Five Questions to Ask Any Agency Before You Pay for AI

Most owners walk into the meeting with the wrong questions, or none, and the agency walks out with a signed deal it did not earn. I am on the other side of that table often enough to know how it goes. The pitch is smooth, the demo looks like the future, and the owner nods along because asking a hard thing feels rude in a room that warm. So here are the five questions I would least like to be asked, handed to you on purpose. Not thirty. Five, because an owner can hold five in their head and a vendor cannot dodge all five. These five questions are the difference between hiring a partner and renting a reseller, and you can carry them into any pitch without a notepad.

I am giving them away because the agencies that deserve to lose a deal are exactly the ones who cannot answer these straight, and I would rather compete on the questions than on the slide deck. A good answer to all five is hard to fake. A bad answer is hard to hide once you know what you are listening for. The list most people bring is the problem: thirty bullet points printed off the internet, which guarantees the owner asks none of them and signs on vibes. Five is the number you actually use. Five is the number that changes what happens in the room.

What questions should I ask an AI agency before hiring one?

Ask five. Who owns the result, the work and the data and the logic. What happens when it breaks, meaning a support plan, not silence after launch. What exactly is delivered, a concrete thing, not "AI capabilities." How you exit without being trapped in one vendor. And how you will both measure that it worked.

That is the whole checklist in one breath. The rest of this is why each one matters, and what a real answer sounds like next to a polished dodge. Read them in order, because the order is part of the design: ownership first, because it colors everything after it, and the measure of success last, because a vendor who skipped the first four will scramble on the fifth.

Question one: who owns the result?

Once it is built, who owns the thing. The automation, the data it runs on, the prompts and logic that make it work, the account it all lives in. This is first because every other answer bends around it. An agency that owns the result by default owns you by default, and a lot of the worst outcomes start right here, quietly, in a sentence nobody read.

A good answer is fast and unbothered, because a real partner has thought about this before you walked in. "You own all of it. The accounts are in your name, the data is yours, and if we part ways you keep everything we built." A weaker answer gets slippery. It talks about "our platform" and "our proprietary system" and the work living inside their tools, on their logins, in a shape you cannot take with you. That is not always malicious. Sometimes it is just how they are set up. For you it lands the same way: you paid to build something you do not own, and the meter never stops.

The ownership question runs straight into where your data actually goes and whether you can ever get it back, which is its own subject worth a hard look; the breakdown of what happens to your business data once an AI vendor is holding it gives you the plainer follow-ups to ask once "who owns it" opens that door.

Question two: what happens when it breaks?

It will break. Not might, will. An AI tool that touches real work runs into a changed website, a new pricing structure, an edge case nobody pictured, an underlying model that updates and behaves a little differently than it did at launch. The question is not whether it breaks. It is who picks up the phone when it does, and whether that was priced in or is about to become a surprise.

The dodge here is a kind of silence. The pitch is all about the build, the launch, the big reveal, and the part after launch is a fog. A real answer is specific about the boring part. "Here is what support looks like after we hand it over. Here is the response time. Here is what is covered and what is billed separately, and here is who you contact." You are not asking for a lifetime warranty. You are asking whether they planned for the day it stops working, or whether they planned to be gone by then. A vendor who only loves the launch is telling you when they intend to disappear.

Question three: what exactly do I get?

Make them name the deliverable. Not "AI capabilities," not "an intelligent solution," not "a transformation." A thing. A working automation that does this specific task. A chatbot that answers these specific questions and hands off to a person at this specific point. A workflow that takes this input and produces that output. If you cannot repeat back in one sentence what you are getting for your money, you are not buying a deliverable. You are buying a vibe, and a vibe cannot be tested, fixed, or owned.

This is the question that exposes the resellers fastest, because a reseller is selling access to AI in the abstract and a partner is selling a result in the concrete. AI platforms like ChatGPT, Claude, and Gemini are genuinely capable, and that is exactly why "we will add AI" is such an easy and empty thing to promise; the capability is real, so the hand-wave sounds credible. Push past it. The good answer is almost dull in its specificity, a plain description of the thing that will exist at the end that did not exist before. The weak answer keeps the deliverable cloudy on purpose, because clouds are hard to hold anyone to.

Question four: how do I leave you?

Ask it plainly, early, before you sign. If this does not work out, or I simply want to take it elsewhere in a year, what does leaving cost me, and what do I walk away with. The answer to this question, asked while they still want your business, tells you more than almost anything else, because it is the one moment their interests and yours are openly not aligned, and you get to watch how they handle that.

A partner has a clean answer because they built for it. "You keep everything, it is all in your accounts, and there is no exit fee because there is nothing to hold hostage." A setup designed to trap you cannot say that without flinching. The work lives on their platform, the data is entangled with their tools, leaving means rebuilding from nothing, and somehow that only becomes clear the day you try to go. The honest version of this is what a clean engagement looks like; the trap is the thing you are screening for. If the answer to "how do I leave" is fuzzy, treat the fuzziness itself as the finding, because the day you want out is exactly the day you will wish you had asked on the way in.

Question five: how will we know it worked?

Last, and most skipped. Before any work starts, how will the two of us know, concretely, whether this did what it was supposed to. Not "you will feel more efficient." A number, or something close to one. Hours saved per week. Response time cut from this to that. Quotes out the door faster. Tickets handled without a person. The point is to agree on the measure before the build, while it is still a shared goal, not after, when it has quietly become a debate about whether you got your money's worth.

A vendor who answers this well is rare and worth keeping, because naming the measure up front is a small act of courage; they are handing you the stick you will measure them with. "Here is the one outcome that defines success. Here is roughly where you are now. Here is how we will check it together at ninety days, and here is what we do if it falls short." A vendor who waves it off ("you will see the difference," "it will pay for itself") is keeping the goalposts loose on purpose, so they can be moved later. Vague success is unfalsifiable, and unfalsifiable is just another word for unaccountable.

What the room tells you

Here is the part that matters more than any single answer. The point of the five is not to grade each reply against a perfect one. It is to watch the pattern. Does the agency meet a plain question with a plain answer, or does every hard thing get redirected to a slide about another client. A real partner can take all five straight, in order, without reaching for the deck, because they have lived these questions from the inside and the answers are just true. A reseller pattern-matches each question to the nearest case study and hopes the warmth carries the room.

You will feel it within the first two questions. Straight answers have a texture: specific, a little boring, occasionally an honest "that part is on you, not us." Dodges have a texture too, and once you have heard it you cannot unhear it. The answers get general right when you need them specific. Things that should be in writing stay spoken. The conversation keeps drifting back to how impressive the product is and away from what you actually asked. None of that proves bad intent. Often it just means they have not thought as hard about your result as you need them to, which for your purposes is close enough to the same risk.

One honest thing. These five are a screen, not a verdict, and a good owner should be able to ask them and correctly decide to walk away from everyone in the room, including a partner who answers well but is simply not the right fit or the right time. Whether to hire an outside team at all is an upstream decision that sits before these questions, not inside them; the rundown of how to weigh doing the AI work in-house against bringing in an agency is where that fork lives, and it does not pretend there is one right answer for everyone. The five questions assume you have decided to take pitches. They are how you tell the pitches apart.

There is a companion to this checklist worth naming, because it does a different job. This piece is the brand checklist: the five questions, in order, that you carry in and ask out loud. The other half of the work is reading the answers closely, decoding the specific phrases that should reassure you and the specific ones that should worry you, the longer red-flag vocabulary you tune your ear to before you sign. That is its own subject, a buyer's field guide to the answers rather than the questions, and it goes deeper into spotting trouble than five questions can. Treat this post as the five you bring and that one as the way to listen to what comes back.

So the action is small and the same as it has always been. Take these five into the next pitch meeting and ask them in order, plainly, and watch the room more than the answers. If you want a partner who treats all five as the floor rather than the obstacle, an AI upgrade run by people who can answer all five without reaching for a case study is the kind of engagement these questions are built to find. Ask the five. The room will tell you the rest.