Writing
How to Choose an AI Partner Without Getting Burned
The questions are short and the wrong answers are obvious once you know to listen for them.
Published · Updated · 8 min read
The short version
Ask who owns the accounts, what happens if you stop paying, and which named person on your team gets trained. If a provider is vague on any of those three, you are buying a dependency rather than a system.
Most bad AI engagements are visible on the first call. Not because the provider is dishonest, but because the shape of what they sell does not match the shape of what you need, and neither side names it.
1. Who owns the accounts and the code when this ends?
If the system runs on their infrastructure under their billing, you are renting. That may be fine, but price it as a permanent subscription rather than a project.
The answer to listen for: everything is in your accounts, on your billing, from the first week.
2. What happens if we stop paying next month?
If the honest answer is that it stops working, the handover was never real.
3. Who on our team will be trained, by name?
If there is no name, adoption is not in the plan. A system nobody owns is dead within a year, and the build quality is irrelevant to that outcome.
4. How did you decide this was the right thing to build?
If the answer is because you asked for it, they are order-takers. You are usually right about the symptom and often wrong about the cause.
Owners arrive certain the problem is support. Sometimes it is. Often support volume is a symptom of a quoting process that sets the wrong expectations, and fixing support makes the noise quieter while the cost stays.
5. What will it be trained on?
Help centre and public documentation is the weak answer. Your actual replies, quotes, and decisions is the strong one, because the gap between published policy and real practice is exactly where systems fail.
6. What is it not allowed to decide?
A provider who has not thought about escalation boundaries has not thought about failure. Ask what happens when it is unsure, and what the value ceiling is on anything with money attached.
7. How will we know if it stops working?
AI systems fail quietly. They do not stop, they get worse. Ask what is logged, what is reviewed, and who looks at it.
8. What does done look like?
If done is defined as delivered, it will be delivered and ignored. Done should be defined as your team running it without them in the room.
9. When would you tell us not to do this?
The most useful question on the list. A provider who cannot name a situation where their service is the wrong answer will sell you their service regardless of what the situation is.
Our answer, for the record: if you have a technical person who will own it, buy a platform. If the work is a clear rule with no judgement in it, buy a connector. If AI is going into your product, hire.
Two signals that are not signals
Case studies from your exact industry. Reassuring, and weak evidence. The operations that cost money are similar across industries because the shape of the work is similar. Quoting is quoting.
A long tool list. Which platforms someone knows says nothing about whether they will pick the right target, and the target is the whole game.
Questions
Should we ask for a fixed price?
Ask for a fixed scope. Fixed price on a vague scope produces a provider defending the boundary instead of solving the problem.
Is a retainer always bad?
No, but know what it buys. Ongoing build capacity is a legitimate purchase. Ongoing maintenance of a system your team should be able to run is not.
How long should the first project take?
Weeks, not quarters. If the first working thing is a quarter away, the scope is too wide.
Read next
If you are comparing options
Find out what your most expensive operation is
The assessment covers the whole company: how you get customers, how you deliver, and how the place runs. You get it in writing, and you can build from it with or without us.