Case study · En Route Luxe · 2025–
A business run on agents
The lifelong travel advisor. One relationship that holds your family's story: the babymoon, the first safari with kids, the anniversary you didn't have time to plan. Every trip that matters, for decades.
En Route Luxe is a boutique luxury travel advisory I founded and run. Behind the human touch run three kinds of agents I built with Claude Code: one drafts every itinerary from the family's profile and runs its own check before I see it, one does the destination research, one answers client questions before and during the trip, all of them working inside one connected advisor workspace that holds each trip from the first brief to the last payment. Agents draft. Guardrails check. The founder edits.

By the numbers
Real clients, real revenue, real margin
- 20+
- clients a year
- 50+
- trips a year
- $700K+
- in bookings, 2026 to date
- 100%
- repeat and referral growth
Grown entirely through repeat bookings and referrals. No paid acquisition. 100% repeat and referral growth is the early proof of lifelong relationships: a client who hands you their next trip, and their sister's honeymoon, is telling you the relationship is the product.
What clients say
“We'd been to Africa before, but never like this. They knew exactly when and where to be — the migration crossing we witnessed wasn't luck, it was planning.”
“Japan was our dream trip for years. We were overwhelmed until En Route Luxe showed us it wasn't about doing more — it was about timing it right. Our kids still call it the best week of their lives.”
“What surprised us most was what they told us NOT to do. That restraint — knowing what to skip — made the trip feel effortless instead of exhausting.”
The operating stack
Agents draft. Guardrails check. The founder edits.
The advisory runs on a stack I built and ship myself: itinerary-building agents that compile a family's profile (explicit intake answers, conversational signals from the call, behavioral signals from past trips) into constraints, draft the day-by-day itinerary, build the client page, and run a check step; destination-research agents over a 190-destination corpus; client-query agents that answer from the trip record (one client on the road asked whether to book his flights; the agent found the same routing for about $2,000 less before I replied); and around them, Claude Code skills for the recurring work, a flight-search MCP server, and the client site and itinerary pages in React and TypeScript on Supabase and Vercel; and underneath all of it the connected advisor workspace, one trip record that every agent reads from and writes to. Every output ends at a decision point: approve, redirect, or kill.
Data and tools
- Trip record: brief, research, itinerary, quotes, proposal, money
- Destination intelligence: 190+ destinations
- Flight-search MCP server
- Markdown knowledge base
- Call transcripts and research ingest
Agents draft
- Coordinator: routes bounded tasks off the trip record
- Itinerary builder: profile to constraints to draft
- Document-and-sourcing agent: DMC quotes to records
- Itinerary QA critic: read-only, fresh context
- Research ingestion and meeting sync
- Post editors with receipt tests
- Lead replies drafted into the inbox
Guardrails check
- Closures on the day scheduled
- Ticket windows on a clock
- Transfer buffers
- Contingency paths
- Jet-lag pacing
- Verified vs estimated pricing
- Stale venues
- Quote totals, dates, required fields: checked in code
- Golden set: checker recall and precision
- Quote and finance acceptance cases
The editor layer: the founder decides
- Approve, redirect, or kill
- Invoices and money movement wait for approval
- Supplier selection stays human
- Disruption response stays human
- Taste stays human
The discipline that makes it work: knowing what NOT to automate. Judgment is the only step that does not scale, so the system spends the founder's time only there.
That boundary was not drawn in the abstract. A general model will book a museum that is closed that day, and it does not know that Shibuya Sky sunset tickets have to be sourced exactly fourteen days in advance. Further out, an African camp is effectively offline: there is no API, and the last mile is a human relationship. Each of those is a guardrail now. Cataloguing where a model is confidently wrong is the same practice as building an eval set, run from the operator side rather than the platform side. Today each class is a check: a golden set of trips with expected findings, a checker scored on recall and precision, a scorecard per draft, and a log of every accepted, rejected, or fixed finding, and every rule the system learns is versioned and dated. The policy is enforced in tooling, not prose: every action is act alone, draft, or escalate; the email permission is draft-only; every site change goes through a pull request. The itinerary builder, its golden set and eval command are public in agent-skills.
The connected workspace
One record per trip, from the first brief to the last payment
Every trip at En Route Luxe lives in one record inside a connected advisor workspace I built and operate. The record carries the client's brief, the destination we settled on, the research behind it with its sources, each version of the itinerary, the quotes the DMCs sent back, the proposal the client approved, and what is owed to whom and by when. A coordinator agent reads that state and hands out bounded tasks: planning work to a planning agent, supplier paperwork to a document-and-sourcing agent, and the finished work to a read-only reviewer that can check but cannot send, commit or invoice. When a preference or a supplier term changes, the record reopens the decisions that depend on it and leaves the rest alone.
Supplier document intelligence
DMC quotes arrive as PDFs, email threads and spreadsheets. The workspace reads them into a commercial record: travel dates, nights, rooms, rate basis, what is included, taxes, currency, how long the price holds, and the payment and cancellation terms. Every material field points back to the page, cell or message it came from, so a total or a deadline can be checked without rereading the quote. Before two quotes are compared, their scope is lined up, same dates, same party, same room category, with optional items kept to one side. Omissions are flagged, and a revised quote shows what changed instead of overwriting the original. Code, not the model, checks the totals, the date arithmetic and the required fields. A parsed quote, an approved proposal and a confirmed booking are three different states, and nothing moves between them on its own.
CRM and the money
The same record carries the client relationship and the money. Each client has a pipeline stage and a next action. Invoices are prepared from the terms the client approved, never from a draft. Receipts are matched to the obligations they settle; a partial payment, a duplicate import or a receipt that matches nothing goes to a review queue rather than being guessed at. Balances and deadlines are derived from the approved terms, so when I open a trip I can see who owes what, in which currency, and by when, without reconstructing it across separate tools.
Act alone, draft, or escalate
Every action class is sorted ahead of time into one of three buckets: act alone, draft for me, or escalate to me. The model drafts; the email permission is create-draft only; anything that issues an invoice or moves money waits for my approval; and the reviewer can flag a problem but has no authority to send, commit or pay. The arithmetic, the date checks and the state transitions live in ordinary application code, where they are deterministic and testable, and the agents are left with the work that needs language.
This is the Zillow discipline applied to a travel business: the model extracts and proposes, code validates, every field keeps its source, and the person approves the step that commits money or a promise to a client.
Destinations
Where the trips go
Itineraries designed across 30+ countries, including:
- Japan
- Kenya
- Tanzania
- South Africa
- Namibia
- Zimbabwe
- Italy
- Greece
- Croatia
- Thailand
- Cambodia
- Spain
- Portugal
- England & Scotland
- French Polynesia
- The Bahamas
- + more






Trips designed across four continents
The judgment story
Six weeks out, the airline cancelled
Kenya and Tanzania: 25 travelers, six families, timed to the Great Migration. Six weeks before departure, the airline cancelled the route.
An agent would have rebooked the same plan on the next available carrier. That is what optimization looks like when the objective function is “restore the itinerary.”
The actual call: re-route the entire group through London, absorb 12 time zones with kids, and turn the disruption into the trip's strongest opening. Camp selection followed the same logic: location beats brand in the reserve.
That wasn't a data lookup. It was taste, timing, sequencing, and knowing when the obvious answer is wrong. For trips that have to count, that judgment IS the product.

The same review pattern that gated 1M+ documents a year at Zillow now runs a travel advisory. Different stakes, same discipline: agents draft, guardrails check, the human decides.