We research the conversations nobody optimises.

For there to be a conversation, you have to understand what was said, what was meant and when to reply. We work on all three at once, and we always start with the hard part.

In machines that converse there is a very well-drawn map and a whole territory left blank: the strong accent, speech affected by a disorder, the voice of an eighty-year-old, the language with few resources. It is not blank because it is impossible, but because it is expensive, slow and makes fewer headlines. We have worked there since 2005.