How We Built a Voice Agent for a Hospitality Chain: Metrics After 90 Days

When a hospitality group with 8 properties — 5 boutique hotels and 3 independent restaurants — asked us to build their AI voice agent platform, we knew immediately that a standard deployment wouldn’t work.

Each property had its own brand, its own menu or room inventory, its own operational quirks. The hotels ranged from a 40-room beachfront resort to a 180-room urban property with a full restaurant and conference facilities. The restaurants were equally varied: a casual Mediterranean bistro, a fine dining establishment with a 3-month advance reservation waitlist, and a high-volume brunch spot that handled 400 covers on a busy Sunday.

What they had in common: all of them were losing business through their phones.

This is the full story of what we built, what the first 90 days looked like, what the metrics show, and what we wish we’d done differently.

The Problem Across All 8 Properties

Before writing a single line of code, we spent two weeks doing intake calls with the GM or manager at each property and pulling three months of call data where available.

The findings were consistent enough to be uncomfortable:

Hotels were missing 18-26% of inbound calls during peak periods (check-in, check-out, weekends). Front desk staff were handling in-person guests and physically couldn’t pick up the phone simultaneously. After-hours, calls went to a voicemail that many potential guests didn’t bother leaving a message on.

Restaurants were losing reservations at dinner rush. The busiest hours for incoming reservation calls (6-8 PM, Friday and Saturday) were exactly when the host was occupied seating guests. Call data showed answer rates as low as 35% during Friday evening peaks at two of the three properties.

Repetitive informational inquiries were consuming disproportionate staff time. At one hotel, the GM estimated that 40% of all calls were questions that could be answered with a fact from the property’s website: check-out time, parking cost, pool hours, pet policy, cancellation terms. Staff were pulling themselves away from guests to answer the same questions hundreds of times per week.

Multi-language calls were being mishandled. Two of the hotel properties are in tourist-heavy markets that see significant international traffic. Staff sometimes struggled with guests who didn’t speak English confidently, leading to miscommunications about reservations and services.

A conservative estimate across the portfolio: $40,000–$60,000 per month in direct and indirect revenue impact from these combined problems.

What We Built: A Multi-Property Dashboard Platform

The client’s requirement was explicit: they didn’t want 8 separate voice agents they had to call us to update. They wanted a dashboard where each property manager could log in, update their own information, configure their own routing, and see their own call analytics — without technical knowledge.

This made the build significantly more complex than a single-property deployment. We weren’t building one voice agent. We were building a platform that could host many, with each property’s configuration isolated from the others.

Platform Architecture

We built on Retell.ai as the underlying voice infrastructure, with a custom Next.js dashboard managing the per-property configuration layer. Each property had its own Retell agent instance, dedicated phone number, and knowledge base.

Each property manager could control: knowledge base updates (menus, policies, amenities, hours), routing rules for escalation, booking integration settings, and voice/language settings. A restaurant manager updating the brunch menu updated the dashboard directly — the agent had the new information within minutes.

Build time: 6 weeks for the full platform and all 8 configurations. A single-property deployment would have been 2-3 weeks; the dashboard infrastructure added the rest.

Reservation Handling

For hotels, the voice agent integrated with each property’s PMS. Two properties ran Opera. Two used Mews. One used a legacy system that had no API, which required a workaround — the agent could check availability (we built a read integration) but couldn’t create reservations in the PMS directly. Instead, it collected the booking details and fired a webhook to the front desk inbox with all the information formatted for manual entry. Not ideal, but it worked.

For restaurants, we integrated with OpenTable and Resy where applicable. One restaurant used a spreadsheet-based reservation system (genuinely). For that property, the agent took reservations and wrote them to a Google Sheet via the Sheets API, which then auto-notified the host team by SMS. Not the most elegant solution, but the restaurant’s owner wasn’t ready to adopt a reservation platform.

Concierge Knowledge Base

Building the concierge layer required the most time per property. Each manager completed a structured knowledge base document covering 80+ common questions, local recommendations, operational details, and edge case policies.

The quality of this input directly determined the quality of the responses. Properties that invested 4-6 hours in their documents had agents handling 85%+ of informational inquiries without escalation. Sparse documents meant constant escalations to staff — defeating the purpose. The AI is only as good as the information it has.

The First 30 Days: What Went Wrong

Honest deployments involve problems. Here’s what we hit in the first month.

The Opera PMS Integration

One hotel had an Opera version that differed slightly from our test environment. This caused reservation writes to fail silently — the agent confirmed bookings that never appeared in Opera. We caught it on day 3 via a guest who arrived for a reservation that didn’t exist. We fixed it same day, combed through the 12 affected reservations, and the GM personally called each guest. The lesson: test integrations on the live environment before launch, not just a test environment.

Language Detection on Low-Quality Mobile Calls

Automatic language detection worked well in testing on clean audio. In production, poor mobile connections confused detection — particularly with accents between language categories. We added an explicit language selection prompt when detection confidence fell below a threshold. Slightly less seamless than automatic detection, but significantly more reliable.

Fine Dining Restaurant Reservation Complexity

The fine dining property’s reservation experience involved discussing tasting menu formats, wine pairing options, specific dietary restrictions, and complex waitlist logic — well beyond “book a table.” The agent escalated 65% of calls to the maître d’ because callers had questions it couldn’t answer competently.

We narrowed the agent’s scope: standard bookings (parties of 2-6, standard tasting menu) handled fully; anything more nuanced routed to the maître d’ with context. Reducing scope increased success rate on handled calls. An agent that knows its limits and routes gracefully outperforms one that tries to handle everything and fails unpredictably.

Metrics After 90 Days

By day 90, the platform was running stably across all 8 properties. Here’s what the data showed.

Hotel Metrics (5 properties combined)

Call answer rate: 98.3% (up from 79.1% baseline). The remaining 1.7% were cases where callers hung up before the agent completed the greeting — typically repeated rapid presses of “0” indicating they specifically wanted a human.

Inbound reservation conversion rate: Increased 19% on average across hotel properties. More answered calls plus consistent upselling behavior from the agent (room upgrades, extended stays, add-ons) both contributed. The best-performing hotel saw a 28% increase in reservation conversion — largely driven by the agent’s consistent offer of the ocean-view upgrade, something front desk staff weren’t reliably doing on every call.

Average incremental revenue per reservation from upselling: $17. Across approximately 1,200 monthly reservations across 5 properties, this added $20,400/month in incremental revenue from upselling alone.

Informational inquiry handling: 72% of informational calls handled completely by the agent without escalation. Staff estimated this freed up 2-3 hours per property per day previously spent on repetitive phone inquiries.

Guest phone satisfaction scores: Improved 16 points on the 100-point scale in post-stay surveys. The relevant survey question: “How satisfied were you with the ease of reaching our team by phone?” Most notable improvement at the urban hotel, which had previously struggled most with call volume during peak periods.

Restaurant Metrics (3 properties combined)

Reservation capture rate during dinner service: Increased from 41% (average during peak hours pre-deployment) to 94%. This was the single most impactful result. Restaurants were losing nearly 60% of calls during the hours when demand was highest. That’s an enormous revenue leak.

No-show rate: Decreased 23% with automated day-of confirmation calls from the voice agent. The agent calls confirmed reservations 4 hours before the reservation time, allows guests to confirm, modify, or cancel with a single key press, and frees the time slot for same-day demand when a cancellation comes in.

Takeout and delivery order accuracy: 96.4% on AI-processed orders versus the pre-deployment rate of 88.1% on staff-processed phone orders. The AI’s confirmation read-back (“Let me confirm your order: one lamb burger, no onions, extra tzatziki, and one Mediterranean salad with dressing on the side — does that sound right?”) was more consistent than staff under dinner-service pressure.

Waitlist utilization: The automated waitlist at the brunch property filled 38 tables per month from the waitlist (previously near zero — the host was too busy to manage it manually). That’s 38 additional covers per month from demand that already existed but wasn’t being captured.

Combined 90-Day Financial Summary

These numbers represent a conservative attribution — we excluded any metrics where causality was ambiguous:

Category	Monthly Impact
Incremental hotel revenue (upselling)	+$20,400
Additional restaurant reservations captured	+$31,000 est.
No-show reduction (restaurant)	+$8,200 est.
Additional takeout orders from improved capture	+$11,400 est.
Staff time redirected from phones	Not monetized
Total monthly revenue impact	~$71,000

Platform cost: custom build ($25,000 upfront, this was a larger project than our standard voice agent subscription given the multi-property dashboard requirement) plus ongoing platform costs across all 8 properties.

The build cost recovered in the first month of operation. The ongoing ROI since then has been consistent.

What We’d Do Differently

Require knowledge base completion before build starts. Three properties submitted incomplete documents and tried to fill gaps during the build. We now require the knowledge base document to be complete before we begin any configuration work.

Test PMS integrations on live environments 2 weeks before launch. The Opera incident would have been caught with a 2-week live-environment test. This is now a hard requirement for all hospitality deployments.

Give high-complexity properties a separate launch timeline. The fine dining restaurant needed a more sophisticated configuration than we initially scoped. It should have had a 2-week standalone testing period before going live with the other properties.

Build manager training into deployment day. Three property managers didn’t understand how to update their knowledge base until a week after launch. A 45-minute session at launch would have prevented outdated information persisting in the agent.

The Ongoing Work

90 days in, the client has a cadence of weekly check-ins with us to review call analytics across the portfolio, flag any agent responses that weren’t ideal, and update knowledge bases for seasonal changes.

The voice agents have become operationally important enough that the client is actively working on adding two more properties to the platform. The speed of each additional property onboarding has decreased as we’ve refined the knowledge base intake process and standard integrations — new properties are now live within 10 business days of kickoff.

For more context on how voice agents work in hospitality, see our broader post on AI voice agents for hospitality businesses.

Frequently Asked Questions

How did hotel guests react to talking to an AI?

Better than we expected. Across 90 days of calls, explicit complaints about AI interaction were under 1% of total calls. The voice quality (we used Cartesia voices) and conversation naturalness meant most callers didn’t identify the agent as AI unless they asked. When they did ask, the agent acknowledged it honestly. The small percentage of callers who preferred a human were transferred immediately, and most of those were existing guests who had a specific request that required operational knowledge.

Did this reduce hotel or restaurant staffing?

No properties reduced headcount as a result of the voice agents. The honest outcome was staff redeployment, not reduction. Phone calls — particularly the repetitive informational kind — had been pulling front desk staff and hosts away from in-person guests constantly. With the voice agent handling the phone, staff were more present, more attentive, and less rushed during in-person interactions. Several GMs noted that employee satisfaction around phone-handling tasks improved noticeably. The phone had been a source of stress. Now it largely wasn’t.

How do the voice agents handle PMS or reservation system failures?

Every integration has a fallback. If the PMS connection fails during a call, the agent detects the failure, apologizes for a “technical issue with our booking system,” takes the caller’s name, contact number, requested dates, and room type, and sends an immediate SMS to the front desk with the information for manual booking. It also tells the caller that the front desk will call to confirm within 30 minutes. Integrations fail in hospitality — often at the worst times. The fallback experience needs to be planned and tested before launch, not improvised when something breaks.

What happens with the fine dining restaurant’s more complex calls now?

The revised configuration works well. The agent handles approximately 55% of reservation calls fully — these are the straightforward ones. The other 45% are routed to the maître d’ with the agent providing context: “I have a reservation inquiry for a party of 6 this Saturday. I’ll connect you with our maître d’ to discuss the experience — please hold for a moment.” The agent’s role in the fine dining context shifted from “close the booking” to “qualify and route” — which it does effectively. Total calls that require the maître d’s time decreased by 40% because the agent screens out and handles the simple bookings independently.

Is a custom multi-property platform the right approach for all hospitality businesses?

No. A single hotel or restaurant can deploy a voice agent without a custom dashboard for around $1,000/month using our standard subscription model. The custom dashboard we built here was justified because the client had 8 properties with ongoing self-management needs and wanted a branded platform they could expand. For a single property or small group, the overhead of custom platform development isn’t warranted. Start with a standard deployment and scale the infrastructure if the portfolio grows.