Canary looked at 10 million AI guest conversations. Two thirds wanted something done, not answered

Canary Technologies published research on 28 October drawn from more than 10 million AI guest interactions across calls, texts and webchats at thousands of hotels in over 100 countries. The headline finding is a useful corrective to how guest-facing AI is usually sold.

Only 35% of conversations were someone asking a question. The other 65% were requests for something to happen: book a table, fix the air conditioning, extend a checkout. Of those action requests, 45% involved bookings or reservations, which is to say they had revenue attached.

That changes what a guest AI tool has to be good at. Answering "what time is breakfast" needs a knowledge base. Taking a restaurant booking or logging a broken unit needs write access to a system and a path to a person who can act. The second is a far harder integration problem, and it is where most of the volume apparently sits.

The other figures

Canary says 60% of its AI responses account for the individual guest's context, meaning reservation details, conversation history and preferences, rather than returning general information.

Guests whose messages were handled by AI were 11% more likely to leave a five star rating on in-stay surveys. Hotels using AI messaging grew ancillary revenue by 54% on average in their first six months.

The speed numbers are the most concrete. AI replies in a median of nine seconds. By Canary's measure of staff responses, 25% take 15 minutes or more and 12% take over an hour. Anyone who has watched a front desk at 4pm on a Friday will recognise those tails, and it is the tail rather than the average that produces a complaint.

The report is called "From Answers to Actions: What Millions of Guest Interactions Reveal About AI Inside the Stay" and sits behind an email gate on Canary's site.

What the research does not say

This is a vendor study of a vendor's own traffic, and it should be read that way.

Canary does not state the time period the 10 million interactions cover, how properties were sampled, or what mix of property types and segments is in there. It serves around 20,000 hotels in more than 125 countries, so the "thousands of hotels" in the sample is a subset chosen in a way the release does not explain.

The satisfaction and revenue figures are correlations, not controlled comparisons. Guests who choose to message a hotel during their stay are already engaged, and engaged guests rate stays higher. Hotels that switch on AI messaging in a six month window are usually changing other things too. The 54% ancillary growth figure is the one most likely to be quoted at you in a sales meeting, and the one worth asking the most questions about: growth from what starting point, at how many properties, and against what the rest of the estate did over the same period.

Why it is still worth reading

The action versus answer split is the part that transfers, because it is a question you can ask about your own channels without buying anything. Pull a week of guest messages, your inbox or your phone log, and sort them into questions and requests. If your own ratio looks anything like two thirds requests, then a chatbot wired only to a knowledge base will deflect a third of the volume at best, and the rest still lands on a person.

The follow-on question is where those requests end up. A tool that can take a dinner booking but cannot create the housekeeping task for the broken unit has solved the easier half. That is a PMS and integration question long before it is an AI question.