Hi, I'm on maternity leave from
Airtable, where I've spent the last year and a half designing AI agents
(field agents, to be exact) and AI credit-awareness experiences.
Every /now update gets my latest view on eval UX, a topic I've cared
deeply about since my ML data training days at
Snorkel AI. Here is my current take:
-
Eval surfaces (sources, citations, confidence scores) aren't
features. They're nudges to get our users to think critically before
they commit to an output or scale it up.
-
So the design question is: given what we know about AI quality and
our product, when and how should we nudge our users?
-
Better AI models have shrunk, but not eliminated, the moments that
need nudging. As designers, we shouldn't stop asking the question out
of fear of introducing friction.
Alright folks, enjoy your ber-month celebrations and we'll check in
again in 2027 🍂 🎃 🦃 🎄 ☃️