Case study · TripWays · 2026
An AI concierge that answers travelers and tour operators 24/7, in their own language
Travelers and tour operators ask TripWays the same questions at every hour, in every language. I built one AI search over the whole catalog and put it behind every place they ask: site search, an assistant on every page, WhatsApp and our emails. Agents keep it current, prices come from the live booking system, and a person can step in at any time.
- Co-Founder & Head of Growth
- TripWays · AI search and concierge
- 2026
Public figures and test results, November 2026.
The problem
The same questions, at every hour, in every language
TripWays sells tours in 120+ countries with a small team. Travelers write from every time zone, and tour operators have questions of their own. Most of it repeats: is there a seat on Saturday, what’s included, where do we meet, what’s the total for three.
Answering by hand doesn’t scale. A chatbot that guesses is worse: one wrong price or date costs the booking and the traveler’s trust. So the brief was an assistant that answers at once, in any language, only from the real catalog and the live calendar, and hands over to a person when it should.
The system
One search, behind every door
Every place a traveler or an AI can ask a question reads the same index. Fix a listing once, and every answer gets better.
01 · Prepare
Agents ready each tour
- Clear titles and itinerariesAI
- Every place a tour visits, with its nicknamesAI
- Photos sized and checked
- Changed tours re-indexed every hour
02 · Index
One record per tour
- Each tour stored by meaning, not keywordsAI
- A short fact card: price, places, length, languages
- Tours that leave the site leave the index
03 · Answer
Rules around every result
- Stays in the city and the budget asked for
- Weak matches dropped, not padded
- Prices and seats from the live calendar
- Says so when nothing fits
04 · Doors
Wherever people ask
- Search on the homepage and tour pages
- An assistant on every pageAI
- WhatsApp, for travelers and partnersAI
- Tour picks in our emails
- ChatGPT and Claude, through MCPAI
Questions with no good answer show where the listings and the FAQs need work next.
How one answer is made
The model talks. The booking system decides.
An assistant that sells has to be right about money and dates. So the AI only does what AI is good at.
- 01UnderstandThe assistant turns the whole conversation into one clear search, in whatever language it was asked.
- 02FindOne search by meaning across the catalog, kept to the place and the budget the traveler gave.
- 03CheckDates, seats and the total come from the operator’s live calendar, in code the model can’t change.
- 04AnswerTour cards and a booking link that must match the quote. A person can take over any chat.
What I built
Five pieces of one AI search
Ingestion agents
Agents keep 16,000 listings ready to be found
Operators write listings in every style. Seven scheduled agents prepare them: clear titles, day-by-day itineraries, the real places each tour visits and the nicknames people use for them, and photos. Each has an off switch, a schedule and a batch size.
Every hour, changed tours are re-indexed. A tour that’s retired, paused or blocked leaves search at once, so the assistant never offers something that can’t be booked. A place counts only when the tour’s own page names it, so “Machu Picchu” finds the right tours however the operator spelled it.
RAG search
Search by meaning, within what the traveler asked
Every tour is one record, stored by meaning, so “a relaxed food tour for two under $100” finds food tours even when no listing uses those words. The price in the question becomes a filter, and a search on a city’s page stays in that city.
Early on, a question about a place we don’t sell got look-alikes: a cathedral in Colombia brought back tours in Argentina. Now the search checks the place against a world gazetteer first. If we have nothing there, it says so plainly and offers the closest real tours in that country.
AI concierge
One concierge on the site and on WhatsApp
The same brain answers the assistant on every page and the TripWays WhatsApp number, which went live in October 2026. Travelers get tour ideas, live dates and a total. Tour operators are recognized and routed to a partner desk, with a person one message away.
It answers in the language it’s asked in, day or night, and takes voice notes too. Payment never happens in a chat: the traveler gets a checkout link on tripways.com.
- 24/7
- for travelers and partners
- Any language
- in text or voice
Guardrails
Prices the model can’t get wrong
A language model will happily repeat an old price or fill in a date nobody gave. So the rules about money live in code, not in the prompt. The dates and group size are taken exactly as the traveler stated them, the total comes from the operator’s live calendar, and the quote is stored.
A booking link goes out only if it matches that stored quote. A tour with an age limit is refused for a group that doesn’t meet it. The model writes the words; it can’t change the numbers.
Testing and handoff
Tested before every release, with a person one message away
Every change to the prompt or the model is scored first on a copy, against 75 scripted conversations: follow-ups, tour details, bookings and WhatsApp, including one in Persian. Only then is it promoted, and only if nothing changed on the live version in the meantime.
Model choice follows the same evidence. One larger model passed every check but took 19 seconds to reply, so the faster one stays. When a traveler asks for a person, staff answer from Slack, the reply goes out from the support number, and the AI steps aside.
- 75
- scripted test conversations
- 99%
- of checks passed in the latest run (Nov 2026)
What changed
Support that’s always on, and a team that sees only what needs a person
Customer and partner support at TripWays is now agentic. The concierge answers around the clock, in the traveler’s language, from the live catalog. Staff step in only where a person adds something.
- Travelers and tour operators get answers at any hour, in their own language.
- Repeat questions about seats, inclusions, meeting points and totals no longer reach the team.
- Questions the AI couldn’t answer show where the FAQs and listings fall short, and they get fixed.
- Customer satisfaction is higher.
What I bring
What this means for your team
The same approach fits any business with a big catalog, customers in many languages, and more questions than people to answer them.
RAG search a business can trust
AI search over your real data, with rules that keep it to what the customer asked, and an honest answer when nothing fits.
Agents that maintain the data
Scheduled agents that clean, enrich and re-index your catalog, so every answer improves without a content team.
Conversational commerce
An assistant on your site and WhatsApp that quotes real prices and hands over a checkout, with a person one message away.
Evaluation before release
Scripted test conversations, staged releases and model choices made on evidence, so changes don’t surprise customers.
Questions
What is RAG search?
Retrieval-augmented generation: before the AI answers, it searches your own data and answers only from what it found. At TripWays it searches 16,000+ tours by meaning, then the live booking system supplies the dates and the price.
Can the assistant invent a tour or a price?
No. It can only show tours that come back from the search, and the total comes from the operator’s live calendar. A booking link goes out only if it matches the stored quote.
What happens when nothing fits?
It says so, and offers the closest real tours instead of look-alikes from somewhere else. Questions like that also show where the catalog needs work.
Does it replace the support team?
It handles the repeat questions, around the clock and in any language. A person can take over any conversation, and the AI steps aside when someone asks for a human.
Can I try it?
Yes: open the assistant on tripways.com or message the WhatsApp concierge. The same search also answers inside ChatGPT and Claude: see the MCP case study.
Want a growth system like this on your team?
I’m open to growth leadership roles, and I run fixed-price AI growth sprints for companies in any industry.