Travel
Agentic AI for travel and hospitality — disruption handling, rebooking and customer care that acts through irregular operations rather than queueing them.
What makes travel hard for AI
Demand is spiky by nature
Disruption produces order-of-magnitude spikes in contact volume precisely when resolution matters most. A system that degrades under load fails at the only moment it was needed.
Rebooking is a constrained optimisation
Availability, fare rules, entitlements and connection feasibility interact. A plausible-sounding answer that violates a fare rule is worse than no answer.
Passenger rights are statutory
EU passenger rights regulations define entitlements precisely. Automated handling has to apply them correctly and evidence that it did.
Where the value concentrates
- Disruption handling. Agents that rebook and resolve autonomously within policy during irregular operations.
- Care and servicing. Routine changes, refunds and entitlement questions handled end to end.
- Ancillary and personalisation. Relevant offers grounded in actual itinerary context.
- Claims and compensation. Correct application of statutory entitlements with an audit trail.
What the economics look like
The measurable outcome is service level held through disruption rather than average handling time on a normal day. Peak capacity is what costs money in travel operations, and it is where automation earns most.
How we build it
We build agents that operate against live inventory and fare rules with hard constraints, so an action is either valid or escalated — never plausible but wrong.
Related work
Starting smaller
Tour operators and smaller carriers typically begin with entitlement questions and routine servicing before granting agents authority to rebook.
Questions we get from travel operators
Can an agent rebook without human approval?
Within defined policy, yes — and disruption is precisely where autonomy pays, because the volume spike is what breaks manual handling. Constrain hard on fare rules, availability and entitlements so an invalid action is impossible rather than merely unlikely.
How do we handle statutory passenger rights?
Encode entitlements as explicit rules rather than leaving them to model judgement, and log the determination. Compensation decisions are audited, and "the model decided" is not a defensible record.
What happens when the system is overwhelmed?
It should degrade predictably: prioritise by disruption severity and connection risk, fall back to human handling on a defined path, and never leave a passenger without an acknowledgement. Behaviour under load is the design requirement in travel, not average-case accuracy.
Where should we start?
Entitlement and status questions. High volume, low risk, and they build the retrieval foundation that rebooking later depends on.