AI Concierge
Designing an AI guest assistant that knows when to answer, when to ask, and when to hand off to a human.
A static guidebook in a 24/7 business
Vacation rental guests have needs at all hours — WiFi passwords, broken TVs, dinner recommendations, the occasional emergency. Breezeway's Guide gave them static documentation; for everything else they texted the property manager. Every "where are the towels?" message pulled an operator away from real work, and every unanswered question was invisible: there was no structured way to learn what guests couldn't find.
The brief: turn Guide from a reference document into a digital concierge — one guests trust and property managers control.
One front door, five flows
Every guest message routes through an intent classifier into one of five purpose-built flows — Question, Issue, Request, Recommendation, Emergency — with a sixth path for unclear messages that asks a clarifying question instead of guessing. Designing the routing layer first kept every downstream flow honest: each solves one guest job, with its own data sources, escalation rules, and tone.
Confidence before answers
When the AI is confident, it just answers. A guest asking for the WiFi network gets the name and password in one tap, with a source link back into Guide to verify — no operator pulled away from real work. The scoring is what makes this safe: every retrieval is rated before the AI speaks, so a high-confidence answer like this is only ever shown when the system actually knows. When confidence is low it does the opposite — says so plainly, offers to check with the property manager, and logs a content gap. Property managers set the threshold and the tone. The screens below show the high-confidence path end to end.




Escalation is a feature, not a failure
Issues start with guided troubleshooting. When that fails, the conversation doesn't dead-end — the AI collects details and photos, confirms entry permission and timing, summarizes for the guest to confirm, then creates a fully-specified task in Breezeway. The ops team sees the task and full context in the messaging thread; the guest never repeats themselves. Urgent issues ("no hot water, ASAP") are auto-prioritized with an immediate PM alert.





Emergencies break all the rules
"There's smoke coming from the kitchen" is not a conversation. The Emergency flow inverts everything: no search, no pleasantries. Call 911 first, then the PM's direct line, extinguisher and exit locations — while the property manager is auto-alerted in parallel, without being asked.


Guardrails guests never see
Trust is mostly invisible. The AI may only draw from the property's Guide, Property Profile, and Blueprint — internal team notes are never surfaced, and access-code visibility rules are enforced so a door code can't leak before check-in day. Recommendations stay PM-curated (enriched with Google Places ratings and details) rather than open web search, keeping property managers in control of what their guests are told.
The right list, not a long one
A guest asking “any good places nearby?” doesn’t want a directory — they want the right answer. The AI narrows by intent first (what are you in the mood for, then what kind of food) before it recommends anything. Results blend two sources the guest never has to think about: the host’s own curated picks and live Google data — ratings, distance, hours — so a property manager’s favorite local spot sits alongside the highest-rated option down the street. The guest gets a recommendation that feels personal and current; the host gets to put a thumb on the scale without it reading like an ad.





Revenue without over stepping
Every flow has a designed upsell moment — strictly after the guest's need is met. Answer the WiFi question, then mention the hot tub. Solve the towel request, then offer the beach kit. Book the repair, then offer late checkout. On day 3, the AI proactively checks in, pairs the weather forecast with bookable local activities, and reserves them in-chat. Helpful first, commercial second — that ordering is what makes guests say yes.



Designed to be measured
I mapped nine canonical interactions — from a WiFi question to a kitchen fire — into a behavior matrix covering trigger, first action, data source, resolution type, upsell, PM involvement, task creation, and logging. It became the team's source of truth for scoping and the rubric for AI evaluation, paired with a full event-tracking spec (sessions, intent detection, flow completion and abandonment, response confidence, link clicks).
Self-service rate
% of guest questions resolved in Guide without escalation
Escalation quality
Fewer follow-ups needed after an issue reaches the ops team
Speed to resolution
Time from issue report → task assigned → staff dispatched
Content health
AI-surfaced Guide gaps reviewed and fixed by admins
Designing for fallibility
The most important moments I designed weren't the happy paths — they were the moments the system says "I'm not sure," hands off gracefully, and leaves the guest more confident, not less. Trust isn't built by the answers; it's built by what happens when there isn't one.