Summary
Travel Intelligence is the decision intelligence platform for the business that runs Voyage, an AI travel concierge. It brings traveller, journey, destination, commercial, product, partner and market data into one place, uses AI to detect signals and explain them, and gives leaders a structured way to make decisions with that evidence.
The case study is about one design problem. AI makes analytics faster to read, and much easier to over-trust. A confident AI answer in a strategy meeting can quietly become the decision. I designed Travel Intelligence so the AI is always useful and never the one deciding: every number says what kind of number it is, every signal shows where it came from, missing data is part of the brief, and the health of the data and models is visible to the people who rely on them.
It is a live prototype on synthetic data and has not been used to run a real business. The final sections set out the gaps and what I would test first.
The problem
A travel platform at Voyage’s scale generates signals constantly: search intent shifts, conversion dips after a price change, a partner’s response times slip, a competitor adds routes, a destination suddenly trends. The leadership team has to decide which of these matter, why they are happening, and what to do about them.
Three ways analytics fails leaders
- Too many dashboards, too few answers. Each function builds its own view. Numbers disagree, and the first half of every meeting is spent reconciling them.
- Signals without a path to action. A dashboard shows that something changed, then leaves the leader to work out why and what to do in a different tool.
- AI that sounds more certain than it is. AI analysts produce fluent explanations and confident percentages. Without context, a correlation reads as a cause, a simulation reads as a forecast, and a confidence score reads as a probability of success.
The third failure is the newest and the most dangerous, because it gets worse as the AI gets better at sounding right. The design question was: how do we get the speed of AI analysis without letting it quietly take over decisions it cannot be accountable for?
How it connects to Voyage
Voyage is the traveller-facing product: a concierge that plans, books and monitors trips. Travel Intelligence is what the business behind it uses to see how that product and its market are performing. The two are designed as one system.
- The same traveller types. Voyage designs for solo travellers, families, business travellers, honeymooners, students, small groups, luxury and budget travellers. Travel Intelligence segments its data the same way, so a business signal about families can be traced to how families actually use the concierge.
- The same journey. The journey funnel in Travel Intelligence follows the traveller from discover and inspire through research, compare, plan and book to prepare, travel and return, the stages Voyage’s screens are built around.
- The product measuring its own AI. AI itinerary adoption is a tracked signal. In the prototype data, AI-assisted planning sessions convert 2.4 times more than sessions without it. Leaders can see whether Voyage’s AI features earn their place.
- One data lineage. Governance starts at the Voyage traveller app and follows events through the data platform and models to the insights leaders read.
The demo story runs across both products. Voyage shows a family planning a trip to Japan. Travel Intelligence shows the business noticing Japan family demand rising 27%, explaining why, and deliberating whether to invest more in it.
Role and principles
I designed and built Travel Intelligence end to end: information architecture, the Command Center, the AI Analyst, the Decision Room, alerts, role views and governance, and the vocabulary the AI uses to describe its own certainty. I built the prototype with AI assistance, and the same discipline applied to the build as to the product: AI drafted, I decided.
Five principles
- The AI facilitates; people decide. The AI can gather evidence, explain and simulate. It does not choose between strategic options.
- Every number states its kind. Actual, forecast, simulation, correlation or confidence: the label sits next to the figure.
- Every signal shows its source. Which model or data feed detected it, and which rule fired, if any.
- Missing data is evidence too. What the analysis cannot see belongs in the brief, not in a footnote.
- Trust is inspectable. Data freshness, quality and model health are visible to the people relying on them, not only to the data team.
Who uses it
The prototype’s main persona is a chief intelligence officer, the person who owns how the business turns data into decisions. Around them, Role Intelligence defines views for nine internal roles and five external partner types.
Leadership
“Where should we invest, and what is at risk?”
CEO and chief intelligence officer. Work across all areas, mostly through the Command Center and Decision Room.
Functional leads
“What is happening in my area, and why?”
Product, revenue, destination strategy, marketing, customer experience, data and AI, partnerships and operations.
External partners
“How are we performing on this platform?”
Airlines, hotels, tour operators, destination organisations and enterprise travel buyers, each with a view scoped to their relationship.
Information architecture
The AI tools I built with proposed the conventional structure: one dashboard per department. I organised the platform around the path from signal to decision instead, with the intelligence areas as the evidence layer underneath.
Eight intelligence areas
Traveler, Journey, Destination, Commercial, Product, Partner, Market and Role Intelligence each go deep on one dimension of the business, and all share the same time range and market filters. They are where a leader goes when a signal needs more context than an AI summary can give.
AI everywhere, but in one voice
“Ask AI” is available from the header on every screen and opens the same analyst with the current context attached. The analyst, the Decision Room facilitator and the AI signals all use the same vocabulary for uncertainty, so a label means the same thing wherever it appears.
Command Center
The Command Center answers one question: what needs my attention across the business right now? It is designed to be scanned in a minute and to lead somewhere useful from every element.
What it shows
- Headline KPIs with trend lines and change against the prior period: active travellers, bookings, travel GMV, conversion, average trip value and the share of AI-assisted planning.
- Travel Pulse, a map of demand by region with trending destinations, and Regional Momentum, which compares each region with its prior period.
- AI Signals: anomalies and trends the AI detected, each with its size, a plain-language explanation, a confidence level and how long ago it was found.
- Emerging Opportunities ranked by demand score, and a Watchlist of metrics running against a threshold, with severity and trend length.
- Traveler Snapshot by segment, and a Journey Funnel from discover to return that highlights the biggest drop.
Every signal starts a workflow
The most important design decision on this screen is behavioural, not visual. Signals, opportunities and watchlist items are all entry points: select one and it opens its workflow, such as an AI Analyst investigation or a triage view. When the funnel shows a drop between plan and book that correlates with a checkout pricing change, the anomaly card says so and offers “start workflow”, rather than leaving the leader to open three tools.
Trade-off: making everything actionable means designing many follow-on paths. A static dashboard is cheaper to build and much easier to ignore.
The intelligence areas
The Command Center and Decision Room sit on top of eight areas of evidence. Each answers a different kind of question, and each connects back to traveller behaviour in Voyage.
| Area | The question it answers |
|---|---|
| Traveler | Which segments are growing, what they spend and how well they convert |
| Journey | Where travellers drop out between discovering a trip and returning from it |
| Destination | Which places are rising or cooling, and why |
| Commercial | How bookings, GMV, margin and trip value are moving |
| Product | Which features, including Voyage’s AI features, drive engagement and conversion |
| Partner | How airlines, hotels and experience providers are performing against their commitments |
| Market | What competitors and the wider market are doing |
| Role | What each leader or partner needs to see from all of the above |
The journey funnel as a shared language
The journey funnel deserves a note. It follows nine stages, discover, inspire, research, compare, plan, book, prepare, travel and return, and shows sessions per 10,000 and the conversion between each step, with the largest drop highlighted. Because Voyage’s own screens map to the same stages, a product leader and a commercial leader can talk about the same drop in the same words. In the prototype, the funnel flags a fall between plan and book that correlates with a checkout pricing change, and offers to start an investigation.
AI Analyst
The AI Analyst is where a leader asks “why?” It is designed as an investigation workspace rather than a chat window, because strategic questions are rarely answered in one exchange.
Context you can see
The active context, for example Japan, family travellers, October 2026 and bookings, is always shown as chips, both in the side panel and next to the question box. Context can be added or removed explicitly, and follow-up questions keep it. A leader always knows exactly what the answer is scoped to, which prevents the most common failure of AI chat: an answer to a subtly different question than the one in the user’s head.
The shape of an answer
Every answer follows the same structure, so leaders learn where to look:
- A one-line headline, then a short summary.
- Confidence, recency and the number of data sources behind the answer.
- Key drivers, each with its size and an “explain” option.
- Who and where: which segments and locations drive the change.
- Momentum over time.
- Risk factors, explicitly labelled as observed correlation.
- Actions: explore in depth, create a report, create an alert, save the insight or share it.
- Follow-up questions that deepen the investigation, including “what if” simulations.
From the prototype. Asked why Japan demand is increasing, the analyst answers that demand is up 27%, driven by family travel, AI itinerary adoption and a favourable currency, with high confidence across seven data sources. It breaks the change into drivers, segments and cities, then lists hotel inventory, airfare and visa friction as risks, labelled as observed correlation rather than cause.
Trade-off: a fixed answer structure is less conversational than free text. It is far easier to scan, compare and challenge.
Decision Room
The Decision Room is the centre of the case study. It is where a signal becomes a decision, and where the AI’s role is most deliberately limited.
A decision is a question, not a dashboard
Each decision starts as a framed question with a status, for example “Should VoyageAI increase Japan family travel investment in Q4 2026?”, in deliberation. The frame states the period, markets and segments in scope, and four headline indicators: the demand index, the hotel supply gap, family NPS and the AI’s evidence confidence.
Evidence in four groups
Evidence is grouped by kind: demand signals, revenue signals, competitor signals and supply signals. Each group includes what is known and, where relevant, what is not. One competitor line reads, in effect, “no intelligence on this competitor’s investment yet — gap”. Putting the gap inside the evidence, rather than leaving it out, was a deliberate choice.
Scenarios with their assumptions
Three scenarios sit side by side: continue the current strategy, increase investment, or reduce it. Each shows projected GMV, revenue, growth, conversion and margin, a segment comparison, and its trade-offs as potential upside, potential downside and key assumptions. The assumptions matter most. Scenario B, for example, only holds if the hotel supply gap is resolved, the family campaign hits its return target and a product release ships on time. Making those visible turns a debate about outcomes into a debate about assumptions, which is the debate leaders can actually resolve.
Data limitations
A dedicated panel lists what the analysis cannot see or has not modelled: competitor spend, the outcome of a supply negotiation, price elasticity at higher demand, a partner’s budget commitment. These are often the things that should decide the outcome.
A facilitator, not an advisor
The AI panel in the Decision Room states its role plainly: it does not recommend a decision; it provides evidence, clarifies assumptions and surfaces missing data. Its suggested questions are about the evidence, such as which assumptions drive the simulation, what data is missing, what changed since the last review and which segment carries the most risk. Below the panel: “AI provides evidence support only. Human judgment required.”
Stakeholders and the decision record
The room also holds stakeholder input and a decision record, so the final call is attributed to people, with the evidence and assumptions as they stood at the time.
Trade-off: an AI that refuses to recommend can frustrate leaders who want a fast answer. It also means the AI can never be blamed for a decision, or credited with one, and the organisation keeps learning how to decide.
One signal, end to end
The clearest way to see the platform is to follow one signal from detection to decision. Every step below uses a real screen in the prototype.
Sense: the Command Center
An AI signal appears: Japan family travel demand up 27% in 30 days, high confidence, detected 12 minutes ago. At the same time, the Alert Center has a critical alert from partner data: Tokyo hotel availability 34% below projected November demand, triggered by the rule “Hotel availability drops by more than 20%”. One signal is an opportunity, the other a constraint on it.
Understand: the AI Analyst
The chief intelligence officer selects the signal, which opens an investigation with Japan, family travellers and October 2026 already set as context. The analyst explains the rise: school holidays aligning with the cherry blossom season, AI itinerary features driving Japan discovery in Voyage, new flight routes and a favourable exchange rate. It shows which segments and cities carry the growth, and lists hotel inventory and rising airfares as risks, labelled as observed correlation.
Decide: the Decision Room
The question becomes a decision: should the business increase Japan family investment in Q4? Evidence arrives grouped by demand, revenue, competitor and supply. Three scenarios show their projected outcomes and, more importantly, their assumptions. The increase scenario depends on resolving the hotel gap through emergency allocation, which the supply evidence says is feasible in four to six weeks but not certain. The data limitations panel notes that competitor investment is not tracked.
The leadership team asks the facilitator which assumptions drive the simulation and what data is missing. It answers both. It does not say which scenario to choose.
Record and follow up
Stakeholder input and the decision are recorded against the evidence as it stood. From the analyst, the team creates an alert so they hear if Tokyo hotel availability worsens further, closing the loop back to sensing.
At no point did the AI make the decision. At every point it made the decision faster and better informed.
The language of uncertainty
The most reusable output of this project is a small vocabulary. Each label is short, means one thing, and appears directly beside the number it qualifies.
| Label | Means | Where it appears |
|---|---|---|
| Forecast, with interval | A model projection, not a guarantee, with its confidence range | Demand charts, Decision Room evidence |
| Scenario | A simulated outcome under stated assumptions; not a forecast | Decision Room scenarios |
| Model confidence | Evidence completeness, assumption certainty and past accuracy; not a probability of success | Scenarios, analyst answers |
| Observed correlation | Things that move together; not proven cause | Risk factors, anomaly cards |
| Signal confidence | How strongly the data supports a detected trend | AI signals, opportunities |
| Gap or limitation | Evidence the analysis does not have | Evidence groups, data limitations panel |
Writing these labels was harder than designing the charts. Each had to be accurate enough for a data scientist and plain enough for a CEO reading on a phone before a meeting.
Alert Center
Alerts turn the platform from something leaders visit into something that finds them. The Alert Center groups alerts by severity, critical, warning, info and success, and by area, and lets each one be resolved, snoozed or analysed.
Rules and models, kept distinct
Every alert names its source, such as the demand model, partner data, the SLA monitor or product analytics. Alerts triggered by a rule show that rule in plain language, for example “Partner SLA drops below 95%”. Alerts detected by a model, such as a conversion dip, show the model as their source. The distinction matters: a rule is a decision someone made in advance and can be reviewed; a model detection is a judgement that needs more scrutiny.
Good news counts
Success alerts, such as AI itinerary adoption crossing 40% or family NPS reaching a new high, sit alongside problems. An alert system that only reports bad news trains people to dread it.
Leaders write their own rules
“Create alert rule” lets a leader turn a recurring question into a standing rule, and the AI Analyst can create an alert directly from an answer, so an investigation can end with “tell me if this happens again”.
Role Intelligence
The AI tools proposed separate dashboards per department. I designed role views on one shared set of data instead, so every leader sees the same truth through their own lens.
Each role view, such as CEO, product leader, revenue leader or partner manager, opens with the questions that role cares about, its key metrics, the AI signals relevant to it, recommended actions, risks and a performance trend. The CEO view, for example, focuses on traveller growth, revenue trajectory, market expansion, platform health and strategic risks, with actions like reviewing an expansion brief or addressing a partner escalation.
External partners in the same model
Airlines, hotels, tour operators, destination organisations and enterprise travel buyers each get a view scoped to their relationship with the platform. Designing partner views in the same model means a conversation with a hotel partner about missed response times uses the same numbers both sides can see.
Trade-off: one shared model forces agreement on metric definitions, which is slow and political. Without it, the platform would simply move the reconciliation argument from spreadsheets into the product.
Data and governance
Leaders are asked to trust insights built on pipelines and models they never see. Data & Governance makes that foundation visible, in language a non-specialist can read.
- Data lineage from the Voyage traveller app through the event stream, data platform, intelligence layer and AI models to the insights on screen.
- Data sources, each with type, freshness, volume, quality and status, and a short note when something is wrong, such as a partner API running slow.
- AI model status for each model, including demand forecasting, journey optimisation, personalisation, partner health, price sensitivity and anomaly detection, with version, accuracy and last training date. A model past its retraining window is flagged with “retraining recommended”.
- Data quality across four measures: freshness, completeness, accuracy and coverage, each with its definition beside it.
- Privacy and consent, and an audit log of activity over the last 30 days.
The design intent is that a leader looking at a surprising insight can answer “should I trust this?” in two clicks: which sources feed it, are they fresh, and is the model behind it healthy.
Design standards for executive tools
Leaders read this platform in short bursts: before a meeting, on a phone, while someone else is talking. That set its own standards.
- Headline first. Every card, answer and alert leads with one sentence that is useful on its own. Detail follows for those who need it.
- Change over level. Numbers are shown with their movement and comparison period, because leaders act on change, not on absolute values.
- Calm by default. Colour signals severity and direction only. Most of the screen is neutral, so a red watchlist item stands out.
- Never colour alone. Severity and status always carry a text label, and trends carry arrows and signs.
- One time range, everywhere. The period and market selector applies across the platform, so a leader who switches to 7 days sees every number on that basis.
- Accessible by design. The prototype includes an accessibility control on every screen and a skip link to main content.
Where I overrode the AI
The AI tools I built with consistently pushed towards confident answers, clean dashboards and more automation. These are the moments I said no.
| AI proposed | I changed it to | Why |
|---|---|---|
| Highlight the recommended scenario | A facilitator that does not recommend | Strategic accountability cannot sit with a model |
| A confidence percentage on every insight | Confidence with its meaning stated | An unexplained percentage reads as a probability of success |
| Show only the evidence that exists | Gaps inside the evidence, plus a limitations panel | Missing data often decides the outcome |
| Risk factors stated as causes | Labelled as observed correlation | Correlation repeated in a meeting becomes a cause |
| One dashboard per department | Role views on one shared model | Different dashboards mean different numbers |
| Hide data and model plumbing | A readable governance section | Trust needs something to inspect |
| A chat window for the analyst | An investigation workspace with visible context | Strategic questions take more than one answer |
Known gaps
A platform that asks leaders to be honest about uncertainty has to be honest about its own. These are the gaps I found reviewing the prototype.
| Gap | Why it matters | Fix |
|---|---|---|
| Metrics that differ across views, such as active travellers and AI adoption on the Command Center and the CEO view | The platform’s core promise is one shared truth | One metric registry with definitions, used by every view |
| Model accuracy as a single number | Accuracy against what, over which period, is not stated | Define each metric beside it, as data quality already does |
| Some alerts have no visible rule | Model-detected alerts need their own explanation | Show the model and the threshold or pattern it detected |
| Decisions are not yet linked to outcomes | The organisation cannot learn how good its decisions were | Revisit each decision record against what happened |
| Synthetic data only | The design is proven on believable data, not real data | Pilot with one region’s real data |
Test plan
- Do labels change decisions? Show leaders the same scenario with and without the uncertainty labels, and see whether their confidence and choices differ.
- Does a non-recommending AI frustrate or help? Run a decision with the facilitator as designed and with a version that recommends, and compare decision quality, time and how accountable participants feel.
- Can leaders answer “should I trust this?” Ask them to judge a surprising insight using only Data & Governance, and time how long it takes.
- Is the answer structure scannable? Compare structured analyst answers with free-text answers for speed and accuracy of understanding.
For teams adding AI to analytics
AI makes analytics faster to read and easier to over-trust. The work that keeps it useful is design work, and it transfers to any domain where AI informs people with real accountability: finance, healthcare operations, supply chain, public policy.
- Give every number a kind: actual, forecast, scenario, correlation or confidence, labelled where it appears.
- Show the source and rule behind every signal and alert.
- Make missing data part of the evidence.
- Let the AI facilitate the decision, and record the people who made it.
- Make data and model health readable by the people who depend 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.