The problem
From the outside, travel planning looks simple. A traveller types "Where should I stay in Lisbon with good transport links and great food, under €100/day?" and the results appear instantly. But that's where the real work starts.
The traveller still has to read the reviews, compare neighbourhoods in Google Maps, cross-check blogs, validate prices, and figure out which advice is actually current and relevant. Then decide whether any of it fits their budget, their travel style, and their dates. All of it manual. All of it taking hours.
Most travel products answer that by handing people more information. This company wanted to hand them a single decision: ask one question, get back one recommendation that's structured, researched, current, grounded in real sources, and written with the judgment of a well-travelled local friend that is good enough to trust before spending any money.
No existing product could do that well enough. So they asked us to build it with them.
Why existing tools weren't enough
The customer didn't want a chatbot. They wanted a research agent that could produce advice: the kind a well-travelled friend gives, grounded in current facts about real places and their own experience.
It wasn't that simple. A standard AI chatbot gave answers that sounded polished but had nothing real underneath them. Search tools returned links, many of them outdated or irrelevant. They couldn't put any of that in front of someone about to book a trip.
Three deeper problems sat under the surface.
- Sources disagree. Good travel advice comes from many places, blogs, reviews, maps, listings, forums, local guides, and they contradict each other constantly. The system had to compare and reconcile them, not just summarise one. That takes real research, not a single shortcut.
- Research takes time. But if users feel like they're waiting on a blank screen, they leave. The wait had to feel earned, with the work happening where users could see it.
- They had to own it. A tool they couldn't change would turn into a liability the day the market moved. They needed control over the product, the code, and the roadmap.
What we built
We worked right next to the founding team, embedded in their day-to-day, shipping alongside them.
Instead of scripting a chatbot, we built an agent that plans its own research. It takes a traveller's question, breaks it into the threads that actually matter, and investigates each one against live web data through Ares, our agentic search engine. For the Lisbon question mentioned earlier, that means researching which neighbourhoods fit the budget, which areas have strong transport links, which places are known for food, what recent reviews and listings suggest, and where the trade-offs are.
It pulls from travel sources, local business listings, and reviews, then reconciles everything into structured destination intelligence in the form of a recommendation the platform could drop straight into its planning and booking flows, closing the loop from question to booking in one journey.
The biggest risk was hallucination. A travel product can't recommend a hotel, neighbourhood, or route on unsupported information; bad recommendations cost users money, waste their time, and break trust.
So we built the agent to stay grounded in what it actually retrieved. Guardrails kept it focused on the user's question and stopped it from drifting into claims the sources didn't support. It was built to tell the difference between what the evidence backed and where real uncertainty remained. The result was a pipeline that produced answers with structure, evidence, and practical judgment.
The customer had no infrastructure team, so this had to be more than a prototype. We designed, deployed, and operated it as a hosted production service their platform could call directly, returning structured output that fed straight into their product. Observability into every run meant the team could see how the agent behaved in production and catch issues before users did. That gave a small team a production-grade research backend without having to build an infrastructure from scratch.
And we made sure they weren't locked into us. They got the working code, the documentation, the deployment setup, and a roadmap for what comes next. The system runs on infrastructure they control, and their team could keep improving it after launch.
The result
After launching the agent, customers using the product spent 30% more through the company's platform.
Travellers could move from research to decision faster, and with more confidence. Instead of leaving the platform to compare blogs, reviews, and maps, they got a grounded recommendation inside the experience they were already using.
This gave the company a real edge. They had a working system, understood exactly how it ran, and owned the infrastructure behind it.
They still run it today, on their own infrastructure, with the freedom to extend it as their product and market evolve.