Key takeaways
- "AI agency" describes three different purchases. Consulting and training, a chat platform on subscription, or software built as a project into your systems. Only the third delivers an agent that takes over a job. Settle which one you mean before the first call.
- A price on the website is the exception. Four of nine vendors publish an entry price, two publish a price for something built, one publishes a price for operations. "On request" means, in practice: budget double the cheapest published price.
- A logo is not a result. All nine have logo walls. Three have a number next to a customer name, and for two of those the number comes from the vendor. Ask for the customer you are allowed to call.
- Almost nobody stays accountable in month four. The prototype ends with a handover, the platform with a maintenance contract for the platform. Who answers for evals, model changes and regressions after go-live is the question this list is sorted by.
- For a chat platform, an agency is the wrong address. Langdock, CompanyGPT and the like solve a different problem from an agent. Which one is explained below.
How we evaluated
The rankings you find for "AI agency comparison" share two traits: the publisher sits at number one, and no competitor carries a price. We keep the first, because otherwise we would be lying, and we disclose it. We break the second. Every vendor below carries its public price or a note that it publishes none, and each gets the case named in which it is the better choice than us.
The list is sorted by what a company with 50 to 1,000 employees needs when one process eats too much time and the internal champion can no longer carry it alone. Four criteria, in this order, and the same four questions you should put to every vendor:
- Does it build the agent as software into your systems? Or does the engagement end with a concept, a prototype, or a platform you fill yourselves?
- Does it stay accountable after go-live? Evals, monitoring, model changes, regressions, with a metric in the contract.
- Is there a price on the website? For the rung you actually buy, not the workshop before it.
- Is there a named customer with a number? Not a logo, but a result someone confirms by name.
A note on perspective: twigbit writes this list, and we are on it. We evaluated against public pages, prices, references and register data, checked on 10 September 2026. We have not spoken with every vendor and have hired none. Large houses with tender experience, TISAX and hundreds of consultants have their own strengths; for a DAX procurement they are sorted correctly, for the buyer described here they are not. Treat the list as a starting point, not a verdict.
The nine vendors in detail
1) twigbit, Berlin
What separates us from the other eight is where our work ends: not at a concept, a prototype or a platform, but at an agent running in your systems that we answer for afterwards. Emil Bergold and I founded twigbit in 2018. We are six people, all senior full-stack engineers, and twigbit is a software agency: we build websites, apps and data infrastructure, with one clear focus, that all of it is ready for agents. In this comparison we compete on one of those offers: an agent with one job, built into your systems as software and kept running by us.
Concretely, it looks like this. The agent has a trigger: an incoming ticket, an order arriving as a PDF, an empty CRM field. It has tools with approval boundaries: what carries risk waits for a person, what carries none runs through. It has an eval set built from your historical cases that checks, on every change to prompt, model or data, whether the output still holds. And it has a metric we report against every month. You keep identity, permissions, data and approvals; we keep code, evals, model migrations and regressions under a maintenance contract, like any software vendor.
The results are named. For Projektron, a German software vendor with hundreds of customers, a support-triage agent reads every incoming request, classifies it, drafts a sourced answer from the knowledge base and routes it; on low confidence it escalates to a person. It runs against the real ticket stream, and the CEO says it paid for itself within months. For Thermondo we cut tracking loss by 80 percent and lifted return on ad spend by 15 percent. For A2Mac1 an AI optimisation tool for supply chains is in use.
Prices are on the website: €2,900 for the one-week discovery sprint, credited in full against a deployment within 90 days, from €7,900 for the built agent, and a monthly operations fee the proposal states next to the variable model costs. We work from Berlin for companies across the German-speaking market.
One honest caveat: we are not the right fit for everyone. If you are running a tender with minimum sizes, need a programme with 40 consultants, are bound to TISAX, or are looking for training under Article 4 of the EU AI Act, the right vendor is further down. And if your people do not have a chat platform yet, buy one first, not us.
Best for: companies with 50 to 1,000 employees that want to get rid of one concrete process and see a number afterwards. Pricing: €2,900 sprint, deployment from €7,900, operations as a monthly fee in the proposal.
2) MaibornWolff, Munich
By far the largest house on the list, around 780 people, with AI agents as one of 13 service lines and a clear seven-step process from workshop to operations. The real strength is enterprise engineering in regulated environments and a clean handover to an in-house IT that takes over operations. That is exactly where the fit divides: for TÜV NORD an internal assistant was co-developed that TÜV NORD now runs itself. The only public price is the workshop; build and operations come as a proposal. For a 200-person company with a process problem, the frame of a 780-person house is rarely the right size.
Best for: corporates and large mid-market firms with tenders, compliance obligations and an IT department that wants to take over operations. Pricing: workshop from about €5,000, everything else on request.
3) statworx, Frankfurt
A Frankfurt AI consultancy with the most transparent price ladder among German consulting houses: impact assessment €4,900, use-case workshop €6,490, agent prototype €14,490 for up to ten person-days. If you want a priced basis for decision before a build and an OpenAI partner badge in procurement, this is the clearest offer. The ladder ends at the prototype, though, whose stated purpose is to help you decide whether an investment is worthwhile. Production and operations are a follow-on project without a public price, and the agent pages carry no customer outcome number.
Best for: companies that want a priced prototype before building and a consultancy with partner status. Pricing: €4,900 / €6,490 / €14,490, bundle €19,490.
4) innFactory, Rosenheim
A cloud and software house with CompanyGPT, Germany's most transparent "your data in your cloud" chat platform: a hardened open-source installation with RAG, an agent catalogue and n8n in your own tenant, at a fixed price with no seat fee. If sovereignty is the requirement, this is a serious option, and innFactory was first to market with it, ahead of everyone including us. It is a platform your team fills: no published definition of agent, no eval feature, automation is built in n8n. There are no customer outcome numbers. Agents as a project innFactory builds without a public price.
Best for: hospitals, municipalities, banks and mid-market firms with a Microsoft tenant, a sovereignty mandate and their own IT, from about 60 users. Pricing: CompanyGPT €14,990 plus €399 a month plus cloud and tokens; projects on request.
5) Merantix Momentum, Berlin
The delivery arm of the Merantix ecosystem, with a ladder from AI strategy through development to operations, and with the strongest named cases on the list in the public sector and industry, including TÜV Rheinland with a 94 percent time saving in claims search, per the vendor. The focus is on corporates and public administration with programme budgets and proximity to research. There are no public prices, and for the buyer described here the entry bar is high.
Best for: corporates and public authorities with research proximity and programme budgets. Pricing: none published.
6) Birds on Mars, Berlin
A Berlin AI consultancy of around 20 people, ISO 27001, with strong references in media, healthcare and mobility, part of the Alexander Thamm group from Munich since July. The strength is Berlin consulting proximity backed by a 500-person group. No public prices, no outcome number with a customer name, and a post-acquisition positioning still taking shape. If you want the former boutique profile, ask about the current one.
Best for: companies that want Berlin consulting proximity backed by a large group. Pricing: none published.
7) One Thousand, Berlin
A Berlin AI consultancy with a four-rung "How to AI" ladder from feasibility hackathons to scaling, and with forward deployed AI engineers who sit with the client. The promise to build capability in-house rather than create dependency is likeable and means, at the same time, that operations end up as your job. The website carries no prices, no named customers and no outcome numbers.
Best for: companies that first need momentum and strategy and want delivery accompanied by embedded engineers. Pricing: none published.
8) Ventum Consulting, Munich
A Munich management and IT consultancy, TISAX and ISO certified, with an AI potential analysis as a workshop, agent consulting on strategy, architecture and governance under the EU AI Act, and MVP builds. The strength is governance and process consulting from one source. Agents are a new service line in a process and automotive house; the agent page ends at the pilot, and named AI customers with numbers are not on the website.
Best for: companies that want AI governance and process consulting from one source and will place delivery internally or elsewhere afterwards. Pricing: on request.
9) AI FIRST, Hamburg
A Hamburg enablement provider with a twelve-month executive programme, video courses and one of the widest-reaching founder voices on LinkedIn, with mid-market references such as ARDEX, CEWE and GEDORE. AI FIRST builds no agents and operates no software. It is on this list because buyers find it in the same searches, and because a good enablement programme can be the right first step before anyone builds an agent. The website names no prices.
Best for: executive teams that want to train their organisation in working with AI before commissioning software. Pricing: none published.
How to spot an offer that ends at the prototype
Four tests you can apply in any first call. They are the same four criteria as above, asked from your side of the table.
- Ask what is handed over on the last day of the project. If the answer is a concept, a maturity model or a prototype with a manual, you are buying consulting. An agent that runs on the last day and is still running the day after is software.
- Ask who is on call in month four when the supplier changes its PDF layout. If the answer is "your team" or "a maintenance contract for the platform", nobody is accountable for the agent. If the answer is a name and a metric, the vendor is.
- Ask for the price of the rung you are buying. A workshop price on the website says nothing about the build. A vendor that does not price the build has either not done it often enough to know, or wants to decide it in the conversation.
- Ask for the customer you are allowed to call. A logo on a website is permission to display it. A number next to a name that someone confirms on the phone is a result. Of nine vendors, three can deliver that.
Platforms and model vendors that are not agencies
Searching for "AI agency" returns Langdock, CompanyGPT, ChatGPT Enterprise and Microsoft 365 Copilot in the same results. They are not agencies, and they solve a different problem: they give your people a compliant place to work with language models, per seat or as a flat fee, with assistants that employees configure themselves. That is a good purchase, and for most companies the first one. It does not replace an agent that takes over a process, and an agency does not replace a platform. If your people do not have one yet, buy the platform first. If they have one and a process is eating the week, you are in the right place on this list.
What an AI agent costs in Germany in 2026
A range can be read from the public prices on this list. Entry, a workshop or assessment, runs from €2,900 to €6,490. A prototype that helps you decide costs €14,490. A built agent running in your systems starts at €7,900. A sovereign chat platform for everyone costs €14,990 plus €399 a month plus cloud. And operating an agent, evals, monitoring, model changes, is a priced item in the proposal at exactly one vendor on the list. Anything labelled "on request" sits, in our experience, at double the cheapest published price for the same rung.
How to choose an AI agency
- Start with a process, not a strategy. An inbox, an intake channel, a CRM field. With measured volume, so that "better" has a number.
- Ask for the definition of agent. If the answer is "a chatbot with knowledge", you are buying a platform. If it includes trigger, tools, approval boundary and evals, you are buying software.
- Demand the price for the rung you are buying. Not for the workshop before it.
- Settle ownership. Who owns code, evals and prompts if you leave?
- Do the maths. An agent from €7,900 plus operations has to save more than it costs. For a process that costs one person two hours a day, in our experience that happens within months. The sprint exists to run that calculation beforehand.
Conclusion
The German market for AI agents in 2026 is full of good consultants, serious platforms and very few vendors who build an agent as software, integrate it into your systems and then answer for the number. Every vendor on this list has the better answer for a particular buyer, and we have named that buyer each time. What matters in the end is not the most impressive website or the longest logo wall, but who can show you a running agent at a customer you are allowed to call.
We sit at number one because we sorted the list by the buyer we are built for: 50 to 1,000 employees, one process, one number, one price on the website. If that is your case, the one-week discovery sprint for €2,900 answers which process it should be and is credited against the deployment. If not, the right vendor is in the table above.
Disclosure: twigbit writes this page and is on it. Every number comes from the vendors' public pages, checked on the date given. Corrections to post@twigbit.com.