Designing Conversational AI Experiences

About the Talk

Most of us have now designed something with a chat box in it. Far fewer of us can say the thing actually helped anyone finish what they came to do. Shivam Sunderam has spent the past years designing AI experiences across search, assistants and conversational interfaces at trivago and eBay, where he works on AI innovation, and he came with the question underneath the hype: where exactly do these experiences break?

His answer was a map rather than a list of tips. A conversational AI experience is won or lost at five moments, and most teams only design one of them.

• ●Problem. It is fine to start from the technology, but you still have to work backwards to a real user problem. Find the intersection of what the model is good at and what someone actually needs, at one stage of their journey, not across all of it.
• ●Entry. A persistent entry point solves access, not value. A floating button in the corner tells nobody what it can help with or why it is better than search. Context is what creates relevance, so the assistant should meet people where the need happens.
• ●Intent. People rarely say what they mean upfront, and a blank prompt field asks them to be good at something nobody trained them for. Shivam's sharpest line: users should not have to become better at prompting, the interface should become better at helping them express intent.
• ●Response. In conversational AI the response is the interface. Treating it as text is the mistake; a long paragraph that cannot be skimmed, compared or acted on is a dead end with good grammar.
• ●Continuity. Do not make every conversation start from zero. The product already knows what the person told it five minutes ago, and behaving otherwise is the fastest way to feel less useful than a human who was paying attention.

The psychology thread running through all of it: when the answer is disappointing, people blame themselves and quietly get better at talking to the machine. That is a design failure being paid for by the user, and it is invisible in every metric except the ones nobody logs.

The room

Eighty-five people, the largest edition the series has run, and the interactive session turned into a swap of the habits people have built up over the last year: what they do differently now when they ask an AI for something, and which of those habits their neighbour was going to steal.