Choosing an AI vendor
Before you compare providers, decide what you are commissioning. This page is written from the buyer's side: what should be settled before the first call, which types of provider exist, which contract model fits when, and how to spot a weak proposal.
- AI vendor
- An AI vendor is a company that designs, builds and runs AI systems for clients, from clarifying the requirements to daily operation. The four provider types (consultancy, single freelancer, platform provider, AI agency with operations) differ less in technology than in what happens after the contract is signed: who owns the code, who runs the system and who is liable when hit quality drops.
Four types of AI vendor, and what happens after signing
None of them is right in general. The type decides what happens to you after the pilot.
A client from the Freising district showed us three quotes for the same chatbot in 2026: EUR 9,000, 27,000 and 31,000. The most expensive one included two years of operation, the cheapest ended with the presentation. The price said nothing, the type said everything.
Five things to settle before the first call
With these five in place, every provider type gives you a comparable quote. Without them you are comparing five different projects.
Ten questions for the first call with an AI vendor
Ask all ten of every provider. An evasive answer is an answer.
How many systems are running at clients today?
Pilots do not count. Ask to see a running system, with a client who confirms it on request. Our three case studies are on the case studies page with numbers.
What do we receive at handover?
Source code, configuration, operations manual, credentials. Ask whether the system keeps running if the vendor is unreachable tomorrow.
How do you measure whether the system works correctly?
Who measures, how often, against what standard. With us, 100 test scenarios in 8 categories run regularly against the production system, not only at acceptance.
What happens when the model makes a mistake?
Every language model makes mistakes. The question is whether the system recognises uncertainty, hands over to a person and logs the case.
Fixed price or time and materials?
Both can fit. What matters is that the proposal says what happens with change requests and from which point a change costs money.
Where do the data and the models run?
EU data centre, your own server, a provider in the US. Anyone who takes the GDPR seriously can say this for every component, including the data processing agreement.
Who writes the EU AI Act documentation?
Article 50 has applied since 2 August 2026. Ask whether the system register, risk classification and technical description are part of the handover or cost extra.
Who is my contact after acceptance?
Name and phone number. If the answer is "our support", ask for the response time in hours and the monthly price of operation.
What do you advise us against?
A vendor who wants to build every idea is selling. One who strikes two out of five ideas and explains why is advising. We say in the first call when a project will not carry.
What does operation cost in the second year?
Hosting, model usage, support. If the vendor cannot name that number, they either do not know their own costs or do not want to say.
Three contract models compared
The contract model distributes the risk. Do not ask which one is best, ask which one fits what you know about the project so far.
| Contract model | When it fits | Risk for the client | Risk for the vendor | What to check |
|---|---|---|---|---|
| Fixed price | Requirement is clear, data has been reviewed, scope is set | Change requests cost extra, because the price rests on the scope | Underestimated effort is on them | Change procedure in the contract, acceptance criteria as numbers |
| Capped time and materials | Requirement still vague, data unknown, pilot | Pays up to the cap even if the result stays thin | Cannot claim anything above the cap | Weekly report, cap as a hard limit, right to stop |
| Retainer for operation | System is live and needs care, model changes and a contact person | Pays monthly, also in quiet months | Has to keep response times, even after changes by others | Response time in hours, service catalogue, notice period |
With us the concept is a fixed price, because we have reviewed the data first. Operation is a monthly fee with a service catalogue. The contract is under German law, the invoice is in euros, and the quote has a number in the client portal.
Six warning signs in a proposal
None of them rules a vendor out on its own. Two together are a reason to ask again.
A delivery date before the requirement
A date before anyone has seen the data is a guess. Ask what the estimate rests on.
References without numbers
"Successfully introduced at a mid-sized company" says nothing. A result has a number, a period and a client who confirms it.
Not a word about operation
A proposal that ends at acceptance shifts the most expensive year onto you. Ask for page two.
Everything is possible
A vendor who contradicts none of your ideas has either not listened or is selling hours.
Data protection as a footnote
If the data processing agreement only appears on request and the location of the models stays unclear, it gets expensive later.
The code stays with the vendor
A licence instead of a handover ties you to the vendor. That can be fine with platforms, but it has to be visible in the price.
AI agency, consultancy or freelancer: five criteria
An open comparison with us in the last column. Where another type fits better, it says so.
| Criterion | Large consultancy | Single freelancer | Kuroko Labs |
|---|---|---|---|
| Contract model | Framework agreement, day rates, preliminary study | Hourly rate, rarely a fixed price | Fixed price in the concept, operation monthly |
| Ownership of the code | Usually yes, often with a licence on their frameworks | Negotiable, often unregulated | Code and documentation pass to you on acceptance |
| Operation after acceptance | Own operations unit, high minimum fee | As long as the person is available | Monthly operation with a service catalogue, same team as in the build |
| References with numbers | Big logos, numbers mostly under wraps | Rarely published | Three case studies with numbers, among them 94 percent less effort at our first client |
| Contact person | Changes with the project phase | One person, no cover | One person, with cover, on-site meetings in the Munich area |
Choosing an AI vendor: common questions
Related topics
AI agency in Munich
Production grade AI agents instead of pilots, with a fixed price and on-site meetings.
AI consulting in Munich
Assessment and roadmap before anything is built. Free, even when we advise against it.
EU AI Act documentation
Risk classes, system register and technical description for operators.
Bring the five decisions. We will tell you whether we belong on your shortlist.
If another provider type fits better, you hear that in the first call. If we fit, you get the first assessment in writing.