Case study
A front-desk agent for a small hotel
A small hotel was missing half of its phone calls. I built the agent that picks them up, and that answers email, web chat and WhatsApp too. This is how it works, how I check it, and what a review of its conversations found.
- Phone calls missed before
- 50%
- Channels, one agent
- 04
- Reviewed answers it can use
- 128
- Conversations in one review
- 60
The problem
Half of the hotel's phone calls were going unanswered. Guests also write by email, web chat and WhatsApp, in English and in Spanish, and they don't keep office hours. A small team can't answer all of it.
Picking up is the easy part. An agent that answers and gets a price or a policy wrong is worse than a missed call, so most of the work went into what it is allowed to say.
What I built
- One agent on all four channels, starting with the phone. A guest gets the same answer by phone, email, web chat or WhatsApp.
- It answers from a reviewed knowledge base. Every answer is waiting for review, approved or retired, and the agent only uses approved ones. Today that is 128 approved answers across 23 topics. Retired answers are kept, so there is a record of what guests used to be told.
- Prices come from the booking system. The agent quotes live rates and availability, never numbers from its own memory.
- When it doesn't know, it asks the team. The question goes to staff, and their answer can become a new entry in the knowledge base once someone has reviewed it.
- Groups and travel agencies are handled in their own modes.
- Staff have an assistant too. It drafts replies to guests and suppliers for them to send.
How I check it
- An automated reviewer reads finished conversations, and a separate judge scores tone.
- Every answer in the knowledge base counts how often it is used, so I can see which ones carry the most weight.
- I read conversations myself. One review covered 60 real ones from production.
What the review found
Seven of the 60 conversations had been passed to the team. Three of those seven were about wheelchair access, and none got a reply.
By the agent's own rules, each of those conversations had gone right: it didn't have an approved answer, so it passed the question on. The problem was what happened after that.
What changed
Accessibility is now its own topic in the knowledge base, with a reviewed answer. It doesn't say yes or no. It says which rooms and areas can be reached without steps, where the ground is hard to roll over, which areas are only reached by stairs, that the room doors fit a wheelchair and that the bathrooms have no grab bars. Then it asks the guest what they need, before they book.
What I took from it
- I found this by reading conversations, one at a time.
- Passing a question to a person only helps if someone answers it.
- The fix was an honest answer in the knowledge base, not a change to the model.