Predictive UI

Predictive UI surfaces the next likely action before the user asks for it. Instead of a menu of options, the screen shows the single decision the user is most likely to make and lets them confirm or reject in one tap.

Helix uses predictive UI for the approval queue: the next draft to approve, the next meeting to confirm, the next stalled deal to nudge. The prediction is informed by the agent’s recent activity, the user’s historical approval pattern, and the priority signals on each thread.

For privacy, Helix runs ranking logic client-side. The ranker does not see the user’s message bodies; it ranks on metadata (recency, sender heat, approval history). Server-side, Helix only stores the metadata needed to run the agent.

Implementation status: v1 uses a heuristic ranking (recency, sender, thread state, approval policy). Per-identity learning — adapting the ranker to each user’s historical approval pattern — is shipping in Iter 2.