How the Compass AI builds an itinerary: architecture and safeguards

Generating a tailor-made travel program in about thirty seconds looks like magic. It isn't. Behind that first draft lies a multi-stage pipeline, explicit consistency rules, and, at the end, an advisor who validates everything. This article lifts the hood: how Compass AI moves from a brief to a structured itinerary, where the safeguards sit, and why the final word always belongs to a human.

From brief to structured intent

It all starts with the brief entered by the advisor: destination, dates, number and profile of travelers, indicative budget, areas of interest. Before any generation, this brief is normalized. The goal is to turn free text and form fields into actionable intent: from it we extract hard constraints (departure and return dates, group size) and weighted preferences (food, nature, desired pace).

This is also the stage where the traveler's CRM profile, when one exists, enriches the brief. If the client has already taken the swipe-based preference test, their radar profile (culture, nature, urban, food, relaxation, adventure) guides the generation. A traveler whose radar leans clearly toward relaxation and food will not receive the same first draft as an adventure profile. The brief is therefore not a simple list of keywords: it is a structured representation of what we are trying to produce.

Structured generation, not a block of text

The core of the pipeline does not produce a paragraph of prose to be cut up afterward. It directly produces a structure: a sequence of days, each broken down into timestamped steps, with geolocated places, activities, and accommodation suggestions. This formatting constraint is deliberate. Asking the AI to fill in a known framework, rather than letting it write freely, reduces randomness and makes each element individually manageable by the advisor.

In concrete terms, the generated first draft is made up of objects that the interface knows how to display and edit:

• Days: one day = one unit, with a guiding theme (arrival, urban exploration, nature excursion).
• Timestamped steps: a morning, midday, afternoon, or evening slot, with an estimated duration.
• Places: points of interest, restaurants, or sites, tied to coordinates so distances can be checked.
• Accommodation: one or more options consistent with the budget and the stay's location.
• Transition times: travel between two steps is not glossed over but treated as a full-fledged element.

This granularity is what later makes it possible to swap out an activity, shift a time slot, or change an accommodation without having to regenerate the entire program. The advisor works on building blocks, not on a fixed text.

The validation layer: consistency and safeguards

A structured AI output is still an AI output: it might suggest a museum open on a closing day, two overlapping activities, or an unrealistic journey within the day. That is why the first draft goes through a validation layer that generates nothing but checks everything. Its role is to test the proposal against explicit, readable, and correctable rules.

Two families of checks structure this stage:

1. Temporal consistency: slots within the same day do not overlap, durations leave room for meals and travel, and the program fits within the brief's date window.
2. Geographic consistency: the places strung together within a day are compatible with realistic travel times, with no absurd round trip from one end of a region to the other.
3. Business safeguards: we verify that the proposed options stay within the announced budget envelope and are suited to the travelers' profile (families, mobility, pace).

The advisor as decision-maker, not spectator

Once the first draft is structured and validated, it lands in the hands of the advisor. This is the fundamental difference from a black-box generator that would hand over a take-it-or-leave-it PDF. At Compass, the generated program is an editable starting point: days are reordered, an experience is substituted, an address known to the agency is added, the tone of the copy accompanying each step is adjusted. The AI proposes a complete structure in a few seconds; the advisor applies their expertise, their on-the-ground knowledge, and their client relationship to it.

This division of roles is not just a stance: it is what makes the gains measurable without dehumanizing the service. In the case of Borealis Travel Group, unifying under a white-label platform translated into roughly six hours saved per tailor-made program and a quote-to-sale conversion rate up 22 %, with no additional hiring. The freed-up time did not vanish: it was reinvested in the client relationship and in the volume of quotes sent.

Why this architecture, and not another

We could have simply asked a language model to write a nice program and display it. That shortcut poses two problems: it makes the output hard to verify, and impossible to edit at a fine level. By explicitly separating the stages - structuring the brief, structured generation, validation, human review - each link has a clear responsibility and can be improved independently of the others.

This modularity is also what makes it possible to keep the core commitment: 92% of time saved on creating a program, without sacrificing quality or the advisor's control. Speed comes from structured generation; reliability comes from the validation layer; the added value, for its part, remains human. An AI that speeds up the agency's work, never an AI that decides in its place.