How to choose an AI provider
Choosing an AI project provider: questions to ask, signals to check and what makes a credible quote.
Reading time: 9 min
The choice of who will carry out an AI project is played on criteria other than those that are usually used. Technology matters little; the way the provider deals with your problem matters a great deal.
There is a signal that is worth more than any other and that is recognized in the first meeting: a supplier who tells you when a thing not You better do it.
In summary
- The best signal: you do not recommend A piece of work.
- It matters more than it does questions about the process that experience in the field.
- Price differences are almost always explained by entries included or omitted.
- Ask for a reference on a project old, not on the one just finished.
- Those who talk about technology and not process are answering another question.
- Write down who responds when something breaks.
The Six Questions That Filter
- "What happens when the system goes wrong?" The useful answer is about the process —who notices, where the error ends —not the accuracy of the model.
- "What exceptions won't you handle?" Those who promise to cover everything will discover them during development and remake them for you.
- "Who is to follow him from us?" If the answer is ‘no one, it is turnkey’, either he has not understood or prefers not to say it.
- "What systems do you connect to, exactly?" We need names of our own, not "your systems."
- "What happens after release?" Assistance, time, cost, for how long.
- "Who owns the code and data?" It needs to be clarified first, you get it with difficulty afterwards.
A supplier who responds accurately to all six has already thought about the project. One who responds well to three is improvising, and the other three will become your problems.
What to Look Beyond Answers
Good Signals
- Ask questions before proposing
- He advises you against something.
- Talk about exceptions and mistakes
- He asks who will follow him internally
- It proposes a narrower perimeter than you asked for
Signals to be explored
- List technologies instead of results
- Promises very short time without analysis
- Never Name Post-Release
- It has only one solution for different problems
- Insist on urgency
The last point on the left is the most counterintuitive: a provider who proposes you do less than you asked is almost always working in your interest. The tight perimeter is the right choice and is not the one that maximizes its immediate turnover.
How to compare quotes
Two offers with very different prices rarely have different margins: they have different perimeters.
| Voice | If there is | If Missing |
|---|---|---|
| Process Analysis | Realistic estimate | The supplier works on assumptions |
| Named Integrations | Clear Perimeter | Surprises Come Later |
| Error Management | System that holds | Prototype disguised |
| Testing on real cases | Fewer problems in production | Customers are testers |
| Post-release assistance | Continuity | Discussion at the first failure |
A quote that omits three of these items is no longer cheap: it is incomplete, and the difference is paid later with the variants. The detail of the costs is in how much AI costs in the company.
References, well asked
The standard reference — a customer happy with a project just delivered — says little. What is needed is a project in production for a year or more.
- Does it still work like the first day?
- How many times did you have to call them?
- How often do they respond?
- Has something changed upstream and how did they manage it?
- If you had to do it again, would you call them back?
The fourth question is the most informative: an integrated project always encounters a change in the first year, and how it was handled says everything about the report.
Practical examples
Supplier who narrowed the perimeter. A company that asked for seven use cases proposed two, explaining that the other five had too many exceptions for the first version. Project delivered on time, extended after six months.
Quotation forty percent lower. Comparing the entries, there was a lack of analysis, testing and assistance. The final cost, with the variants, exceeded the highest offer.
Supplier with great industry experience. He proposed the same solution already sold to others, without asking anything about the specific process. The system covered well the industry standard cases and badly the particularities of the company, which were the reason why a tailor-made project was needed.
The first meeting, what to observe
Much is decided in the first conversation, and not so much by what the provider says as by what he asks.
A supplier who after half an hour knows how to describe your process better than you told it he listened. One who after half an hour presented three successful cases of other customers he sold.
- How many questions did he ask about your specific process?
- Did you ask for special cases and exceptions?
- He asked who will use the system and who will follow it?
- Did he mention something that he doesn't think is worth doing?
- Did he talk about what happens when things go wrong?
Three out of five positive responses indicate an interlocutor who is evaluating your problem. Fewer than two indicate that you are receiving a standard proposal with your name on it.
When it is best to involve more suppliers
Asking for three quotes is practice, but for AI projects it only makes sense under certain conditions.
It makes sense if you have prepared a common card with process, numbers and systems: then the offers are comparable and the difference says something. Without that board, three suppliers will propose three different projects for three slightly different problems, and comparison will be impossible.
It doesn't make sense if the project is small: the time you and they spend in the commercial phase can exceed the price difference. For a first project contained, it is often better to choose one that convinces you and use the project itself as a test, instead of making a race.
The Six Questions and What They Reveal
| Question | Good Answer | Response to be deepened |
|---|---|---|
| What happens when he makes a mistake? | Talk about the process | Talk about accuracy |
| What exceptions remain outside? | Concrete list | "We Cover Everything" |
| Who's following him for us? | Ask for a referent | ‘Keys in Hand’ |
| What systems do you connect to? | Names and Versions | "Your Systems" |
| What happens after release? | Time and Cost | Silence |
| Whose code is it? | Immediate response | Resend |
Mistakes to Avoid
- Choosing on price without comparing perimeters.
- Evaluating technologies instead of the questions it asks.
- Not asking which exceptions remain outside scope.
- Accepting "turnkey" without internal referent.
- Asking for references on newly delivered projects.
- Postponing the issue of code ownership.
How to apply it in the company
Prepare the same card for all vendors: the ten-line process, today's numbers, systems involved with names, and exceptions you know. With the one in hand the offers become comparable and the estimates much closer.
Then you ask the six questions to everyone, in the same order, and write down the answers. The difference between the candidates emerges more from the questions they ask them than from the answers they give.
Contractual aspects are in contract with an AI provider, the choice between interior and exterior in DIY AI or agency.
After the choice
The relationship with the supplier is built in the first few weeks, and there are two behaviors that make it productive.
The first is to give quick access to those who know the process. A supplier waiting days for an answer on a detail works on assumptions, and hypotheses become code to be redone.
The second is to say immediately when something does not convince, instead of waiting for delivery. A correction reported mid-work costs little; the same signaled at the end is a variant.
The opposite is also true: if the provider reports that something you had asked for is not worth doing, it is worth listening to it before insisting. It's the behavior you've been looking for in the selection phase, and ignoring it discourages it for the rest of the project.
The contractual aspects to be closed first are in contract with an AI provider.
The right supplier for the right stage
There is no best provider ever: there is the one suitable for where you are.
For a first small project you need someone quick, who does not impose heavy processes and who accepts a tight perimeter. Instead, a system that becomes central to processes needs someone structured, with guaranteed continuity and the ability to intervene quickly.
They are different profiles and rarely coincide. Changing supplier from one stage to another is normal and is not a failure of the relationship — provided that the points of contract with an AI provider have been closed at the beginning, otherwise that passage becomes impossible.
To go deeper with verified price lists and benchmarks, When to Change Your AI Model and Why Vendor Benchmarks Are Not Comparable.
Conclusion
Choosing an AI provider means choosing who understands your process, not who uses the latest technology. The questions it asks you say more about the answers it gives you.
And the most reliable signal remains the initial one: Who advises you against something is looking at your result. It is rare, it is counterintuitive for those who sell, and it is the best predictor of a project that will be fine.
If you want to understand what processes in your company can be automated with AI, Giallo Studio can help you turn the problem into a real workflow.
Finally, it is worth remembering that the choice of supplier is not final: many relationships start with a small project of their own to occur with each other. A first contained job, with a clear perimeter, says more than any selection stage — and if it doesn't work, the cost of the error remains low. It is also why it agrees that the first project is small, as described in AI project: how to get started.
Giallo Studio guides: methodology and sources
How we apply it in the company
Related resources
FAQ
What is the best signal that a supplier is serious?
That they tell you when something is not worth doing. A provider who advises against a piece of work shows that they are focused on your outcome rather than the size of the quote, and that is the most reliable indicator there is.
Does experience in my field count?
Less than you think. It matters much more that he knows how to ask questions about your process. A provider who has worked in your industry but proposes the same solution to everyone is less useful than one who does not know it but listens and understands.
How do I compare two very different quotes?
Looking at what they include: analysis, integrations listed one by one, error management, testing, after-release assistance. Almost always the price difference is explained by entries present in one and absent in the other, not by different margins.
Do I have to choose who uses the most advanced technology?
No. The choice of model affects the result much less than the integrations and exception management. A supplier who talks mostly about technologies and little about your process is answering the wrong question.
Is it important that it is close geographically?
Less than before, but availability matters. What is really needed is to know who answers when something is not working and with what timing. It is a question to be asked explicitly and to be put in writing.
How do I verify references?
Ask to speak to a client who has had a similar project in production for at least a year. Not a demo, not a success story that has just concluded: an older project, because that shows how the follow-up was handled.