What to ask an AI agency before you hire one

Ask what happens when the system is wrong, who owns the code, what it costs to keep running, what it decides without asking, who gets told when it fails, and what happens if the agency disappears. Capability questions get a yes from everyone. These six have a cost behind them, which is why a confident bluffer struggles with them.

We get asked whether we can build a thing far more often than we get asked what happens after we build it.

That is the wrong way round, and it is the reason so many AI projects are quietly abandoned a few months in. The build is the easy part to judge, because you can see it. Everything that determines whether you still have a working system next year is invisible at the point you sign.

These six are about the build and the arrangement around it. The ones about where your data ends up are a separate set, and they are written out here.

Who owns the code when it is finished?

Ask for the answer in three parts: the repository, the documentation, deployment access.

"You own it" is a sentence anyone can say. The test is whether you could hand the whole thing to a different developer next month and have them make a change. That needs the code in a repository you control, documentation written for somebody who has never met the agency, and the ability to deploy without asking permission.

Our own answer, so you can hold us to the same standard: you keep all three. The system runs on infrastructure we manage, against a yearly fee. Stop paying and you keep the code while it stops running, until you host it somewhere else. That is a real limit, and it is better said out loud than discovered in month nine.

What happens to the bill if your volume triples?

It usually rises, because one of the three parts of a running cost moves with your traffic. The build price is the easy number; ask for the other one.

Every AI system has a running cost and it has three parts: the hosting, the maintenance, the model usage, which is the one that moves with your volume. Ours is a single yearly fee covering all three, which means the number does not move when your traffic does, and we carry that risk instead of you. An agency that quotes you a build price and nothing else has either not worked this out or is leaving you to find it in month three.

So ask it as a number rather than a principle. If the running cost triples with your traffic, you want to know that before you sign and not in the quarter it happens. And if the answer is a flat fee with no mention of a usage limit, ask where the limit sits, because there is one.

Can they show you the list of what it does alone?

Ask for the list. If there is no list, there is no answer, there is only a hope.

Every system that acts on your behalf has a boundary: things it does alone and things it escalates. Issuing a refund sits on a different side of that line from drafting one for approval. Cancelling an order sits on a different side from flagging a cancellation request.

OWASP ranks this sixth in its 2025 Top 10 for LLM Applications, as Excessive Agency: systems granted more autonomy, more permissions or more tools than the task needs. It is the failure that costs the most and shows the least, because nothing looks broken until something expensive happens quickly.

A good answer is written down before anything is built. We write ours as two columns, and the conversation that produces it is usually the most useful hour of a project.

What happens when it is wrong, and who finds out?

Not whether it will be wrong. It will. Ask what the system does on that day and whose name is on noticing.

The answers worth hearing are specific. It says it does not know rather than guessing. It writes the case to a log a person reads. It tells somebody by name, through a channel they already check. And somebody can switch it off in minutes without opening a support ticket.

Ask how long from noticing to stopping. If the answer involves messaging a developer and waiting, the real number is hours, and you should know that before the day you need it rather than during.

What if the agency shuts down or moves on?

This is the question people feel rude asking and it is the most useful one on the list.

Small studios close. They get busy, or they lose the one person who understood your build. The honest version of this answer has no reassurance in it, only arrangements: the code sits in your repository, the documentation assumes no knowledge of us, the credentials live in your accounts rather than ours, and nothing in the contract takes the work back.

An agency that answers this well is telling you it has made itself replaceable on purpose. That is a strange thing to sell and the right thing to buy. Ours is written on the build page, in the same words we would want read back to us.

Does any of this fall under the EU AI Act?

Ask early. It changes what gets built, not the paperwork afterwards.

The EU AI Act names certain uses as high risk, employment and worker management among them, and those obligations apply from 2 December 2027 for standalone systems and 2 August 2028 for AI built into regulated products. Both dates moved back from August 2026 under the AI Omnibus, which entered into force on 27 July 2026.

Whether any of it reaches you depends on what the system decides and for whom. An agency that has never considered the question is not disqualified by it, but one that waves it away has told you something about how it handles the parts of a build that are nobody's favourite.

If a system is already live and you cannot answer these six about it, that is what an audit is for, and it is the same exercise whoever built the thing. None of the six is about whether an agency is clever, either. Cleverness is the part you can see in a demo, and it is the part everybody optimises for, because the demo is what gets shown in the meeting where the decision actually gets made.

These are about the year after it.