Most agencies asking "where do we start with AI" are asking a question that already has an answer in their own data. They just have not looked at it.
I have 17 years in travel operations and I now build software for the desks I used to run. When an owner asks me where to start, I do not answer from a product roadmap. I answer from the shape of the demand — and on a Gulf wholesale desk, that shape is very specific.
More than 60% of bookings were created outside office hours. The median gap between booking and travel was nine days. Those two facts, from more than 36,100 real bookings, decide the order of everything below.
The order
1. Measure where the hours actually go — before you buy anything.
One quarter of created-at timestamps from your own back office, counted by hour and by day. An afternoon of work, no budget, no vendor. You will find out whether your peak is 20:00 or 23:00 and whether your quiet day is genuinely quiet or merely unstaffed. Every decision after this one depends on it, and buying software before you have done it is buying an answer to a question you have not asked.
2. Automate the acknowledgement.
Not the quote. The acknowledgement. The cheapest useful automation in a travel agency has nothing to do with AI: a rule that ensures no inbound request sits unanswered overnight. It costs almost nothing, it is not clever, and on a market where most requests arrive after hours it protects a large share of your book. Do this before anything expensive.
3. Automate the assembly of the quote.
This is the real one. Producing a complete, priced, branded quotation is assembly work — gathering, formatting, applying the owner's margin, presenting options in a comparable form. It is the single largest consumer of desk time and it is almost entirely mechanical. It is also the thing that decides whether your answer exists while the client is still in the conversation, or the next morning.
4. Automate the follow-up.
On a nine-day median, a quote that gets no reply within a day is usually dead. Chasing is a task with no judgement in it and it is the first thing a busy desk drops.
5. Only then, look at anything that touches the customer directly.
Everything sold as "AI for travel" starts here. It should be last. It is the most visible, the most demoed, and the least connected to why you lost the booking.
The one thing to never automate
The decision about what to sell and at what margin.
Not because a machine could not produce a number. Because the moment that decision leaves the owner, three things follow: the margin drifts, the exceptions stop being visible, and nobody can explain to a client why a price is what it is.
An automation that assembles the quote and hands it to a human for approval makes the desk faster without making it dumber. An automation that decides and sends removes the only person who understands the relationship from the only decision that carries risk.
I built the approval step into my own product before I built the automation around it, and it is a constraint I would not remove. The owner approves every quote. That is not a limitation of the software. It is the point of it.
What "AI in travel" usually means, and why it is the wrong end
The public conversation about AI in travel is almost entirely aimed at the traveller — planners, itinerary generators, chat concierges, personalisation. That territory is crowded, funded, and it does not touch the constraint that actually costs an agency money.
The constraint is not that customers cannot find hotels. It is that the demand arrives when the desk is closed and the answer takes work that nobody has time to do twice.
Start where the hours go. The hours are in the assembly.
Source: the Gulf Outbound Hotel Index, version 1.0 — more than 36,100 B2B wholesale hotel bookings from one Gulf desk. Findings and full tables · Methodology · CC BY 4.0 · 10.5281/zenodo.21796038