Travel Intelligence: AI that advises leaders, never decides for them
The decision intelligence platform behind Voyage. It turns traveller, booking, partner and market data into signals, explanations and decision briefs, and it is designed to be honest about what the AI does not know.
- Type
- Business intelligence platform for the company running Voyage
- My role
- Product designer, UX AI engineer and product strategist
- Users
- Leadership, functional leads and travel partners
- When
- 2026
- Reference product
- Voyage AI travel concierge
- Status
- Live prototype on synthetic data
Should VoyageAI increase Japan family travel investment in Q4 2026?
Evidence from demand, revenue, competitor and supply signals, with three scenarios side by side.
The problem
Voyage, the AI travel concierge, produces a constant stream of data: what travellers search for, how they plan with the concierge, what they book, where they drop out, how partners perform. The leadership team running the business needs that data to make decisions such as where to invest, which partners to escalate, and which product bets are working.
Most analytics tools give leaders more dashboards. AI analytics tools add something riskier: confident answers. A model that says “increase Japan investment, 87% confidence” sounds helpful, but it hides the assumptions, the data gaps and the fact that 87% is not a probability of success. In strategic decisions, overconfident AI is worse than no AI, because it shifts accountability to a system that cannot carry it.
My role
I designed and built Travel Intelligence end to end as the business side of Voyage: the information architecture, the AI analyst, the decision workflow, the governance views and the language the AI uses to describe its own certainty. It reads the same traveller segments and journey stages as Voyage, so a leader can trace any business signal back to traveller behaviour in the consumer product.
From signal to decision
The platform is built around one path. Each step has its own surface, and each surface hands off to the next.
Four decisions that shaped the platform
The AI facilitates decisions; it does not make them
In the Decision Room, a strategic question is framed with evidence in four groups: demand, revenue, competitor and supply. Three scenarios sit side by side with their upside, downside and key assumptions. A data limitations panel lists what the analysis cannot see, such as competitor spend that is not tracked. The AI facilitator answers questions about the evidence, but it is designed not to recommend an option.
Trade-off: leaders sometimes want the AI to just say what to do. Refusing makes the tool feel less magical, and keeps accountability with the people who own the outcome.
Every AI number says what kind of number it is
The platform never shows a figure without its status. Forecasts are labelled projections with a confidence interval. Scenario outputs are labelled simulations, not forecasts. Model confidence is explained as evidence completeness and assumption certainty, “not a probability of success”. Risk factors are labelled observed correlation. The labels are short, consistent and placed right next to the number they qualify.
Trade-off: labels add visual noise to clean charts. They are the cheapest protection against a leader repeating a simulation in a board meeting as if it were a forecast.
Signals show where they came from
AI signals on the Command Center carry a confidence level and how recently they were detected, and each one opens a workflow rather than a dead end. Alerts name their source, such as the demand model or partner data, and the rule that triggered them, for example “Hotel availability drops by more than 20%”. In AI Analyst, the active context is always visible as chips, such as Japan, family travellers and October 2026, so an answer can never silently drift to a different question.
Trade-off: showing sources and rules exposes how simple some detections are. That is a feature: a leader who understands the rule can judge the alert.
Make the health of the AI visible
Leaders are asked to trust insights built on data pipelines and models they never see. Data & Governance shows them: data lineage from the Voyage traveller app to insights, every source with its freshness, quality and status, each AI model with its version, accuracy and last training date, and warnings such as “retraining recommended”. Privacy, consent and an audit log sit in the same place.
Trade-off: showing a model in a warning state invites doubt about every insight it feeds. Hiding it would be worse when the doubt arrives later.
Where I overrode the AI
The AI tools I built with kept pushing towards confident answers and polished dashboards. These are the moments I said no.
A recommended option
AI proposed: highlight the best scenario in the Decision Room.
Removed. The facilitator lays out evidence and assumptions; people choose.
Unlabelled confidence
AI proposed: a confidence percentage on every insight.
Kept, but always with what it means and what it does not mean.
Hiding data gaps
AI proposed: show only the evidence that exists.
Added an explicit data limitations panel, so missing evidence is part of the brief.
Dashboards by department
AI proposed: one dashboard per business function.
Built role views on one shared truth, so every leader sees the same numbers through their own lens.
Outcome
Travel Intelligence is a live prototype on synthetic data, designed as the business layer of Voyage. It has not been used to run a real business.
What exists
- Command Center, eight intelligence areas, AI Analyst, Decision Room, Alert Center, Role Intelligence and Data & Governance
- A consistent vocabulary for uncertainty across the product
- A decision workflow with evidence, scenarios, trade-offs and data limitations
- Role views for internal leaders and external partners
Still open
- Testing with leadership and partner users
- Reconciling metric definitions across views
- Defining what each model’s accuracy figure measures
- Connecting decisions to their outcomes after the fact
If you are adding AI to analytics
AI makes analytics faster to read and easier to over-trust. The work that keeps it useful is design work:
- Label every number with what kind of number it is
- Show the source and rule behind every signal
- Put missing data in the brief, not in a footnote
- Let the AI facilitate the decision, and record who made it
- Make data and model health visible to the people who rely on them
Voyage shows how I design what an AI may do for its users. Travel Intelligence shows how I design what an AI may claim to the people running the business.