AI agency, consultancy or your own team: who should build and run your AI

AI agency, consultancy, freelancer or own team: where is the difference?

The options differ in what you hold at the end and in who runs the system after go-live. The labels overlap in the market: some agencies advise, some consultancies build, and many firms call themselves both. So ask what you will receive and who will operate it. What the provider calls itself tells you little.

Buying a finished product is a further option, and often the right one when the work matches what the product already does. Our guide on build or buy covers that choice. Here the question is who builds and runs whatever you do not buy off the shelf.

Should a mid-sized company build AI in-house or outsource it?

In-house pays off when there is enough AI work to keep a team busy and when that team can also run what it builds. Outsourcing pays off for the first system, while you learn what kind of team you need. The hard part is finding the people. In Bitkom's 2026 survey (14 September 2026), companies that do not use AI yet named a lack of technical know-how (85%) and a lack of personnel (51%) as barriers. Bitkom reported on 3 September 2026 that 79000 IT specialists are missing in Germany and that 97% of companies have trouble filling IT posts. Owners see the gap but rarely plan to close it by hiring: in the DZ Bank Mittelstand study reported by Börsen-Zeitung on 24 September 2026, more than a quarter of owners reported missing skilled staff, but only 7% want to hire AI experts.

A team that runs AI in production needs more than one model specialist. Someone owns the platform and the access to models, someone keeps the test set current and runs it after each change, someone handles security reviews and incidents, and someone in the business owns the process and the review queue. Count those roles before you compare a salary with a project fee. So far few companies do this: in ifo's business surveys of May 2026, as reported by retail-news.de, only 18.7% of companies build their own AI systems.

• Own team first: AI is in the product you sell, the work never runs out, and you can offer engineers interesting problems and a career path.
• Partner first: AI supports internal processes, you have one or two clear candidates, and nobody in the company has run an AI system in production before.
• Both: a partner builds the first system with two or three of your people, who then become the core of your own team.

Consulting or implementation: a PDF or something that runs?

If the goal is less routine work, you need something that runs. A recommendation alone changes nothing in the process. Bitkom's president Ralf Wintergerst put it this way on 14 September 2026: "Wer nur eine Lizenz kauft, hat KI eingekauft, aber noch nichts verändert", whoever only buys a licence has bought AI but changed nothing yet. The same applies to a roadmap. The step that most often fails is the one from pilot to operation. In Lünendonk's study of DACH upper Mittelstand and large firms, reported by CIO.de on 17 September 2026, 66% said fewer than a quarter of their piloted agents leave the test lab, and the author's verdict was "Was fehlt, ist die operative Konsequenz".

A consultancy that only advises still has a place: an independent second opinion before a large investment, a portfolio decision across many business units, or the preparation of a tender. Then ask how its recommendation will be tested on your own work and by whom. Consultants make the same point about their own trade. One wrote in Business Punk on 19 July 2026 that "Ein Prototyp, den man anfassen kann, sagt mehr als 90 Folien", a prototype you can touch says more than 90 slides.

How do we choose an AI partner? Questions to ask and evidence to request

Choose the partner that can show, on your data, that the system works and that you can run it without them. German IT buyers already weigh evidence over claims: in Akima's 2026 study of 309 IT decision-makers, reported by Computerwoche on 29 June 2026, the top trust factors were visible in-house experts (36%), verifiable reference customers (32%) and certifications and partnerships (30%). The table lists the questions that separate a system you will own from a dependency.

The questions for the first call itself, and how to judge a small first step, are in our guide on choosing an AI consultancy. Published checklists point the same way: ADVISORI's ten criteria (24 April 2026) include production and MLOps competence, technology neutrality, and references with measurable results. Bring purchasing, IT, legal and the business unit to the table together. The purchasing association BME's practice forum on buying AI (July 2026) puts exactly that on its agenda, along with "Lock-in früh erkennen und aktiv begrenzen", spot lock-in early and limit it actively.

Can we move from an agency to our own team later?

Yes, if the handover is planned from the start. Bolted on at the end, it rarely works. Planning it early also answers the other worry buyers raise: what happens if the provider is no longer around? A system that your team can take over also survives the provider. A build with handover has five parts.

1. 1

   Your accounts and repositories from day one

   Code, configuration, prompts and infrastructure live in your repositories and your cloud account or on your own servers. The partner works with access you grant and can revoke.

2. 2

   Your people build alongside

   Two or three of your staff work on real tasks with the partner's engineers, review every change and make some of them. That way the knowledge moves over during the work itself.

3. 3

   Tests and measurements come with the system

   The test set built from your cases, the acceptance test results and the measurements before and after go-live are handed over and run in your pipeline.

4. 4

   Runbooks your team has rehearsed

   Your team handles a model update, a failed integration and a quality drop with the runbooks while the partner watches, before it handles them alone.

5. 5

   Operation passes over in agreed steps

   The contract names the steps and what has to be true before each one, and keeps the partner available for questions afterwards.

Who owns which part after the handover, and how to write the exit into the contract, are covered in our guides on AI system ownership and on the AI vendor exit plan. One consultant's line in Business Punk (19 July 2026) sums up the risk on the other side: "Wer da nur ein Abo verkauft, verkauft eine Abhängigkeit, kein Werkzeug", whoever only sells a subscription sells a dependency, not a tool.

Which option fits your situation?

Start from your own situation and pick the provider type after that. The matrix covers the cases we see most often. Where two rows apply, the stricter answer wins.

Where Cloudsail fits, and where it does not

Cloudsail is an engineering consultancy with engineers in Germany and Poland. We build and run AI from the servers to the agents: infrastructure in your own cloud account or on your own servers, company data brought together with its permissions, and agents that automate work with human approval, each tested before go-live and after every change.

The way we work follows the build with handover described above. It starts with a free 30-minute call that needs no access to your systems or data. We pick one process, measure how the work runs today, build with your people in your repositories and check the result in an acceptance test on your own cases. Then you choose: your team runs it with the code, tests, runbooks and measurements, or we run it for you. Our engineers can also join your team for longer. Each step has a fixed price, agreed in writing after the first call, and no model or tool vendor pays us.

We are not the right choice for a strategy document without a system behind it, for a programme that needs a large team across many countries at once, or for automating decisions about applicants or credit without people. If one of those is what you need, the table above shows which kind of provider to look for.

AI strategy and rolloutAI engineers for your teamHow we work