Building a multi-sided AI commerce network for hospitality.

Lycia AI started as a guest-facing autonomous agent. The thesis was simple: hotel staff are stretched thin, guests want immediate answers, and a domain-trained AI sitting on top of the existing tech stack could close that gap without forcing a rip-and-replace. We built the agent end-to-end — voice, chat, multilingual, integrated to the hotel's PMS — and brought it to market.

The product worked. Customers shipped it, guests used it, and the metrics moved. But on the way to market we ran into the same wall over and over: the long tail of independent 5–50-room hotels — the hotels that would benefit most from autonomous staff augmentation — often didn't have a PMS to integrate with. Our overlay model required a system that, for half our potential market, wasn't there.

We resolved it in two moves. First, we leaned into what was already working — the hotel deployments — and re-pointed the revenue model so the agent didn't sit alone. Hotels run Lycia for free; revenue comes from commissions when the agent recommends and books at restaurants, music venues, and activities in its network. The product becomes a multi-sided commerce engine, and every new participant makes the network more valuable for everyone already in it.

The second move is the one this portfolio leads to: an AI-native PMS, bundled below the agent, designed for the hotels we currently can't reach. The research, the customer evidence, the vertical pricing experiments, and a working concept demo are all here.

The product is an autonomous guest agent — voice, chat, and multilingual — domain-trained on hospitality with persistent guest memory, MCP integrations to PMS / payments / OTAs, and a four-to-six-week deployment cycle. Each vertical gets its own persona: Emma for hotels, Olivia for airports, Mia for malls, Nick for restaurants, Lucy and Alex for offices. One AI brain underneath, vertical-specific workflows on top.

The hardest design problem was trust at the edges of the conversation. The agent had to know when it should answer (most cases), when it should escalate to staff (the right cases), and when it should refuse (the regulated cases). The AI-first inbox — the staff-side surface where AI-drafted replies, sentiment-driven escalation, and load-balanced auto-assignment all converge — is the answer we shipped. The case-study deliverable below the outcomes section walks that decision in detail.

A subscription priced at the value Lycia delivers would be expensive. Hotels would resist it, sales cycles would extend, and the long tail would never adopt. The pivot: give the agent away free on the demand side, and capture revenue on the supply side — when the AI recommends a restaurant, books a music venue, or sells an activity, commission flows back into the network. Hotels deploy with no friction; partners pay for a steady stream of qualified, in-destination demand.

Each supply-side vertical has a paired pitch. The partnership deck sells the partner on joining the network ("we bring the guests to you"). The revenue engine deck upsells them their own Lycia AI for direct customers — five revenue levers, costed and modeled. Hotels themselves have two pricing experiments: free-with-voice-add-on, and a flat $10/room/month variant that pays for itself with one or two upsells per day.

The current product needs a PMS to integrate with. That's fine for chains and for the boutique segment that already runs Mews, Cloudbeds, or StayNTouch. It's not fine for the long tail of 5–50-room independents — the segment with the most acute staffing pain, and the one most receptive to a single bundled solution. The expansion thesis: build an AI-native PMS underneath the agent, sell the bundle, and capture the segment that's been left out. The deliverable below is the full argument.