Where AI agents actually work, and where they embarrass you

There are two confident positions on AI agents right now and both are wrong. One says they will replace your sales team this year. The other says they are a novelty that annoys customers. In practice an agent is a tool with a narrow competence, and the businesses getting value from them are the ones who worked out where that competence ends.

The three jobs they do genuinely well

Responding instantly, at three in the morning, on a Sunday. Speed to lead is the single most reliable predictor of whether an inbound enquiry converts, and no human team can be first every time. An agent can. This alone justifies most deployments.

Asking the boring qualifying questions. Budget, timeline, location, what they have already tried. These are the questions a salesperson finds tedious and sometimes skips, and they are exactly the questions a model asks consistently and without irritation.

Booking. Once someone is qualified, an agent that can read a live calendar and confirm a slot removes the back-and-forth that loses a meaningful share of otherwise good leads.

The three jobs they do badly

Anything where the customer is upset. A model will match the tone of a complaint and produce something that reads as sincere, and it is not, and people can tell. Route anger to a person immediately, this is the failure mode that costs you a customer permanently.

Anything where being wrong is expensive. Pricing on a bespoke job, contractual terms, technical feasibility, anything regulated. A model will produce a fluent, confident, wrong answer, and your customer will treat it as a commitment you made.

Anything requiring genuine judgement about a person. Whether this client is going to be difficult, whether this project is worth taking, whether someone is a serious buyer or three months from being one. That is pattern recognition built from consequences, and a model has never experienced a consequence.

The rule that makes it work

The agent handles the first four minutes and hands over. Not the first conversation, not the qualification-to-close journey. The first four minutes: acknowledge, ask the qualifying questions, book or route. Everything after that is a person.

Every deployment we have seen fail did so by widening that scope. Someone notices the agent is handling eighty percent of conversations and decides it should handle the rest too, and the twenty percent it inherits are precisely the cases where being wrong is most costly.

Say that it is an agent

There is a persistent belief that disclosure kills conversion. We have not seen it. What kills conversion is a customer working it out for themselves halfway through, which they will, and then retrospectively distrusting everything the agent said, including the parts that were accurate.

A single line at the start costs nothing and buys you the ability to be obviously automated without being deceptive. It also means the handover to a human reads as an escalation rather than as a bait and switch.

What to measure

Not deflection rate. Deflection measures how many humans the agent stopped from being involved, which is only a good number if the outcomes were good, and it will look excellent right up until the moment you check the outcomes.

Measure the same things you measure for a person: median response time, qualification accuracy against what the salesperson concluded, booking rate, and how often a conversation had to be escalated after the agent had already answered. That last one is the number that tells you whether your scope is right.