Three questions to ask before an AI agent touches real work

In the demo, an agent glides across the screen, books the meeting, drafts the email and lands every step in nine clean minutes. What you don’t see is the run before it, where the same agent stalled, guessed wrong or sent the thing to the wrong person. I’d treat every demo as a performance with the failures cut, and that gap is what the right AI demo questions are for. The agent spectacle everyone was watching this month turned out to have people behind much of the work.

Canadian buyers are seeing more of these pitches. Among businesses using AI, virtual agents or chatbots were the third most common application in Statistics Canada’s spring 2026 survey, at 28.2%.

The tools are real, and they are getting better fast. The gap is not between real and fake. The gap is between what a staged run proves and what a working week asks. A demo proves an agent can do a job once, on clean data, in front of a patient audience. Your firm needs it to do that job on your worst Tuesday, with your messy files, while nobody is watching it.

Why does every agent demo look flawless?

Because somebody edited it. A demo runs on a task picked to flatter the tool, on data cleaned the night before, with every failed attempt left on the floor. When the agent pauses, guesses, or takes a wrong turn, the recording skips it.

You are shown the one run in ten that behaved.

This month a staged network of agents, set up to look like agents working on their own, drew a wide audience. Within days researchers questioned who was really posting, and it came out that people stood behind much of the work the agents seemed to do alone. The show still proved something real. An agent can look capable on a stage, in front of a crowd, on a task built for it. It proved nothing about the same agent loose in a real office.

A demo shows you the best run. Your firm has to live with the average one.

What did the agent touch while it worked?

This is the first question, and it comes before price, features, or speed. Ask what the agent opened, changed, sent, or spent while the demo ran. A demo agent that reads a sample inbox and drafts replies it never sends has touched almost nothing. Ask too whether it ever notices anything on its own, the test I’d use to tell a proactive agent from a fast chatbot.

An agent that books, buys, deletes, or writes to clients touches your firm’s name every time it acts.

The touch list is the risk list. Every system the agent can reach is a system it can reach wrongly, on a bad day, with a real client on the other end. If the seller cannot tell you plainly what the agent touched during the demo, the demo was staged to hide the reach, not to show the work.

An agent that touches nothing real has proved nothing real.

What went wrong that the edit hid?

Second question. Every honest run has a failure in it somewhere. Ask what the agent got wrong in testing, how often it happens, and what the failure looks like when it lands. A stalled task is cheap. A confident wrong answer sent to a client is not.

The size of the failure matters more than the count of the successes.

Ask to see a failed run, or a record of one. A seller who can show you a failure and explain it plainly is showing you the part of the product you will actually live with. And read whatever the agent produces the way your client will read it, because the words an agent sends in your name may already be labelled as machine made, and your client may know it before you do.

Ask to see the run that went wrong before you ask what it costs.

Who could have stopped it mid-run?

Third question. While the demo ran, could a person pause it, override it, or kill it halfway through a bad action. If the answer is a dashboard nobody watches, the honest answer is nobody. An agent you cannot stop is not a helper on a long leash.

It is an unsupervised worker acting in your name. Before any agent runs for you, I’d put its job and its stop rule in writing.

Decide your stop rules before the first real run, while you are calm, and write them down. I make the same point about pricing tools, where owners do better to write your rules first than to improvise them after the first surprise. Who watches the agent, what needs a human yes before it sends, and what shuts it down. Three lines, written once, save the argument later.

Keep this beside you at the next demo.

Question	A good answer	A warning sign
What did it touch?	A plain list of systems, reads and sends	Vague talk about integrations
What went wrong that the edit hid?	A failed run, shown and explained	A claim that it rarely fails
Who could stop it mid-run?	A named person with a pause or kill switch	A dashboard nobody watches

Which agent should face the three questions first?

Bring the three questions to the next demo you are shown, and ask them in order. What did it touch. What did it get wrong that the edit hid. Who could have stopped it. A good tool survives all three without the seller getting vague. I’d walk away from any seller who can’t answer the first.

A weak one gets vague at question one.

Buy the agent that survives your worst Tuesday, not the one that trended this month.

Then run the same questions on any agent already inside your firm, because plenty arrived without a demo at all. The answers tell you what the agent is allowed to reach, what its failures look like, and if a person can step in. Working through those answers for a specific firm is part of the AI Strategy Roadmap.

What people ask about AI demo questions

Which demo are you being shown next? Bring the three questions to it. I write about stop rules like that each Friday in the AI Newsletter with ZAK.