Back to Research

White Paper · 06 · June 2025

Proactive AI

Surfacing What You Need Before You Ask

Nick Brandt & Leo Gestetner 12 min read AI Architecture

Abstract

Every AI assistant today works reactively: user asks question, AI processes, AI responds. But the most valuable information is often what you didn't know you needed. Proactive AI monitors context — calendar, screen, location — and surfaces relevant information before users ask. This paper presents an architecture for anticipatory AI that feels intelligent rather than merely responsive.

1. The Reactive AI Problem

Every AI assistant today works the same way:

  1. User asks question
  2. AI processes
  3. AI responds

This is fundamentally reactive. The AI waits for you to know what to ask.

But the most valuable information is often what you didn't know you needed.

Reactive AI: User asks "What's on my calendar?" AI responds "You have a meeting with John at 2pm."

Proactive AI: Without any prompt, AI surfaces "You're meeting John at 2pm. Last time you discussed the Henderson project. He mentioned concerns about timeline. Here's what you promised to follow up on."

2. Context Triggers

Proactive surfacing is triggered by context changes:

TriggerWhat to Surface
Calendar event approachingRelated facts, previous interactions, commitments made
Email from contactRelationship history, pending items, shared context
Document openedRelated documents, previous versions, collaborator notes
Location changeLocation-relevant information, nearby contacts
Time-basedDaily review, upcoming deadlines, expiring commitments
Screen contentInformation related to what you're currently viewing

3. The Meeting Prep Example

Traditional Workflow

  1. See meeting on calendar
  2. Wonder what it's about
  3. Search email for context
  4. Search notes for background
  5. Try to remember what was discussed last time
  6. Go into meeting unprepared anyway

Proactive Workflow

  1. 15 minutes before meeting, notification appears
  2. "Meeting with John (Henderson Project)"
  3. Summary of last interaction
  4. Commitments you made ("send updated timeline")
  5. His recent concerns (from your notes)
  6. Documents you both referenced

The Key Insight

No search. No recall effort. Context is ready. The AI knew what you'd need because it understood the relationship between calendar events, contacts, and your knowledge base.

4. The Commitment Tracker

People make commitments constantly:

  • "I'll send that by Friday"
  • "Let me follow up on that"
  • "I'll think about it and get back to you"

Most are forgotten. A proactive system:

  1. Detects commitment language in conversations
  2. Extracts deadline (explicit or implied)
  3. Tracks status (pending, completed, overdue)
  4. Surfaces at relevant moments

Proactive Notification

"You told Sarah you'd send the proposal by Friday. It's Thursday and you haven't sent it."

5. The Knowledge Graph Foundation

Proactive surfacing requires understanding relationships. When you have a meeting with John, graph traversal finds:

  • John → works_on → Henderson Project → has_deadline → March 15
  • John → mentioned_concern → "timeline slipping"
  • You → committed_to → John → "send updated timeline"

6. Privacy Implications

Proactive AI requires deep context awareness: screen content, calendar, email/messages, documents, location.

This is precisely why on-device processing matters:

All Analysis Local

Processing happens on your device, not in the cloud

No Screen Content Transmitted

What you're viewing stays on your machine

No Email Content Sent

Your communications remain private

Context Stays Private

Your knowledge graph never leaves your device

Proactive AI without privacy guarantees is surveillance. With on-device, it's assistance.

7. The Interruption Balance

Proactive doesn't mean intrusive. Design principles:

Do Surface:

  • Information with clear relevance to current context
  • Time-sensitive items (deadlines, meetings)
  • High-importance facts at appropriate moments

Don't Surface:

  • Low-relevance associations
  • Information you've recently seen
  • Trivial facts unless specifically relevant
  • Anything during focus time

8. Implementation Architecture

┌─────────────────────────────────────────┐ │ CONTEXT MONITOR │ │ ├── Screen capture (OCR) │ │ ├── Calendar events │ │ ├── Active application │ │ └── Time/location │ └────────────────┴─────────────────────┘ │ ▼ ┌─────────────────────────────────────────┐ │ RELEVANCE ENGINE │ │ ├── Graph traversal │ │ ├── Temporal relevance (decay) │ │ ├── Importance scoring │ │ └── User preference weighting │ └────────────────┴─────────────────────┘ │ ▼ ┌─────────────────────────────────────────┐ │ SURFACING DECISION │ │ ├── Threshold check │ │ ├── Interruption budget │ │ ├── Deduplication │ │ └── Timing optimization │ └────────────────┴─────────────────────┘ │ ▼ Notification / Sidebar

9. Measuring Proactive Value

MetricWhat It Measures
Surfacing precision% of surfaced items user found useful
Information discoveryItems surfaced user didn't search for
Time savedAvoided searches due to proactive surfacing
Commitment trackingReminders that prevented missed items
Dismissal rateItems surfaced but ignored (too aggressive?)

10. The Shift in AI Interaction

Generation 1: Keyword search

Generation 2: Natural language query

Generation 3: Proactive surfacing

Each generation reduces the burden on the user to know what to ask.

11. Conclusion

Proactive AI is the realization that the best interface is no interface — information appears when relevant.

By combining context monitoring, knowledge graphs, temporal relevance, and on-device privacy, we can build AI assistants that don't just answer questions but anticipate needs.

The goal isn't an AI that responds better. It's an AI that makes asking unnecessary.

Further Reading