Three questions before you buy an AI agent

Before I look at any AI agent demo, I ask three questions. They take two minutes and eliminate most of the proposals that reach my desk. Not because the tools are bad, but because many of them were built to impress in a meeting room, not to move the number at the end of the month.

The demo always works. The lead sends a message, the reply goes out in seconds, the text is polite and well written. Everyone in the room agrees it's impressive. But response speed isn't the problem your sales operation has. The problem is what happens after the response.

What decisions does this agent make on its own?

This is the first question, and the most revealing. Ask the vendor to list the decisions the agent makes without asking anyone's permission. Qualifying the lead as hot or cold. Choosing which project to present. Booking the visit in the broker's open slot. Moving the card to the next stage in the CRM.

If the answer is none, you bought a search box with a nice voice. The customer asks, the system looks through the material and returns a paragraph. It's useful, but it's an FAQ with good copywriting. It's not an agent, because an agent is something that acts.

Imagine a developer that gets three hundred leads over a launch weekend. A system that only answers leaves three hundred open conversations for the team to handle on Monday. An agent that decides has already delivered, by Monday, a list of booked visits and another of leads who don't fit that product.

What does it do when it doesn't know?

Every AI will hit a question it can't answer. A customer asking about a payment option that isn't on the price list. Someone who wants to know if the building allows large dogs. A question about construction progress.

What matters is the behavior in that moment. Guessing silently costs more than handing off to a human. An invented answer about price or timing becomes a complaint, a lost lead or, in the worst case, a legal problem. And nobody finds out, because the conversation looked normal.

Ask the vendor how the agent recognizes that it doesn't know, who it hands the conversation to and how quickly. If the answer is vague, the risk is yours, not theirs.

Where does it keep what it learned?

The third question is about memory. During a conversation, the agent picks up several things: the customer wants three bedrooms, wants to be near the subway, has a property to put down as a trade-in, prefers to be called in the morning. Where does that information end up?

If the conversation dies at the end of the interaction, the next contact starts from zero. The customer comes back the following week and has to repeat everything. The broker who takes over knows nothing of what was said. The CRM still has an empty record, just a name and a phone number.

The right agent records what it understood where the operation works, which is usually the CRM. That way the next message, the next call and the next proposal start from what the customer said, not from a generic script.

What changes results

A fast reply impresses in the meeting room. What changes results is the next action being different because of what the AI understood. If the agent answered in three seconds but the operation keeps doing exactly what it did before, you paid for speed, not intelligence.

These three questions separate an agent from an answering tool. Which one have you never asked a vendor?

If you want to understand where an agent would make real decisions in your operation, our diagnostic starts there.