What Should an AI Phone Agent Be Allowed to Promise?

There is a product category now for AI that answers your business phone, and the demos are genuinely good. Before you shortlist vendors, it is worth separating what the technology solved from what it did not, because only one of those is on their roadmap.

Can AI actually answer your business phone now?

Yes, and convincingly. The capability arrived in October 2024, when OpenAI released its Realtime API at DevDay, enabling speech-to-speech conversation at roughly 300 to 500 milliseconds of latency. Before that, voice agents chained speech-to-text, then a language model, then text-to-speech, and the delay made every call feel like a bad hold system. That chain is gone. The current wave of AI receptionist products exists because of it.

Why is answering the phone the easy part?

Because latency was a technology problem, and technology problems get solved by vendors. What the voice is allowed to commit you to was never a technology problem, and nothing in the last two years has moved it.

Watch how the pitch has shifted. It started at "it can answer." It is now closer to "it can handle it." Those are different claims, and the second one quietly assumes you have already done work that most businesses have not.

What does an AI phone agent need to know before it picks up?

Three things, and they are the same three in almost every service business:

• What you charge, including the jobs you price by looking at them rather than from a list.
• Which appointment slots are genuinely open, as opposed to open because nobody updated the calendar.
• Which work you turn down, and how you turn it down without losing the referral.

In most shops none of that exists in writing. It lives with whoever has answered the phone for the last nine years, and it has never needed to be written because that person has never been unavailable at the same time as a hard question.

What goes wrong when those rules are not written down?

The agent does not fall silent, which is what people expect. It fills the gap. It quotes a number that sounds reasonable, books a slot that looks free, and accepts a job you have no crew for, at two in the morning, very politely. You find out days later, from a customer who has a confirmation and an expectation.

A human at the front desk hedges when they are unsure. Hedging is a learned social behavior, not a default, and the software does not have it.

How do you decide what the AI is allowed to commit to?

Write the three commitments down before you evaluate a single vendor. For pricing, that means the range you will quote sight-unseen and the trigger that forces "we will have to come look at it." For scheduling, it means which calendar is authoritative and who maintains it. For scope, it means the two or three job types you decline and the sentence you decline them with.

That document is worth having whether or not you ever buy anything. It is also the specification any vendor will ask you for in week two, when the pilot starts producing quotes you did not sanction.

Should a small business wait on AI phone agents?

Not necessarily, but the sequence matters more than the timing. A business that has written its three commitments can deploy an AI phone agent in an afternoon and it will behave. A business that has not will get a confident, articulate, always-available version of its worst improvisation, running at volume.

What is the first step?

Take the last twenty calls that turned into work and read what was promised on each one. Where the promise came from a written rule, you are ready. Where it came from judgment, write the judgment down. That list is the actual project, and it is the same list whether the phone gets answered by software or by the next person you hire.