What to ask before you hire anyone to build AI agents

Most of this market sells the same four things: agents, gates, file ownership, flat pricing. The differences that matter are in the answers to questions people rarely think to ask until the build is half done.

I have put my own answers under each one. Not because mine is the only good answer, but because a guide that dodges its own questions is an ad.

Who checks the output before it reaches me?

A single assistant hands you a confident answer and leaves the checking to you, which is the work you were trying to remove. Ask who reviews, not whether it is accurate.

What a good answer sounds likeA named reviewing step inside the system, described without you in it. If the answer is "the model is very accurate", that is not a review step.

MineEvery department has a team lead that reads its agents' output before anything moves up.

What stops an action that cannot be undone?

Deleting records, cancelling a service, granting access, sending to a customer. The failure mode that hurts isn't a wrong sentence. It's an action nobody can take back.

What a good answer sounds likeA specific mechanism, and a list of what triggers it. "It asks for confirmation" is weaker than "here are the three triggers and here is where you approve".

MineThree triggers: spends money, reaches a customer, cannot be undone. The work stops and waits for your yes.

What does it cost to run, on top of what I pay you?

Model usage is billed by the provider. If nobody mentions it, you'll find out on a card statement.

What a good answer sounds likeA clear statement of who bills it and whether the vendor marks it up.

MineBilled by Anthropic or OpenAI directly, on your account, at their price. I add no markup and never handle it.

Do I own what you build, or am I renting it?

This decides what happens the day you stop paying. It's the single question with the largest gap between vendors.

What a good answer sounds likeFiles on accounts you control, and a plain answer about what switches off if you leave.

MineSetup files sit on your own Claude Code, ChatGPT or Codex account. I host nothing. Stop working with me and nothing switches off.

What is the price, and when will it be finished?

A vendor who won't price before a call has priced you, not the work. A date is the harder half: most will give a number and no date.

What a good answer sounds likeBoth, in writing, before you commit. Watch for a price that arrives without a delivery window.

MineEvery price is published with its delivery window. A free thirty-minute consult, then builds from $295 to $1,650, delivered in 5 to 10 days depending on size.

What is not included?

Scope arguments happen at the end of a build, when leverage is worst. The vendors worth hiring name the exclusions first.

What a good answer sounds likeA written list, and a statement of what happens if the work turns out bigger than quoted.

MineEvery build states a cap in departments and agents. If intake shows the work exceeds it, I tell you before starting.

What happens when the model changes?

Providers ship new versions. Behavior shifts. Something that worked in March can drift by June without anyone touching it.

What a good answer sounds likeAn honest description of what drifts and who is responsible for noticing. Beware anyone who says nothing will change.

MineI do not sell a retainer, so you are not paying monthly for the option. The files are yours and documented; when something stops working you can come back to me and I help.

What does it look like when it goes wrong?

Everything is demonstrated working. The useful question is what the failure looks like and who picks it up.

What a good answer sounds likeA written failure plan handed over as a deliverable, not a promise to be responsive.

MineA failure plan is one of the four artifacts in every engagement. It says what happens, and who picks it back up.

Can my team run this without you?

If the answer is no, you haven't bought a system. You've bought a dependency.

What a good answer sounds likeA handover document written for the person who runs it, not for a developer. Ask to see the format.

MineA plain-language guide, written for the operator. It is the fourth artifact, delivered every time.

Can I see it working before I commit?

Slides and screenshots are cheap. A thing you can click isn't.

What a good answer sounds likeSomething you can operate yourself. A recorded demo is better than nothing and worse than access.

MineBefore you pay, the free consult maps your process and where the approval gates go, and /work shows three organizations drawn the way they are built. There is no live demo account yet.

Print this, take it to three vendors, and compare. If my answers stop being the better ones, hire the other vendor: my prices are on one page