The gap is already closing
So, what's an AI agent then?
Mark M
21 January 2026
Agent 2 Agent Booking
This small moment, the gap between the recommendation and the booking, is where a lot of the interesting questions about AI agents actually live.
Last week, someone asked ChatGPT to find them a hotel in the Cotswolds for their parents' anniversary. Something quiet, with good food, a proper garden. The AI came back with a recommendation. Not a list. A recommendation. One place, with an explanation of why it fit.
Then the person clicked a link, left the conversation, went to a booking site, and made the reservation themselves.
This small moment, the gap between the recommendation and the booking, is where a lot of the interesting questions about AI agents actually live.
The word "agent" is doing a lot of work
Most of what people currently call AI is a chat interface. You ask, it answers. You ask for options, it offers some. Helpful, certainly, but you're still doing the deciding and the doing. The AI responds; you act.
An agent, in the fuller sense of the word, acts on your behalf. Not just recommends. Completes. The conversation that ends with "here's what I found" becomes "I've booked it, confirmation's in your email."
The infrastructure for this exists. Protocols for agent-to-agent communication, structured data exchange, verification and trust, completing transactions. These aren't theoretical. They're built and working.
To choose, you have to be sure
This is the key difference between responding and acting.
A chat interface can hedge. It can show you five options and say "these all look promising." It can present information and leave the judgment to you.
An agent that's going to act can't hedge in the same way. If it's going to book something on your behalf, it has to choose. And choosing requires a kind of confidence that depends on the quality of information available.
Confidence that it understood what you actually meant. "Quiet" and "good food" and "proper garden" are interpretive. An agent needs to be reasonably sure it's interpreted them correctly before it acts on them.
Assurance that the place it's choosing genuinely fits. Not just that it matches keywords, but that someone celebrating an anniversary would actually be glad they ended up there.
Certainty that it can complete the action. That the room is available, that the booking will go through, that nothing will go wrong between recommendation and confirmation.
This is why agents are selective in ways that search engines never were. A search engine shows you everything that matches. An agent only offers what it's willing to stand behind.
Not everything can be chosen
For an agent to recommend somewhere with confidence, it needs structured ways to understand what that place is actually like. Not marketing copy, not a list of amenities, but genuine signal about character, about fit, about what kind of person would be happy there and who might be better served elsewhere. In practice, this means a machine-readable identity that captures not just facts, but character, constraints, and booking reality.
It needs ways to verify that claims are accurate. Not just that a venue says it's dog-friendly, but that this can be corroborated by a tourism board, by observable patterns, by multiple sources agreeing.
And it needs ways to complete the transaction. Availability that's current, booking terms that are clear, a path from "yes, this one" to "done."
The venues that have this in place get recommended. The venues that don't may still exist on the web, but they're functionally invisible to the agents people are increasingly using to find places.
Seen this way, the changes underway aren't abrupt. They follow a fairly clear path from where things are to where they're heading.
First: AI as search. The AI finds options and presents a list. You compare, you decide, you book elsewhere. This is what most people experienced until recently.
Then: AI as assistant. The AI makes a recommendation and you confirm it. "I found this one. Want me to book it?" You say yes, it completes the action. This is happening now.
Eventually: AI as agent. The AI decides and executes within parameters you've set. You get a confirmation. "Booked. Details in your Wallet." You weren't involved in the decision at all. You'd already told it what kind of thing you wanted, and it handled the rest.
The infrastructure required for the second and third stages is essentially the same. Structured identity, so the agent knows what a venue actually is. Trust signals, so it can verify claims. Booking capability, so it can complete a transaction. Build for assisted, and you're ready for autonomous.
This is already happening
What's clear, without overstating it, is that a shift is underway in how recommendations get made.
When you search for "quiet hotel Cotswolds good food," you get a list ranked by a mix of signals: keywords, backlinks, ad spend, reviews. You do the work of comparing, deciding, and booking.
When you ask an AI the same question in natural language, something different happens. The AI interprets your request, applies some judgment about what would fit, and offers a smaller set, often just one or two options, with reasoning attached. You still do the booking in most cases, but the shape of the interaction has changed. The AI made a choice. You're confirming or rejecting it, not selecting from an undifferentiated list.
This is already observable. Millions of people are using AI this way, right now. ChatGPT apps launched. Bookings are happening.
Being seen isn't the same as being chosen
For venues, there's a practical question buried in all this: what does it take to be the recommendation an agent makes?
It's not the same as what it took to rank well in search. Search rewarded visibility: keywords, optimisation, ad spend. An agent making a recommendation needs something different. Enough structured information to understand what a place actually is. Enough verifiable evidence to trust that the information is accurate. Enough connectivity to complete a booking if that's what the user wants.
A venue that's visible on the web but not legible to an agent, not structured, not verifiable, not bookable in the systems agents use, may find itself steadily less discoverable. Not because it's been removed from anything, but because the way people find places is shifting.
The person looking for a hotel in the Cotswolds got a recommendation. They still had to book it themselves, clicking through, leaving the conversation, completing the reservation elsewhere. But the infrastructure that closes that gap, that turns the recommendation into the transaction in a single conversation, already exists. The question is which venues are ready for it.
Mark M