from the outside before you sign anything. The first is where the conversation starts. A consultancy that opens with the business problem and asks about your approval chains, your error tolerance and your data access is working on the thing that decides the outcome. One that opens with a model comparison has skipped it.

The second is what you are handed at the end. A ranked list of use cases with one of them costed, owned and scheduled is a decision. A strategy deck is a description of a decision that somebody else will have to take later.

The third is what happens after the workshop. Implementation is where the 33 percent cost overrun in the Bitkom data is generated, because that is where integration, permissions and exception handling turn up. A consultancy that leaves before that point has sold you the easy half.

What does AI consulting cost in Germany?

Day rates and project prices are documented well enough to plan against. Sven Jagata's German pricing guide, published on 20 May 2026 and last updated on 11 August 2026, puts hourly rates at 80 to 120 euros for a junior or freelance consultant, 140 to 220 euros for an experienced certified consultant, 220 to 350 euros for a senior with ERP depth, and 320 to 480 euros at the large firms.

Priced by package rather than by hour, the same guide lists a strategy workshop at 1,500 to 4,500 euros for one to two days, a pilot project at 3,500 to 8,000 euros over four to eight weeks, a mid-sized implementation at 12,000 to 25,000 euros over two to four months, a full integration at 25,000 to 80,000 euros over four to nine months, and ongoing support at 350 to 1,800 euros a month.

Read those ranges as a test rather than a price list. If a proposal lands at the top of a bracket, the thing to ask is which of the brackets below it was ruled out and why. The answer tells you whether anyone did the ranking work at all.

Which subsidies apply, and what do they change?

Consulting for small and mid-sized companies is frequently co-funded in Germany, and the share is large enough to move a decision. The same pricing guide records the BAFA framework as covering 50 percent of the fee up to 1,750 euros in western Germany and up to 80 percent up to 2,800 euros in the east.

What that changes is the size of the first step, not the shape of it. A funded engagement still has to produce a named use case with a number attached, and the funding conditions change often enough that they belong in the conversation before the contract rather than after it. Check the current terms for your programme, then decide the scope.

How do you choose, and what should you ask?

Ask for one example from a company that looks like yours, and ask what it cost, what it returned and how long it took to return it. A consultancy that can name a process, a metric and a date has done the work before. One that answers with a client logo has not.

Then ask two harder questions. Ask which use case they would advise you against, because a ranking that contains no rejections is not a ranking. And ask who owns the result inside your company once the engagement ends, because that is the seat the Bitkom cost overruns are created in.

If somebody offers you a large programme in the first meeting, the ranking step has been skipped, and you will be the one paying for the discovery that should have happened first.

How long before anything moves?

Three to six months to break even is the realistic planning figure for a clearly defined use case, provided the engagement ends with something running in production rather than a concept deck. That condition carries most of the variance.

Sven Jagata's guide, updated on 11 August 2026, puts numbers against individual cases. Invoice processing is the fast one, at a 200 to 400 percent return with amortisation in 8 to 14 weeks. An AI phone assistant breaks even at 50 to 70 calls a month and saves 50 to 80 percent against an external answering service beyond that point. No-code pilots, in the author's own practice, show positive results after two to four months.

The planning consequence is unglamorous and worth stating to a board. Plan year one to break even, and expect the margin improvement from year two. A business case that promises a return inside the first quarter is either an unusually narrow use case or an unusually optimistic spreadsheet, and it is worth finding out which before you sign.

Where should the first project sit?

Put it where the work is repetitive, the rules are written down somewhere and the error is cheap to catch. Offer preparation, inbound classification, document extraction and first level support all qualify, and the KfW data shows why: the applications that spread fastest in the Mittelstand are the broadly usable ones, language generation and text recognition, rather than the ambitious ones.

Set one number before you start, and set it small. Handling time per case, share of tickets closed without escalation, days from enquiry to quote. If the number does not move within the first engagement, you have learned something worth the fee. If nobody agreed on the number, you have bought a pilot that can never end, which is the most expensive outcome available and the easiest one to avoid.