Background
Lighthouse is the leading commercial intelligence platform for the hospitality industry, used by revenue managers, distribution directors, and marketing teams to understand their market, price their rooms, and grow their business. It is a product built on data. The question facing design and product leadership in early 2026 was a hard one: what happens to a data product when AI makes data frictionless?
The industry signals were clear. Capabilities that had been a competitive advantage two years earlier, including automated rate recommendations, anomaly detection, and market share tracking, were fast becoming table stakes. The next wave would not be about surfacing information faster. It would be about acting on it autonomously. The question was not whether Lighthouse would become an agentic platform, but what that would look and feel like, and whether design could shape that direction before engineering did.
Three UX leaders were tasked with producing an answer in one week.
The challenge
This was not a standard brief. We were not designing features for an existing roadmap. We were asked to create a visiontype: a functional, high-fidelity prototype of a product that did not yet exist, representing a plausible three-year future. The purpose was not to hand off a spec, but to give the company a concrete artefact to react to. Something senior leadership, product managers, and engineers could hold up and say "yes, this is the direction" or "no, this is wrong, and here is why."
The constraint was a single week. Three designers in a room in Barcelona, with a rotating cast of stakeholders from product, engineering, revenue management, and data science joining remotely at key moments.
The sprint
Day 1: Understand & Inspire
We started by compressing weeks of context into hours. I used Dovetail AI to surface patterns from existing UXR, synthesising months of user interviews, session notes, and feedback into a structured map of commercial hotel teams' core jobs, frustrations, and unmet needs. What emerged was a consistent theme: hotel commercial teams spent the majority of their time managing data rather than acting on insight. The "commercial manager" persona was, in practice, a data wrangler.
We ran a Lotus Blossom exercise to break the problem space open, generating eight themes from a single central question: what does a hotel commercial team look like if AI handles 80% of routine decisions autonomously? By end of day we had a shared vocabulary and a scope. We would prototype the shift from passive dashboards to an agentic commercial workspace.
Day 2: Ideate
We used Gemini to run rapid scenario explorations, prompting the model to describe future-state workflows across different hotel types, team sizes, and commercial contexts. This was not about generating design ideas; it was about stress-testing assumptions quickly. Which agent actions felt helpful versus invasive? Where would a hotel GM need to stay in the loop? What does "autonomous" actually mean at 2am when a competitor drops their rates?
By afternoon we had moved from scenarios to wireframes. We split the prototype work between a web application targeting the commercial team's primary workspace and a mobile companion for the on-the-go revenue manager. We aligned on the core interaction model: natural language as the primary interface, proactive agent cards replacing passive data tables, and an approval-and-delegate pattern for autonomous actions. Then each of us took a section.
Days 3 & 4: Prototype
We used Claude Code connected to a shared git repository to build the prototypes collaboratively. Working from the same codebase meant we could see each other's decisions in real time, resolve inconsistencies as they appeared, and move faster than would have been possible with separate files and manual merging. AI accelerated the build: generating component scaffolds, translating wireframe logic into working interactions, and helping iterate on copy as we went.
The web prototype centred on a unified commercial workspace. The home view was not a dashboard but an agent feed: prioritised opportunities surfaced by the system, each with a recommended action and a one-click approval path. Agents could monitor pick-up, flag rate displacement, draft promotional offers, and update channel distribution. The human reviewed and approved rather than operated.
The mobile prototype focused on the manager who needed to stay in the loop without being at a screen. Push notifications for anomalies, voice-to-action for quick approvals, and a digest view summarising what the agents had handled overnight.
On Day 3 we presented the in-progress prototype to the Design & Research team and incorporated their feedback. On Day 4 the Data Science team joined. Their input on the feasibility of the agent behaviours sharpened the claims we made in the prototype and removed the ones we couldn't support.
Day 5: Present
We presented to senior leadership with both prototypes live in the browser. The session ran long, which was the right sign. The debate was not about the design. It was about the strategic implications. Which product areas would the agentic layer touch first? What did this mean for Lighthouse's positioning against the major OTAs and PMS players moving in the same direction? What was the 12-month path to making any of this real?
The vision we designed
The central concept was a commercial OS: a single workspace where AI agents handle the operational layer of revenue management, distribution, and marketing, freeing the commercial team to focus on strategy and judgment. The product metaphor shifted from "analytics tool" to "intelligent teammate."
Three design principles held the vision together:
- Proactive over passive. The product comes to the user with prioritised signals, not the reverse. The default state is an agent briefing, not an empty dashboard.
- Natural language as the interface. Querying data, approving actions, and delegating tasks all happen through conversation. The interface adapts to the request rather than requiring the user to navigate to a specific view.
- Human-in-the-loop by design. Autonomous execution for routine, bounded decisions. Mandatory human review for anything with material revenue impact. The approval UI is as important as the agent itself.
Design principle
The "commercial manager" persona we inherited from existing research was a data wrangler. The persona we designed for was a commercial architect: someone who sets intent, evaluates agent recommendations, and makes the calls that require judgment. The design sprint's deepest contribution was articulating this shift clearly enough that it could drive hiring, product, and positioning decisions.
AI as a creative collaborator
Using AI throughout the sprint, not just as the subject of the design work but as a tool within the process, changed how we worked in ways worth noting. Research synthesis that would normally take a week happened in an afternoon. Scenario generation that would have required a workshop ran in a focused two-hour session. Prototype code that would have taken a week to produce was functional by day four.
The speed was useful, but the more important effect was on the quality of debate. Because we moved faster through low-level decisions, we had more time for the questions that actually mattered: what should the agent be allowed to do autonomously, and what requires human approval? How does the product communicate uncertainty? What does "confidence" mean in an AI recommendation presented to a revenue manager at midnight?
These are the design questions that will define agentic products. The sprint gave us a week to sit with them rather than deferring them to engineering.
Outcome
The prototypes became a reference point across the organisation. Several concepts were picked up directly by product teams as starting points for new initiatives. The vision document and prototype recordings circulated well beyond the original stakeholder group.
The clearest signal of impact came later. Ernest, Lighthouse's new AI workspace for hoteliers at askernest.ai, has ended up closely resembling what we designed that week. A unified commercial intelligence workspace, proactive agent surfacing, natural language interaction. The vision sprint did not just produce a prototype. It produced a direction that the company subsequently shipped.
Impact
A week of focused, AI-accelerated design work became a reference artefact that shaped product direction across multiple teams. The new Ernest product, Lighthouse's commercial AI workspace for hotels, reflects the core concepts developed during the sprint. It is a rare case where a vision prototype directly preceded what the company built.