Travel and hospitality / Lodging marketplace
Airbnb: Classifying and routing voice support calls
- Company
- Airbnb
- Country
- United States
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
The existing IVR was a fixed menu tree. Callers had to press buttons and follow a set path. General-purpose speech recognition models often got Airbnb's own vocabulary wrong. They turned listing into lifting and help with my stay into happy Christmas Day.
Technology and data
When the IVR answers, it asks the caller to say in a few sentences why they are calling. The reply is transcribed by an ASR model built for Airbnb. Instead of a general pre-trained model it was switched to one tuned for noisy phone audio, and a domain phrase list was optimised so Airbnb terms are recognised. A Contact Reason Detection model sorts the call into categories such as cancellation and refund or account problems. Help articles are indexed in a vector database as embeddings, up to 30 are pulled by cosine similarity, and an LLM reranking model orders them again. The link to the top article is sent by SMS and app notification. Before the link goes out, a Paraphrasing model reads back a one-line summary of the intent. When the caller asks for an agent, a separate intent detection model catches it and connects them to a person.
Results
Across several hundred voice samples the word error rate fell from 33 percent to about 10 percent. Intent detection latency stays under 50ms on average and article retrieval usually finishes within 60ms. Matching of the summary phrasing was evaluated by people and exceeded 90 percent precision. Airbnb said that as the error rate came down, the accuracy of help recommendations rose, the NPS of customers who used the ASR menu improved, and reliance on agents and handling time fell.
Limits and open questions
The resolution rate for all support contacts and the cost saving are not given as numbers in the source. The improvement in self-resolution was only confirmed directionally in an experiment with English-speaking hosts.
Sources
Compiled from public sources. These are not results from ATF Works customers.