Don't Hire an AI Consultant Until You Ask These 10 Questions

Question 1

How do you handle prompt injection?

A hacker hides instructions in an email or invoice your AI reads, and the AI obeys the stranger instead of you. One crafted email made Microsoft's Copilot leak files with zero clicks.

A good answer sounds like"It can't be fully prevented, by anyone. So we assume it happens and limit the blast radius: least access, human approval, and here's how that looks in your setup."

Walk when you hear"Our filter blocks it." Or "Never heard of it." Or any promise that it's solved.

Question 2

Where does my agent run, and what can it touch?

An agent loose on your main systems with admin rights is a data-destruction machine with a random seed.

A good answer sounds like"Isolated sandbox, not your production machines. Zero access by default. Internet only where allowlisted. Destructive actions wait for a human yes."

Walk when you hear"It runs on your server and needs admin to work."

Question 3

Whose keys does it hold, and where do they live?

Whoever holds the credentials owns your business.

A good answer sounds like"Your accounts, not ours. Keys scoped per task, stored in a secrets vault, MFA on everything, rotated."

Walk when you hear"We use our master key for all clients." Or "It's in the config file."

Question 4

It breaks my data at 2 a.m. How do I get it back, and how fast?

Insurance mails a check in eight months. A tested restore has you open Monday.

A good answer sounds like"Backups the agent itself can't touch, and we tested a restore on [date]. You'd be back in X hours."

Walk when you hear"The cloud handles that." Or they have never actually tested a restore.

Question 5

Show me the log.

If nobody can say what the agent did, nobody can say what went wrong. And no insurance claim survives without evidence. Then ask the harder half: show me where a human told it no. In May 2026 a federal court found "not a scintilla of evidence" of real human review, and the tell was that nobody had ever rejected the AI. A log of nothing but approvals is not proof of oversight. It is evidence against you.

A good answer sounds like"Here's everything it did last Tuesday: every file, every message, every call. And here's the decision log on top of it: what your people approved, what they changed, and what they sent back. The rejections are what proves the oversight was real."

Walk when you hear"It doesn't really keep logs." Or an approval log that has run for months and never once records a human saying no.

Question 6

If I fire you tomorrow, what happens?

Vendor lock-in with a person attached.

A good answer sounds like"You own the accounts, keys, and code. Here's the off switch you can flip yourself, today."

Walk when you hear"It's all under our accounts, but don't worry."

Question 7

Where does my data actually go?

One shared system means one bug leaks your books to a stranger. It happened to billion-dollar Asana in June 2025. Regulated industry? This is a legal question.

A good answer sounds like"Your own isolated environment. The provider doesn't train on your data. And what industry are you in? That changes the compliance answer."

Walk when you hearYour data shares a system with their other clients, or they answer before asking your industry.

Question 8

How do you version and back up what you build me?

No history means no undo and no paper trail.

A good answer sounds like"Private repository per client, you get access, every change reviewed, rollback tested. Secrets never stored in it."

Walk when you hear"It's all on my laptop."

Question 9

Does your insurance cover an AI agent wrecking my business?

Since January 2026 insurers attach generative-AI exclusions at renewal, and a generic business policy never covered this anyway.

A good answer sounds like"Tech E&O plus cyber, no AI exclusion endorsement. Here's the certificate of insurance."

Walk when you hear"I've got a business policy, we're fine." (Bonus move: ask your own AI whether your policy covers AI-caused losses.)

Question 10

Whose account does my agent run on?

A consultant running every client off one personal login breaks the AI providers' account-sharing rules. Your business can be shut off for someone else's behavior, on an account you don't own.

A good answer sounds like"Your own account, in your name. A subscription or API key that you control. And you can switch AI providers when a better or cheaper model ships."

Walk when you hear"It all runs on my account." Or anything hardcoded to one provider forever.

Don't hire an AI consultant until you ask these 10 questions.

The 10 questions.