A real-time copilot that rides a live client call — designed around restraint. The hard problem wasn’t surfacing help. It was building a system whose primary job is staying silent, so the one cue it does give is trusted, not ignored.
PolicyAdvisor
Designer + build (design & front-end)
Designer-first · restraint-led
Shipped · in production
There are regulatory beats that must be hit — disclose pre-existing conditions, confirm the stability period, give the suitability rationale, explain exclusions, confirm limits — and a real client on the line who needs to feel heard. The instant the advisor is reading a wall of AI notifications, they’ve stopped listening, and the tool has become the problem.
So I didn’t design a dashboard of suggestions. I designed restraint — a context engine whose primary job is deciding what not to say, so the rare cue it does surface is the one the advisor trusts and acts on.
A live-call assistant that talks too much is the thing that loses the call. So the resting state is — nothing.
The resting state isn’t an empty dashboard waiting to be filled — it’s a deliberate silence. The advisor runs the call; the tool earns the right to speak.
Most copilots are designed around what they can show. This one is designed around what it refuses to.
The engine tracks everything — compliance beats, client signals, what’s been covered. Then it suppresses almost all of it.
The funnel is the product: many signals in, a suppression stage that drops almost all of them, exactly one cue out. The engineering exists to stay quiet.
Give the suitability rationale — explain why this plan fits Daniel’s specific needs before moving to close.
A context system that surfaces everything is just noise with a model attached. The intelligence is in the suppression.
When it does speak, it’s a single prioritised nudge — glanceable in the half-second an advisor can spare.
A strict queue of one. The next cue does not appear until the current one clears — the advisor is never reading a list while a client is talking.
Walk through the key policy exclusions so Daniel understands what is and isn’t covered — before he commits.
Two cues at once means the advisor reads instead of listens. The one-at-a-time rule is the whole point.
Every regulatory beat is watched and checked off as the advisor naturally covers it — nothing to manage mid-call.
The advisor never touches the checklist — it fills from the conversation. It’s ambient by design: present when glanced at, silent otherwise. Only an un-covered required item can escalate.
A checklist the advisor has to manage mid-call is another thing stealing attention from the client. So they don’t manage it.
It assumes competence. It doesn’t nag, repeat, or escalate — it stays silent so that the one time it speaks up, the advisor listens.
Daniel’s about to commit — and exclusions haven’t been explained. This is a required disclosure. Cover it before you take the close.
Trust isn’t a feature you add. It’s the silence you keep — so the one cue that matters is the one they act on.
Everything the engine does serves one outcome: many signals in, almost nothing out.
Designing from the bombardment failure instead of the feature list set what every screen had to do: stay silent by default, suppress 95% of what it tracks, surface one ranked cue at a natural pause, track compliance without asking the advisor to manage it, and break silence only to prevent a real miss. The restraint pattern — trusting the user enough to say nothing — is the spine of the whole PolicyAdvisor AI suite.
The hardest thing to design was everything it doesn’t do.
A live-call assistant is the rare product where the obvious version — helpful, present, surfacing insight — is the one that fails. Every notification is a moment the advisor isn’t listening to a human being who’s deciding whether to trust them.
The decision I’m proudest of has no UI: the resting state is silence, and the engine’s primary function is suppression. It would have demoed better with more on screen. It works better with almost nothing — because trust is the silence you keep, so the one cue that matters is the one they act on.