“Autonomous” has become the most overloaded word in travel technology. It is currently applied to chatbots that answer FAQs, online booking tools with a search filter, and systems that genuinely complete transactions end to end without human involvement.
These are not the same thing, and the difference determines whether a TMC gets a productivity gain or an expensive interface change.
Here is a practical way to tell them apart.
The five levels of booking autonomy
Borrowing the framing used for self-driving vehicles, because it maps unusually well:
Level 0 — Manual
An agent performs every step: receives the request, searches, applies policy from knowledge or a document, selects, books. The system is a set of screens.
Level 1 — Assisted
The system provides shortcuts. Saved traveller profiles, quick-search templates, a policy reference panel. The agent still makes every decision. Most legacy TMC desktops sit here.
Level 2 — Augmented
The system offers recommendations. It surfaces ranked options, flags likely policy issues, pre-fills traveller data. The agent evaluates and confirms every transaction. Most tools marketed as “AI booking” today are Level 2.
Level 3 — Conditionally autonomous
The system completes routine bookings end to end within defined parameters, and escalates anything outside them. A straightforward domestic return for a known traveller within policy is booked without a human. A complex multi-city itinerary with a policy exception is routed to an agent.
This is where meaningful capacity change begins, because it is the first level where transaction volume decouples from agent headcount.
Level 4 — Highly autonomous
The system handles the great majority of transactions including changes and disruption rebooking, learns from every resolution, and involves humans only on genuine novelty or relationship-sensitive cases.
Level 5 — Fully autonomous
No human involvement in any circumstance. This does not exist in corporate travel and arguably should not — some decisions warrant human judgement regardless of technical capability.
The useful question for any vendor is simply: which level, and on what proportion of transactions? A vendor claiming “autonomous booking” who cannot answer that is describing Level 2.
What Level 3 requires that Level 2 does not
The jump from recommendation to execution is larger than it appears. Four things must be in place.
Encoded policy, not referenced policy. At Level 2 the system can display the policy and let the agent judge. At Level 3 it must decide, which means the policy has to exist as unambiguous logic — including what happens when no compliant option is available.
A defined confidence boundary. The system needs to know what it does not know. Booking autonomously requires a reliable measure of when a transaction falls inside its competence, and honest escalation when it does not. A system that is confidently wrong is worse than one that asks.
Write access to real inventory. Recommending is read-only. Booking means holding, ticketing, issuing — and handling failure states when a fare disappears mid-transaction. This is where a lot of “autonomous” products quietly stop.
Reversibility. Autonomous action requires the ability to detect and correct its own errors, including voiding within the void window.
What it looks like operationally
A request arrives — email, chat, or through the client’s own channel. The system parses intent: traveller, origin, destination, dates, purpose, any stated preference.
It resolves the traveller against stored profile and history. It searches connected fare sources, evaluates each option against the client’s encoded policy, and ranks by a blend of cost, policy fit, and the traveller’s demonstrated preferences.
If the best option is within policy and confidence is high, it books. The traveller receives an itinerary. No agent was involved.
If the fare is out of policy, or the itinerary is unusual, or the traveller is new, it produces a ranked recommendation and routes it to an agent with the reasoning attached.
The agent’s role shifts from assembling every booking to handling the cases that genuinely need judgement. TMCs running the VantisCorp Booking Agent see roughly 60% faster booking turnaround and handle around 40% more bookings with the same team — the capacity gain comes from removing the routine, not from replacing the agent.
Common objections, addressed honestly
“Our travellers have complex requirements.” Some do. Most do not, most of the time. Measure the split before assuming. In most portfolios a substantial majority of transactions are routine repeat patterns — the complex ones are memorable, which distorts perception of their frequency.
“An error would be unacceptable.” Correct, which is why the confidence boundary matters more than the capability. A well-configured Level 3 system books only what it is certain about. The relevant comparison is also not against perfection but against current human error rates under time pressure at 5pm on a Friday.
“Our clients want to speak to a person.” They want their problem solved. Many will happily accept an instant correct itinerary. The ones who genuinely want a relationship should get agent time — which is easier to give them when agents are not processing routine bookings.
“It will not understand our exceptions.” Initially, some it will not. Systems that learn from agent resolutions improve; static rules engines do not. Ask which kind you are buying.
How to start without betting the business
Run it in shadow mode first. Let the system process requests in parallel with your agents without executing anything. Compare its choices to theirs over a few hundred transactions.
This does three things: it quantifies accuracy on your real transaction mix rather than a demo; it surfaces policy-encoding errors safely; and it builds agent trust, which is usually the actual constraint. Agents who have watched a system make good decisions for a month are far more willing to let it act.
Then enable autonomous execution on your narrowest, cleanest segment — typically domestic, single-leg, known traveller, within policy — and widen from there as measured accuracy justifies it.
The honest summary
Autonomous booking AI is real, but it exists on a spectrum, and most of what is sold under the label sits at Level 2. The commercially meaningful threshold is Level 3, where routine transactions complete without human involvement and volume stops being limited by headcount.
When evaluating, ignore the adjective and ask the operational question: what proportion of our transactions would this complete end to end, without a human, on day one — and what proportion after six months?
See it in practice: the Booking Agent performs AI-ranked search across every connected fare source and books within policy autonomously, escalating what needs judgement. For the wider context, read our guide to agentic AI in corporate travel.
Talk to VantisCorp about your TMC
See how the workflows in this article could work for your operation. Share a few details and our team will follow up.