Proactive AI: Building Suggestion Systems That Don’t Annoy

The most valuable thing an AI agent can do is speak up before you ask — the unbilled project, the contract sitting unsigned for nine days, the client who went quiet. It's also the fastest way to get muted. Every proactive system lives on a knife edge between indispensable and Clippy, and the difference is design discipline, not model quality.

The Interruption Budget

Treat user attention as a hard budget. Every proactive surface spends from it; every useless suggestion overdraws it; an overdrawn account gets the feature disabled. We design around three channels with very different costs:

The core rule: default every new suggestion type to the cheapest channel that works. A signature reminder starts in the briefing. It earns interrupt status only when it's about to breach an actual deadline. In our production systems, interrupts are rare enough that users read every single one — which is the entire point.

Anatomy of a Suggestion That Gets Accepted

1. Evidence first. "The Meridian SOW has been out for signature for 9 days; their last three took 2–3 days" — not "you might want to follow up on some documents."
2. One concrete action, pre-built. The nudge email is already drafted. Accepting costs one tap through the approval queue; the agent did the work before asking for attention.
3. Stated confidence. Suggestions carry the same confidence percentage as everything else the agent says. Users calibrate fast and forgive a hedged 70% far more than a confident miss.
4. A visible dismiss path that teaches. Dismiss options are specific: "not now", "not for this client", "never suggest this type." Each maps to a rule the agent actually applies.

Clippy failed on all four counts: no evidence, no prepared action, no confidence signal, and dismissals that changed nothing. Every annoying AI feature since has failed the same checklist.

The Feedback Loop Is the Feature

A proactive system that doesn't learn from responses is a spam generator with good grammar. We track acceptance rate per suggestion type per user, and enforce floors: a type accepted under ~20% of the time gets automatically demoted a channel — interrupt drops to briefing, briefing drops to ambient — and the user is told why. Suggestion types above ~60% acceptance become candidates for the opposite conversation: "you've approved this nudge 14 times in a row — want me to send these automatically?" Automation is offered as a graduation, never assumed.

Suppression Rules We Ship by Default

• Never repeat a dismissed suggestion within its cool-down (type-specific, days not hours)
• Cap concurrent open suggestions — ours is 5; the queue holds the rest
• Collapse related items ("3 unsigned documents" is one card, not three)
• Silence during detected focus: back-to-back meetings, late evening, weekends — batched channels absorb everything
• New suggestion types launch in shadow mode first: we log what would have fired for two weeks and review precision before a user ever sees one

Measuring the Knife Edge

• Acceptance rate per type — the health metric (target: 40%+ overall)
• Time-to-action on accepted suggestions — proves the pre-built action works
• Mute events — one user disabling a channel outweighs fifty acceptances; treat each as an incident
• Caught-value log — money and deadlines saved by proactive catches; this is the number that renews contracts

Proactive AI is a trust product. Users grant the agent the right to interrupt them exactly as fast as it proves the interruptions are worth it — and revoke it faster. Spend the interruption budget like it's your own attention, and the agent becomes the colleague who only taps your shoulder when it matters.