How We Built AI Agents for a Travel Agency, From First Enquiry to Priced Offer

One of our clients is a travel agency that designs tailor-made trips. Their travel designers are great at their job. The trouble is how little of their day actually goes into designing trips. The rest goes into chasing missing details, retyping supplier quotes and pricing by hand. One proposal took about eight hours. We built four AI agents that take that work off their desk. Here is how they work, and the choices that mattered most.

The problem

Every trip starts with an enquiry. An email, a phone call, a web form. Most of them are vague: “Two weeks in Japan next spring, something special.” Getting to a usable brief meant several emails over several days.

Once the brief is clear, the designer sketches a trip and asks a local partner on site for a quote. In the trade, that partner is called a DMC. The quote comes back in any shape. An Excel file with five tabs. A PDF. A scan. Sometimes just the body of an email. Someone has to read it, pull out every hotel, room, transfer and date, and turn it into a proposal the traveller wants to read.

Then comes the price. The margin rules lived partly in Excel, partly in old tools and partly in people’s heads. Two designers could price the same trip differently. Errors in the figures were found late, or not at all.

Four agents, one chain

We did not build four separate tools. We built one chain, where each agent picks up what the previous one left.

1. The intake agent. It reads the enquiries already sitting in the CRM, whether they came from an email or a recorded phone call. It fills in a structured trip card and gives the enquiry a quality score, with the two or three reasons behind it. Then it writes follow-up questions for whatever is missing, in French, Dutch or English. The designer reviews them and sends them from the CRM. The answers go straight back into the card. Nothing is typed twice.

2. The design agent. From a clear brief, it builds one to three trip concepts. It works from past trips the agency actually sold and delivered, and from destination content the team has approved. Each version comes with a short note on why it made those choices. When something is unclear, it flags it instead of making it up. It never mentions a price or a named hotel.

3. The proposal agent. This is the workhorse. It reads the DMC quote in any format, including messy tables and workbooks with several tabs. Every stay, room type, service, transfer and date lands in the same set of fields, with a link back to where it appears in the original file. Then it writes the proposal in the brand’s own voice. It also flags anything missing, contradictory or uncertain in the quote. When a revised quote arrives, the old proposal is marked out of date and a new one is generated.

4. The pricing agent. It prices the same items. It is also the one place where the AI is not allowed to touch the numbers.

The AI never writes a price

This is the most important choice in the whole project.

A language model is great at reading a messy quote and writing a warm paragraph about a lodge in Botswana. You don’t want it anywhere near your margins. So we split the work in two. The AI handles the text. Plain, tested software handles the money.

In practice:

• The commercial owner writes down and signs the pricing rules: margins, seasons, taxes, currency, supplier specifics. At our client, this was the first time those rules existed in one place.
• Every change to the rules is kept, so you can always see which rules an offer was priced on.
• An offer below the minimum margin is blocked. Only the commercial owner can release it, and that is recorded.
• Every figure in the offer is checked against the number it came from. If they don’t match, the gap is shown.
• Net rates and margins are never sent to the AI service.

The trip and the price are built separately, then put together in one document. The designer checks it once and sends it once.

Built on what they already had

We did not sell them a new AI platform. The agency already paid for Microsoft 365 with Copilot and already kept its enquiries in a CRM. We set up and extended those tools, connected the agents to the CRM, the shared drive and the mailboxes, and put guardrails around them.

That choice removed the heaviest part of the work, and the cost that goes with it. It also means the team works in tools they already know.

We started in one part of the business, where the CRM was already live. That let us measure the whole chain, from first enquiry to priced offer. The agents are built so that rolling them out to the rest of the business is a matter of setup and content, not a rebuild.

Guardrails from day one

Travel data is personal, and some of it is sensitive: a health condition, a food allergy, a passport number. So we designed the guardrails in from the start instead of adding them at the end.

• Everything runs in Europe. The AI service works under a data processing agreement and does not train on the client’s data.
• Health and dietary details are removed before any text goes to the AI. The real values stay in the client’s own storage.
• Passports and ID documents never go to the AI.
• Automatic checks run on every result before a person sees it. Figures must match their source, required fields must be there, and every statement must link back to where it came from.
• Every run goes into a log nobody can edit, kept for at least six months. Spending on the AI is capped.
• Nothing reaches a traveller without a named person approving it.

The quality score got extra care. It scores the enquiry, never the person, and it never uses health or dietary information. It helps the team decide what to handle first. It does not decide who gets served, and no enquiry is ever dropped because of its score. Scoring people, guessing sensitive traits and pricing on someone’s personal behaviour are all ruled out by design. That keeps the agents in the lighter tier of the EU AI Act, with the required AI labels built in.

How we work with the team

The technical side is only half the job. The other half is working side by side with the people who will use the agents every day.

• Measure the starting point. Before building anything, we measured where things stood: time per proposal, days to a usable brief, errors in the figures, price gaps between designers. Every agent is judged against those numbers.
• Real work from week one. No demo data. Every milestone ends with a working demo on real enquiries and real quotes, and a clear decision on what comes next.
• Designers shape the tools. They defined the trip card with us in workshops. They tuned the quality score by comparing it with their own judgement on recent files.
• Size the hard part early. Reading quotes is the biggest piece of work, and the hardest to estimate. So we asked for 50 real quotes up front. That turned a guess into a measurement. Then we added formats one at a time, most common first.
• Keep a test you can rerun. Twenty real quotes, each paired with the offer that came out of it, form a test set. We rerun it after every change, so quality never slips without anyone noticing.
• Build shared pieces once. The design and proposal agents work from the trip card the intake agent builds. The quote reader will be reused by the next agents, for travel documents and booking checks.
• Leave the team in charge. Training, a user guide, and a handover pack on how it is built, how it runs, how to configure it and where its limits are.

What you can take from this

This is not really a story about travel. The same problem shows up wherever skilled people spend their days reading messy documents, retyping them and checking numbers: insurance claims, freight quotes, energy contracts, public tenders. The same four rules apply:

• Let the AI read and write. Keep the numbers in code.
• Build on the tools you already pay for.
• Work on real data from the first week.
• Put a named person in front of every result that leaves the building.

Where Ozymind comes in

This is what we do. We sit with your business team, pick the work worth handing to an agent, build it with them and take it to production. Along the way, we turn what we learn into pieces your own team can reuse: connectors, templates, test sets and documentation. That way the second agent comes faster than the first. More on our Agentic AI & Automation work. If you run a travel agency, the complete offer is ARIADNE.