01The problem
An online shop gets the same questions every day: where is my order, how much is shipping, which product should I choose. They arrive in the evening, at weekends, from several countries. If one person answers, while also running the business, the customer waits hours for information that already exists.
02How it is used
- The customer writes on WhatsApp, as they would to the shop. No app to install, no form to fill in.
- The assistant recognises the language and replies with the shop's own information: orders, shipping, products.
- On delicate cases it hands over. Returns, complaints and out-of-scope requests go to a person, and from that moment the assistant stops replying.
- It says what it is. Inside the conversation it states that it is an automated assistant, rather than hiding it in the terms of use.
03How it is built
- ChannelWhatsApp, on the shop's own number
- Understandinga language model interprets the question and the language of the person writing
- Knowledgeanswers are based on the shop's information; when it doesn't have it, it looks it up there
- Handoverrules that recognise the cases to give to a person; after the handover the assistant goes quiet
- Infrastructurean always-on server: nights, weekends and public holidays included
04Results
Over six months of real conversations the first reply arrives in 4 seconds (median), against 23 hours when a person answered, and 80% of conversations close without human intervention. These are medians, not averages: the averages would have looked more impressive, but they were skewed by conversations picked up again days later.