AI that answers fast doesn't sell. AI that knows when to stop does.

Every AI agent demo opens with the same number: response time. Two seconds. Twenty-four hours a day. No lead left unanswered.

It is a true number and an easy one. It is also the least important on the list.

Speed solves a real problem. A lead who waits two hours for a reply goes cold, and in a high-volume operation that costs money every day. But speed is a floor, not a differentiator. As soon as everyone in your market has an agent answering in two seconds, and that is happening faster than most people think, answering fast stops buying you any advantage. It becomes hygiene, like having a website.

What separates an agent that helps from an agent that just takes up space is something else: it knows when to stop talking.

Three places where stopping is the right call

When the customer has already said yes. An agent trained to be helpful keeps offering information after the person has decided to move forward. One more detail, a comparison, a question about another development. Every extra message at that point is a fresh chance for the customer to have second thoughts. Once the yes shows up, the only thing the agent should do is book the conversation with a human.

When the question is about a decision, not information. Construction deadlines, payment terms, trade-ins, what is negotiable. It is not that the agent can't answer. It is that the answer is a commitment, and commitments belong to someone with authority. An agent that answers these on its own creates expectations the company will have to unwind later, with the customer already annoyed.

When it doesn't know. This is the easiest to fix and the most ignored. A language model was trained to produce a plausible answer, not to admit a gap. If you don't force the agent to say it doesn't know, it makes something up. And it makes it up with the same confidence it has when it is right, which is exactly what makes the mistake expensive. In customer service, an invented figure about square footage, timelines or price is not a typo. It is a commercial problem and sometimes a legal one.

The number I would look at first

If I could track only one indicator for an agent in a sales operation, it would not be response time, volume handled or resolution rate.

It would be this: of the leads the agent passed to a human, how many did the human consider good?

That number is uncomfortable because it does not depend on the technology alone. It mixes the quality of the agent, the clarity of the qualification criteria and the trust of the sales team. That is exactly why it is useful. It is the only one that measures the whole thing.

When that number is low, improving the prompt won't help. Either the definition of a sales-ready lead was never agreed, or it was agreed and never written down, or it was written by someone who doesn't talk to customers.

What this changes when you buy

When a vendor opens the presentation with response time, ask three questions before you look at the price.

What does the agent do when it doesn't know the answer?

How does it decide it is time to call a human, and who defined that rule?

What does the human receive along with the lead?

Those three answers tell you more about the result you will get than any latency figure. And a good vendor enjoys answering them, because they set it apart from anyone who just wrapped a model in an interface.

The takeaway

An AI agent in sales is not a robot that talks. It is a filter that decides.

Answering fast is what it does all day. Knowing when to go quiet and hand over to whoever closes is what it does once per lead, and that single decision is the one that shows up in revenue.