Choosing an AI service provider: 14 questions to ask before you sign
When choosing an AI service provider, focus less on models and demos and more on three interfaces: where the system runs, whether the implementation actually reaches production at the agreed price, and who then operates the system with written commitments. Add to that a defined exit, references with figures and a clear data classification.
Image: Nacho Kamenov & Humans in the Loop / Data annotators discussing the correct labeling of a dataset / Licenced by CC-BY 4.0
For decision makers in brief
- Effort: A structured selection with a checklist takes only a few meetings per provider. It does not replace an assessment of your starting point, but it prevents you from buying a pilot that never reaches operations.
- Cost: According to the industry report, there are no reliable public benchmarks for project costs in the German Mittelstand, mid-sized companies with roughly 50 to 1,000 employees; the figures in circulation come from marketing copy. Compare scope of services and pricing logic per phase instead of individual amounts.
- Risk: The biggest risk lies in the handovers: between data centre and implementer, between project and operations, between contract and exit.
- Result: A checklist with 14 questions, each with the mark of a good answer and the typical red flag.
Why selection fails at the interfaces
Artificial intelligence has arrived in industry. 58.7 % of companies in manufacturing use it (ifo Business Survey by the ifo Institute, May 2026). According to the consultancy study Deloitte AI in Manufacturing 2026, however, only one in five use cases has been scaled across the company. In our analysis of 52 published practitioner conversations (September 2025 to August 2026), 42 of 206 pain point statements concern data, IT/OT and integration. Technology as such hardly appears as a problem.
If you have to choose an AI service provider, draw one conclusion from this: the question of the best model is rarely the decisive one. What matters is who is responsible for the connection to your systems and who operates the system after rollout. This is exactly where the market splits into roles. There are data centre operators that provide capacity but implement nothing. There are consultancies that deliver concepts but operate nothing. There are implementers that build a pilot and leave after acceptance. Each of these roles can do good work on its own. The problems arise where one role ends and the next begins: who is liable when the connection to the ERP breaks after an update? Who retrains when the results decline?
Our analysis reveals a blind spot here. There is a lot of talk about use cases, and almost none about where the model runs, who operates it and how operations are safeguarded. We have therefore shaped our own offering so that infrastructure, implementation and operations sit with one responsible party. Our overview of AI implementation for mid-sized companies shows how this fits together. Whether it suits you, however, is something you should check with the same questions you ask every other provider.
Eight criteria for choosing an AI service provider
Infrastructure: where does the model run?
77 % of companies name data protection requirements as the biggest obstacle to AI and digitalisation (Bitkom, the German digital industry association, 2026). At the same time, 71 % use cloud services from US providers, although only 8 % would prefer them (Bitkom Cloud Report 2026). Ask specifically: in which country is the hardware located, who operates it, and which subcontractors have access? Is every use logged so that an audit can trace it? A good partner answers with names and locations, not with the word sovereign.
Implementation at the agreed price
A proposal should show assessment, pilot, rollout and operations separately, each with scope of services, duration and price range. Pay particular attention to one item that is often missing. According to the consultancy study Capgemini and Microsoft 2025, 80 % of data and AI initiatives in industry require upstream investment in data architecture, OT/IT integration and interfaces. These costs rarely appear in provider proposals because they are not part of the product. Ask whether they are included. We disclose our own pricing framework under AI project fixed price: 72 hours to the first internal pilot, 4 to 10 weeks to productive implementation.
Operations with commitments
Models deteriorate over time. If no one is assigned operational responsibility, the benefit is lost again after about twelve months, according to the industry report. Ask who monitors model quality after rollout, who retrains and who decides on a shutdown. Response times per severity level belong in the contract, not in the presentation. Distinguish between a general operations target and a contractual commitment. We work with a general operations target of 99.9 % availability, which is expressly not an individual guarantee per customer. On our operations page, we describe what well organised AI operations as a managed service look like.
Exit: what do you take with you?
Most contracts describe the entry in detail and the exit hardly at all. Clarify in advance in which format you receive model weights, data, configuration and documentation, within what period and at what cost. The industry report expressly names interchangeability of the model as a criterion for sovereignty, so that the dependency does not simply shift. We have published our own process for switching AI provider so that you can compare it.
References with figures
Only 14 of the 52 analysed practitioner conversations mention a reliable figure at all. Anyone who presents measured results stands out. So do not just ask for logos; ask for a named case with a starting value, result and measurement method. Pay attention to the type of source: independently verified, company statement or provider statement. Also, do not confuse published industry examples from large corporations with a provider's own customer projects. We too only publish references with measured results once the customer has approved the content and the figure.
Handling the baseline
44 % of machinery and plant manufacturers name an unproven return on investment as an obstacle (VDMA, the German Mechanical Engineering Industry Association, 2025). Proof only works if the starting value is fixed before the pilot: unplanned downtime minutes, processing time per case or scrap rate. A good partner asks for it on their own initiative and insists on it. Anyone who tries to reconstruct success after the fact delivers a figure that controlling will not accept.
Data classification
Which data may leave the company and which may not shapes architecture and cost more than any choice of model. In 39 % of industrial companies, AI is merely tolerated (ZEW, the Leibniz Centre for European Economic Research, 2026), and there nobody knows which data flows where. A serious provider clarifies this question with you before proposing an architecture, not afterwards.
Co-determination
AI systems that can record the performance or behaviour of employees touch on co-determination rights. Bitkom published a guide on this in February 2026. In practice, it has proven effective to involve the works council before the pilot, because a works agreement negotiated after the fact is regularly more expensive. Ask whether the provider supplies documents for this, for example a description of which data the system processes.
Checklist: 14 questions to take with you
This AI vendor checklist is designed so that you can ask every provider the same questions, ideally in writing. The table shows how to recognise a good answer and which answer is a red flag.
| No. | Question for the provider | How to recognise a good answer | Red flag |
|---|---|---|---|
| 1 | Where does the model run, and who operates the hardware? | Country, location and operator are named, own servers or a contractually fixed zone in the EU | General statements such as secure cloud without a location |
| 2 | Which subcontractors have access to data or metadata? | Named list in the contract, changes are announced in advance | Reference to general terms and conditions |
| 3 | Is every use logged, and do we get access? | Logs are accessible to your audit function | Logs only internal at the provider |
| 4 | What does each phase cost up to production? | Separate packages with scope, duration and price range | Only one price for the pilot, open after that |
| 5 | Is preparatory work on data and interfaces included? | Explicitly named, with assumptions about the data situation | We will clarify that later |
| 6 | How long does it take to the first result and to production? | Specific time frames per phase, with prerequisites | Only a total duration without interim milestones |
| 7 | Who monitors, retrains and decides on a shutdown? | Named responsibility after rollout, with a reporting line | Operations are not part of the proposal |
| 8 | Which response times apply per severity level? | In writing in the contract, with an escalation path | Availability figure without contractual basis |
| 9 | What do we receive on exit, in which format and at what cost? | Model weights, data, configuration and documentation with a deadline | Export only as an extra service charged by effort |
| 10 | Can the model be replaced? | Models with open weights or a justified alternative | Tied to a single model without a plan B |
| 11 | Which named case proves your work with a figure? | Starting value, result, measurement method and type of source | Logos without figures or third party industry examples |
| 12 | How do we fix the baseline before the start? | One metric, taken from the system before the pilot | Success is estimated after the pilot |
| 13 | Which data may leave the company, and who decides? | The question is asked before the architecture proposal | Architecture is set before the data is clarified |
| 14 | How do we involve the works council and data protection? | Documents on the data processed, involvement before the pilot | The topic only comes up at rollout |
A single red flag on questions 7 to 10 weighs more heavily than a gap in the choice of model. Operations and exit can hardly be renegotiated after the contract is signed, whereas a model can be replaced if the architecture provides for it.
How to use the checklist in your selection
- Clarify internally first: Define data classification, the first use case and its baseline before you approach providers. Otherwise you will be evaluating proposals for different projects.
- Ask the same questions: Send all providers the 14 questions in the same form. This is the only way to make the answers comparable.
- Rate the answers: Classify each answer as proven, committed or open. Proven means there is a document, contract wording or a verifiable case.
- Weight the red flags: Open points on operations, exit and price up to production are reasons for exclusion; open points on technical details usually are not.
- Start narrow: Agree on a single use case measured against the baseline. Start narrow, finish measurably, repeat: this is how the industry report summarises the proven sequence.
Frequently asked questions
Should we choose one provider for everything or several specialists?
Both can work. With several providers, you have to manage the interfaces yourself, which means defining who is responsible when an error occurs at an interface. If you do not have the internal capacity for this, there is a strong case for an AI partner with bundled responsibility.
How can we tell whether a reference is reliable?
By a named starting situation, a result figure with a measurement method and the type of source. A company statement is worth more than a pure provider statement, and an independent review more than both. Also ask whether the provider was involved itself or is quoting someone else's example.
Is a fixed price realistic for AI projects?
Yes, for clearly defined phases such as assessment and pilot. The price for rollout and operations depends on the data situation, the number of use cases and existing infrastructure, and it should be fixed in writing after the assessment.
What role does the choice of model play in the selection?
A smaller one than many presentations suggest. It is more important whether the model can be replaced later and who monitors quality in operation. A good model without operations loses its benefit; a replaceable model with operations can be improved.
