Agentic booking, and what actually ships in 2026.
The first wave of agentic booking, from late 2024 through early 2026, was a demo wave. A camera pointed at a laptop, a voice asking for a beach trip in September, a chat window populating with flights, hotels, and a car. Impressive on stage. Impossible to explain to a revenue manager. The second wave, the one shipping now, is narrower and quieter and, for the businesses that pick their spot correctly, worth an order of magnitude more.
Why the general agent lost the argument
The general agent, the one that plans and books an end-to-end trip in one conversation, cannot answer three questions the industry demands. Who owns the booking record. Who is liable when the itinerary breaks. Who reprices when a supplier changes inventory. Every one of those answers routes back into a system of record that the general agent does not own and cannot fix. So the general agent stays on stage and the business quietly ships something narrower.
The four places agentic actually ships
First, servicing. A guest with a booking, asking to change a date. The agent has a booking number, a schedule of change fees, a live inventory feed, and a single, bounded outcome. Second, reshopping. A ticket exists, the fare drops, the agent detects the delta and rebooks inside policy. Third, disruption. A flight cancels, the agent finds the next viable option and holds it while the traveller decides. Fourth, corporate booking inside a policy. Not open-ended trip planning. A defined universe of fares, a defined set of preferences, a defined approval flow.
The bar the agent has to clear
The bar is not accuracy on a benchmark. The bar is behaviour under a bad input. When the inventory feed is stale, does the agent hold or does it invent. When the fare changes mid-transaction, does the agent tell the traveller or hide the delta in a footer. When a supplier goes down, does the agent fail loudly or fail silently. The businesses shipping agents in production are the ones who spent the first six months not on the model, but on the failure modes.
The stack, in one paragraph
A retrieval layer that returns the exact booking, fare rule, and inventory state, not a paraphrase. A guardrail layer that refuses to book without a confirmed price, a confirmed availability, and a confirmed policy match. A model layer that plans, reasons, and calls tools, but never writes to the system of record without a signed transaction. An audit layer that reconstructs every decision for a regulator or an ops team. The model is the interesting part on stage. Everything else is what keeps the business running.
Where to start
Pick the servicing lane. It has a bounded outcome, a paying customer with a specific problem, and a measurable saving on contact centre volume. Ship it into a real queue, with a human backstop, for ninety days. Measure containment, escalation, and NPS. If the numbers move, expand the agent's authority. If they do not, you have learned something about your data before you invested in the reshopping or disruption use cases.
We build agentic systems that ship into servicing, reshopping, and disruption for travel operators. If you are past the demo and stuck on the failure modes, that is where we live.