The plain definition
An AI agent is a program that takes a goal from a person ("find me a table for four on Friday") and carries out the steps itself: searching, reading, asking questions and, where it is allowed, committing to something. An AI agent booking is a booking made that way, either through a website, through an integration, or by phoning you with a synthetic voice.
How it differs from a customer booking online
- Who fills in the form. Online, the customer does. With an agent, software does, based on what the customer said earlier. It may leave out details the customer would have thought of, such as allergies, or fill in guesses.
- Who reads your terms. A customer might skim your cancellation policy. An agent reads only what is exposed to it in a form it understands, and may not mention the terms to the customer at all.
- Which door it uses. The customer uses your booking page. An agent may use your page, a phone call, an email, or a standard interface such as MCP or A2A, explained in our guide to those terms.
- How you know who it is. Online, you have a name and an email. With an agent there are two parties to identify: the agent and the person it acts for. Some say they are an AI; some do not.
- Volume and timing. Software does not mind ringing five places at once or retrying at 19:00 on a Friday. A single customer rarely does.
What can go wrong
- A booking that ignores your rules (party size, deposit, required details) because the agent never saw them.
- Wrong information in the other direction: the agent quotes a price or time from an outdated listing.
- Several holds at once for the same customer at different businesses, and cancellations afterwards.
- Uncertainty about who agreed to what, if there is a dispute about a deposit or a no-show.
None of this is inevitable, and most of it is the same kind of problem that online booking raised a decade ago. What is new is that the party on the other end is software that may not behave like a person.
What a business can set up
- State your rules in text where both people and programs can read them: deposit, cancellation notice, required details, who may book.
- Choose your doors. Decide whether you accept bookings by phone from automated callers, only through your website, or through a standard interface. Publish that choice.
- Confirm with the person. A confirmation by email or text to the customer, not just to the agent, shows who is really coming.
- Keep a record of who asked, for whom, and under which rules, so a dispute has something to stand on.
- Write bookings into the calendar you already use so a human and an agent cannot take the same slot.
The Model Context Protocol, described on its introduction page, is one open standard for connecting AI applications to outside tools; the A2A protocol, described on its own site, covers how agents talk to other agents. You do not need to adopt either yourself to be reachable, but it helps to know they exist when someone mentions them.