Unclassified
IntegrateWise — AI Workbench Specification
Version: 1.0.0 | Date: 2026-07-03 | Status: Human-First AI Management Surface
Philosophy: The AI Workbench is where humans manage the AI's work. It is separate from the human workbench. It is organized. It is human-controlled.
Table of Contents
- What the AI Workbench Is
- What the AI Workbench Is NOT
- The Seven AI Workbench Panels
- Navigation & Access
- Human Control Model
- Data Flow
- States & Interactions
- Accessibility
- TypeScript Interfaces
- Appendix: Difference from Audit Logs
What the AI Workbench Is
The AI Workbench is a dedicated management surface where humans can:
- Review all AI proposals in one place, organized and filterable
- Organize AI actions into categories, projects, or workflows
- Monitor AI performance — accuracy, acceptance rates, outcomes
- Explore AI memory — what the AI knows, how it learned it, confidence levels
- Adjust AI behavior — thresholds, scope and review-priority rules
- Audit AI decisions in context, not just as a raw log
- Manage AI workflows — what the AI is running, what it will run, what it completed
The AI Workbench is the human's window into the AI's operations. It is not the AI's window into itself. The AI does not "use" this interface. Humans use it to manage the AI.
┌─────────────────────────────────────────────────────────────────┐
│ HUMAN WORKBENCH │ AI WORKBENCH (separate tab) │
│ ───────────────── │ ───────────────────────────── │
│ Entity list │ Proposals Queue │
│ Detail pane │ AI Actions History │
│ Human actions │ AI Memory Explorer │
│ Timeline │ AI Workflow Monitor │
│ Search │ AI Performance Dashboard │
│ Projections │ AI Governance Settings │
│ │ AI Model Management │
│ → Human does work here │ → Human manages AI here │
│ → AI is subtle, ambient │ → AI operations are surfaced │
│ → AI helps when asked │ → Human controls and organizes │
└─────────────────────────────────────────────────────────────────┘
What the AI Workbench Is NOT
| NOT This | Because |
|---|---|
| AI Home / Chat | The AI Home is for conversation. The AI Workbench is for management. |
| Audit Logs | Audit Logs are passive historical records. The AI Workbench is active operational control. |
| AI Dashboard | The AI does not own this. The human does. It is not "AI's dashboard" — it is "human's AI management surface." |
| Default View | The human workbench is the default. The AI Workbench is a secondary, opt-in surface. |
| Auto-Executing | Nothing in the AI Workbench auto-executes. Every action requires human approval. |
| AI-Proposed UI | The AI does not redesign this interface. Humans configure it. |
The Seven AI Workbench Panels
Panel 1: Proposals Queue
Purpose: All AI proposals, organized and actionable.
Layout:
┌──────────────────────────────────────────────────────────────┐
│ Proposals Queue [Filter ▼] │
│ ─────────────────────────────────────────────────────────── │
│ ┌─────────────────────────────────────────────────────────┐│
│ │ [○] All [○] Pending [○] Approved [○] Rejected [○]││
│ │ Executed ││
│ └─────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ ⚠️ PENDING (3) ││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Acme Corp health dropped to 62. Escalate to VP? │││
│ │ │ Confidence: 82% │ Source: Twin │ Time: 2h ago │││
│ │ │ [View Context] [Approve] [Modify] [Dismiss] │││
│ │ └────────────────────────────────────────────────────────┘││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Schedule QBR with Northwind for next week? │││
│ │ │ Confidence: 91% │ Source: Twin │ Time: 5h ago │││
│ │ │ [View Context] [Approve] [Modify] [Dismiss] │││
│ │ └────────────────────────────────────────────────────────┘││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Draft follow-up email to Meridian's champion? │││
│ │ │ Confidence: 76% │ Source: Twin │ Time: 1d ago │││
│ │ │ [View Context] [Approve] [Modify] [Dismiss] │││
│ │ └────────────────────────────────────────────────────────┘││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ ✅ APPROVED (12 today) ││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Escalated Acme Corp to VP CS — executed 2h ago │││
│ │ │ Confidence: 82% │ Approved by: Sarah │ Outcome: ✓ │││
│ │ │ [View Details] │││
│ │ └────────────────────────────────────────────────────────┘││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Scheduled Northwind QBR — executed 5h ago │││
│ │ │ Confidence: 91% │ Approved by: Sarah │ Outcome: ✓ │││
│ │ │ [View Details] │││
│ │ └────────────────────────────────────────────────────────┘││
│ └──────────────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────────────┘
Key Features:
- Filter tabs: All / Pending / Approved / Rejected / Executed
- Batch actions: Select multiple proposals → Approve All / Dismiss All / Modify All
- Context view: Click "View Context" → see Entity360 + reasoning trace + evidence
- Confidence visible: But not dominant. Small badge, not large score.
- Time-ordered: Most recent first. Human scans quickly.
- No auto-execute: Every proposal requires human action. No "auto-approve after X hours."
States:
- Empty: "No pending proposals. The AI is watching."
- Loading: Skeleton cards for proposals.
- Error: "Could not load proposals. Retry."
- Success: Proposal list with actions.
Panel 2: AI Actions History
Purpose: What the AI did, when, why, and the outcome.
Layout:
┌──────────────────────────────────────────────────────────────┐
│ AI Actions History [Filter ▼] │
│ ─────────────────────────────────────────────────────────── │
│ │
│ ┌─────────────────────────────────────────────────────────┐│
│ │ [○] All Actions [○] Proposals [○] Executions [○] ││
│ │ Enrichments [○] Memory Updates [○] Tool Calls ││
│ └─────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Timeline ││
│ │ ││
│ │ Today ││
│ │ ───────────────────────────────────────────────────────── ││
│ │ 09:14 AM ├─ Escalated Acme Corp to VP CS ││
│ │ │ Confidence: 82% │ Approved by: Sarah ││
│ │ │ [View Details] [View Entity360] ││
│ │ ││
│ │ 07:30 AM ├─ Scheduled Northwind QBR for 2026-07-10 ││
│ │ │ Confidence: 91% │ Approved by: Sarah ││
│ │ │ [View Details] [View Entity360] ││
│ │ ││
│ │ Yesterday ││
│ │ ───────────────────────────────────────────────────────── ││
│ │ 04:22 PM ├─ Enriched Acme Corp entity with 3 new ││
│ │ │ contacts from LinkedIn ││
│ │ │ Confidence: 88% │ Auto-approved (enrich) ││
│ │ │ [View Details] [View Entity360] ││
│ │ ││
│ │ 02:15 PM ├─ Twin analyzed Q2 churn pattern ││
│ │ │ Confidence: 95% │ Read-only (no action) ││
│ │ │ [View Details] [View Report] ││
│ │ ││
│ └──────────────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────────────┘
Key Features:
- Timeline view: Chronological, grouped by day.
- Action types: Proposal (human-approved), Execution (AI performed), Enrichment (data added), Memory Update (pattern learned), Tool Call (external API), Read-only (analysis only).
- Context links: Every action links to the entity, the proposal, the reasoning trace.
- Outcome tracking: Did the action succeed? What was the result?
- Filter by type: Human can see only "executions" or only "proposals" or only "enrichments."
Difference from Audit Logs:
- Audit Logs = passive compliance record (what happened for legal reasons)
- AI Actions History = active operational record (what the AI did for operational understanding)
Panel 3: AI Memory Explorer
Purpose: What the AI knows, how it learned it, and how confident it is.
Layout:
┌──────────────────────────────────────────────────────────────┐
│ AI Memory Explorer [Search...] │
│ ─────────────────────────────────────────────────────────── │
│ │
│ ┌─────────────────────────────────────────────────────────┐│
│ │ [○] All Memories [○] Personal [○] Work [○] Org ││
│ │ [○] AI Patterns ││
│ └─────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Memory Graph (visual network) ││
│ │ ││
│ │ ┌──────────┐ ││
│ │ │ Acme Corp│ ││
│ │ │ (entity) │ ││
│ │ └────┬─────┘ ││
│ │ │ ││
│ │ ┌───────────┼───────────┐ ││
│ │ │ │ │ ││
│ │ ▼ ▼ ▼ ││
│ │ ┌────────┐ ┌────────┐ ┌────────┐ ││
│ │ │Champion│ │ Health │ │ Risk │ ││
│ │ │John S. │ │ Score │ │ Level │ ││
│ │ └────────┘ └────────┘ └────────┘ ││
│ │ ││
│ │ [Zoom In] [Zoom Out] [Reset View] ││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Memory List ││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Acme Corp's CEO prefers phone calls over email │││
│ │ │ Confidence: 95% │ Source: user input │ Permanent │││
│ │ │ Learned: 2026-01-15 │ Last reinforced: 2026-07-01 │││
│ │ │ [Reinforce] [Challenge] [Delete] │││
│ │ └────────────────────────────────────────────────────────┘││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Health drops for Acme correlate with unresolved bugs │││
│ │ │ Confidence: 89% │ Source: AI pattern │ Active │││
│ │ │ Learned: 2026-05-20 │ Repetitions: 7 │ Decay: 23d │││
│ │ │ [Reinforce] [Challenge] [Delete] │││
│ │ └────────────────────────────────────────────────────────┘││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Q1 churn pattern: health < 60 at renewal = 73% churn │││
│ │ │ Confidence: 92% │ Source: AI pattern │ Cold │││
│ │ │ Learned: 2026-04-01 │ Repetitions: 12 │ Decay: 67d │││
│ │ │ [Reinforce] [Challenge] [Delete] │││
│ │ └────────────────────────────────────────────────────────┘││
│ └──────────────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────────────┘
Key Features:
- Memory Graph: Visual network of entities, relationships, and knowledge nodes. Human can explore connections.
- Memory List: Text list with confidence, source, tier, decay status.
- Human actions: Reinforce (strengthen), Challenge (reduce confidence), Delete (remove).
- Source attribution: Every memory shows where it came from (user input, AI pattern, external data).
- Tier indicators: Active (hot), Warm, Cold, Permanent — color-coded.
- Decay visibility: Human sees when memories will fade, can choose to reinforce.
The Human Controls the AI's Memory:
- The human can delete memories the AI got wrong.
- The human can reinforce memories that are important.
- The human can challenge memories that seem questionable.
- The AI does not delete or modify its own memory without human oversight.
Panel 4: AI Workflow Monitor
Purpose: What AI workflows are running, scheduled, or completed.
Layout:
┌──────────────────────────────────────────────────────────────┐
│ AI Workflow Monitor [Filter ▼] │
│ ─────────────────────────────────────────────────────────── │
│ │
│ ┌─────────────────────────────────────────────────────────┐│
│ │ [○] Running [○] Scheduled [○] Completed [○] Failed ││
│ └─────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ RUNNING (2) ││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Daily Standup Summary │││
│ │ │ Trigger: Cron (daily 8:00 AM) │││
│ │ │ Status: ● Running │ Progress: 3/5 steps │││
│ │ │ Next: Send Slack summary to #standups │││
│ │ │ [Pause] [Stop] [View Details] │││
│ │ └────────────────────────────────────────────────────────┘││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Health Score Update │││
│ │ │ Trigger: Event (Salesforce update) │││
│ │ │ Status: ● Running │ Progress: 2/4 steps │││
│ │ │ Next: Write updated score to Spine │││
│ │ │ [Pause] [Stop] [View Details] │││
│ │ └────────────────────────────────────────────────────────┘││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ SCHEDULED (5) ││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Weekly Churn Analysis │││
│ │ │ Trigger: Cron (weekly, Monday 6:00 AM) │││
│ │ │ Next run: 2026-07-07 06:00 AM │││
│ │ │ [Run Now] [Edit] [Delete] │││
│ │ └────────────────────────────────────────────────────────┘││
│ └──────────────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────────────┘
Key Features:
- Running workflows: Real-time progress, step-by-step, human can pause/stop.
- Scheduled workflows: Cron or event-driven triggers, human can run now/edit/delete.
- Completed workflows: History with outcomes, success/failure, logs.
- Failed workflows: Error details, retry options, human investigation.
- Human control: Pause, stop, run now, edit, delete. The human is the operator.
Panel 5: AI Performance Dashboard
Purpose: How accurate is the AI? How often are its proposals accepted?
Layout:
┌──────────────────────────────────────────────────────────────┐
│ AI Performance Dashboard [Date ▼] │
│ ─────────────────────────────────────────────────────────── │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Proposal │ │ Accepted │ │ Rejected │ │ Accuracy │ │
│ │ 47 │ │ 38 │ │ 9 │ │ 81% │ │
│ │ this week│ │ 81% │ │ 19% │ │ ↑ 5% │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Acceptance Rate Over Time ││
│ │ ││
│ │ 100% │ ││
│ │ 80% │ ██ ││
│ │ 60% │ ██ ██ ││
│ │ 40% │ ██ ██ ██ ││
│ │ 20% │ ██ ██ ██ ██ ██ ██ ██ ██ ██ ││
│ │ 0% └────────────────────────────────────────────────── ││
│ │ W1 W2 W3 W4 W5 W6 W7 W8 W9 W10 W11 ││
│ │ ││
│ │ [View Details] ││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Top Performing Capabilities ││
│ │ ││
│ │ Capability │ Proposals │ Accepted │ Accuracy ││
│ │ ────────────────────┼───────────┼──────────┼──────────────││
│ │ Health Scoring │ 12 │ 11 │ 92% ↑ ││
│ │ QBR Scheduling │ 8 │ 7 │ 88% ↑ ││
│ │ Churn Prediction │ 15 │ 11 │ 73% ↓ ││
│ │ Email Drafting │ 12 │ 9 │ 75% → ││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Recent Failures ││
│ │ ││
│ │ ┌────────────────────────────────────────────────────────┐││
│ │ │ Churn prediction for Dataway was wrong (predicted │││
│ │ │ churn, but they renewed). Confidence was 89%. │││
│ │ │ [Investigate] [Retrain Model] [Adjust Threshold] │││
│ │ └────────────────────────────────────────────────────────┘││
│ └──────────────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────────────┘
Key Features:
- Summary cards: Proposals, accepted, rejected, accuracy.
- Trends: Acceptance rate over time, accuracy changes.
- Capability breakdown: Which AI capabilities perform well, which need tuning.
- Failure analysis: What went wrong, why, how to fix.
- Human actions: Investigate, retrain model, adjust threshold, disable capability.
Panel 6: AI Governance Settings
Purpose: Human controls AI thresholds, scope and review-priority rules.
Layout:
┌──────────────────────────────────────────────────────────────┐
│ AI Governance Settings │
│ ─────────────────────────────────────────────────────────── │
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Confidence Thresholds ││
│ │ ││
│ │ Auto-approve proposals with confidence ≥ [85%] [─●────] ││
│ │ Queue for review: 70% ─────────────────────── 85% ││
│ │ Discard proposals with confidence < [70%] [────●──] ││
│ │ ││
│ │ [Save Changes] ││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ AI Scope (What the AI Can Do) ││
│ │ ││
│ │ ☑ Read entity data ││
│ │ ☑ Propose actions ││
│ │ ☑ Send notifications ││
│ │ ☐ Execute actions without approval (DANGER) ││
│ │ ☑ Prepare entity enrichment proposal (approval required) ││
│ │ ☐ Access sensitive data (PII, financials) ││
│ │ ☐ Delete entity data ││
│ │ ││
│ │ [Save Changes] ││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Per-Capability Settings ││
│ │ ││
│ │ Capability │ Threshold │ Auto-Approve │ Enabled ││
│ │ ────────────────────┼───────────┼──────────────┼──────────││
│ │ Health Scoring │ 85% │ Yes │ ☑ ││
│ │ QBR Scheduling │ 85% │ Yes │ ☑ ││
│ │ Churn Prediction │ 80% │ No │ ☑ ││
│ │ Email Drafting │ 90% │ No │ ☑ ││
│ │ Escalation Routing │ 95% │ No │ ☑ ││
│ │ Budget Adjustment │ 99% │ No │ ☐ ││
│ │ ││
│ │ [Save Changes] ││
│ └──────────────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────────────┘
Key Features:
- Confidence sliders: Human adjusts thresholds. Real-time preview of what would happen.
- Scope toggles: What the AI can and cannot do. Explicit opt-in for dangerous actions.
- Per-capability settings: Different thresholds for different capabilities.
- Save changes: Changes are versioned, audited, reversible.
- Danger warnings: Red highlights for high-risk settings.
Panel 7: AI Model Management
Purpose: Which AI models are used, per-capability, cost tracking.
Layout:
┌──────────────────────────────────────────────────────────────┐
│ AI Model Management │
│ ─────────────────────────────────────────────────────────── │
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Default Model ││
│ │ ││
│ │ Model: GPT-4o Mini (via OpenRouter) ││
│ │ Provider: OpenRouter ││
│ │ Cost: $0.15 / 1M tokens ││
│ │ [Change Model] ││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Per-Capability Model Assignment ││
│ │ ││
│ │ Capability │ Model │ Cost (week) │ Status ││
│ │ ────────────────────┼──────────────┼─────────────┼────────││
│ │ Health Scoring │ GPT-4o Mini │ $0.02 │ ● ││
│ │ QBR Scheduling │ GPT-4o Mini │ $0.01 │ ● ││
│ │ Churn Prediction │ Claude 3.5 │ $0.12 │ ● ││
│ │ Email Drafting │ GPT-4o Mini │ $0.03 │ ● ││
│ │ Escalation Routing │ GPT-4o Mini │ $0.01 │ ● ││
│ │ ││
│ │ [Reassign] [View Usage] [Test Model] ││
│ └──────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────────────────────────────────────────────────────┐│
│ │ Model Usage This Month ││
│ │ ││
│ │ Total tokens: 2.3M ││
│ │ Total cost: $0.34 ││
│ │ ││
│ │ [View Breakdown] ││
│ └──────────────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────────────┘
Key Features:
- Default model: Change the global model.
- Per-capability assignment: Use cheap models for simple tasks, strong models for complex ones.
- Cost tracking: See how much each capability costs.
- Usage breakdown: Token usage, cost per capability, trends.
- Model testing: Test a model on a sample before assigning it.
Navigation & Access
How the Human Reaches the AI Workbench
┌─────────────────────────────────────────────────────────────────┐
│ [IntegrateWise Logo] [Workbench] [AI Workbench] [Settings] │
│ ▲ ↑ │
│ │ │ │
│ Default (human) Secondary (AI management) │
│ │
│ Clicking "AI Workbench" opens the AI management surface. │
│ Clicking "Workbench" returns to the human workbench. │
│ │
│ The AI Workbench is NEVER the default. The human must choose. │
└─────────────────────────────────────────────────────────────────┘
Sub-Navigation Within AI Workbench
┌─────────────────────────────────────────────────────────────────┐
│ [AI Workbench] │
│ [Proposals] [History] [Memory] [Workflows] [Performance] │
│ [Governance] [Models] │
│ │
│ ← Sub-navigation tabs, one per panel │
│ ← Default tab: Proposals Queue │
│ ← Human switches between panels │
└─────────────────────────────────────────────────────────────────┘
Human Control Model
Every action in the AI Workbench follows this model:
┌─────────────────────────────────────────────────────────────────┐
│ HUMAN CONTROL MODEL │
│ │
│ 1. OBSERVE → Human sees AI operations surfaced │
│ 2. REVIEW → Human evaluates proposals, memory, performance │
│ 3. DECIDE → Human approves, rejects, modifies, deletes │
│ 4. EXECUTE → Human triggers action (or confirms AI action) │
│ 5. VERIFY → Human checks outcome, provides feedback │
│ 6. ADJUST → Human tunes thresholds, scope, models │
│ │
│ The AI never executes without human confirmation. │
│ The AI never modifies its own settings. │
│ The AI never deletes its own memory without human approval. │
│ The AI never changes governance thresholds. │
│ │
│ The human is the operator. The AI is the tool. │
└─────────────────────────────────────────────────────────────────┘
Data Flow
┌─────────────────────────────────────────────────────────────────┐
│ AI WORKBENCH DATA FLOW │
│ │
│ Spine (source of truth) │
│ │ │
│ ├──→ Proposals Queue ←── action_proposals table (D1) │
│ │ │ │
│ │ └──→ Human reviews → approves/rejects/modifies │
│ │ │
│ ├──→ AI Actions History ←── spine_audit_log (D1) │
│ │ │ │
│ │ └──→ Human explores → filters → investigates │
│ │ │
│ ├──→ AI Memory Explorer ←── memory table (D1) │
│ │ │ │
│ │ └──→ Human reinforces/challenges/deletes │
│ │ │
│ ├──→ AI Workflow Monitor ←── workflow state (DO) │
│ │ │ │
│ │ └──→ Human pauses/stops/edits │
│ │ │
│ ├──→ AI Performance Dashboard ←── metrics (KV) │
│ │ │ │
│ │ └──→ Human tunes thresholds/models │
│ │ │
│ ├──→ AI Governance Settings ←── policy_registry (D1) │
│ │ │ │
│ │ └──→ Human adjusts scope/thresholds │
│ │ │
│ └──→ AI Model Management ←── model config (KV) │
│ │ │
│ └──→ Human reassigns models │
│ │
│ All writes go through the Pipeline (§5.2, §6.6). │
│ All changes are audited (§11.4). │
│ All changes require human authorization (§6.3). │
└─────────────────────────────────────────────────────────────────┘
States & Interactions
Empty State
┌─────────────────────────────────────────────────────────────────┐
│ Proposals Queue │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ │ │
│ │ 📭 No pending proposals │ │
│ │ │ │
│ │ The AI is watching your workbench and learning. │ │
│ │ When it has a suggestion, it will appear here. │ │
│ │ │ │
│ │ [Go to Workbench] │ │
│ │ │ │
│ └──────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Loading State
┌─────────────────────────────────────────────────────────────────┐
│ Proposals Queue │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ ┌──────────────┐ │ │
│ │ │ ██████░░░░░░ │ Loading proposals... │ │
│ │ └──────────────┘ │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ │ │
│ │ ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ │ │
│ │ ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ │ │
│ │ ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ │ │
│ │ ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ │ │
│ └──────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Error State
┌─────────────────────────────────────────────────────────────────┐
│ Proposals Queue │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ │ │
│ │ ⚠️ Could not load proposals │ │
│ │ │ │
│ │ The AI management service is temporarily unavailable. │ │
│ │ Your workbench and data are not affected. │ │
│ │ │ │
│ │ [Retry] [Contact Support] │ │
│ │ │ │
│ └──────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Accessibility
Screen Reader Experience
"AI Workbench. Proposals Queue. 3 pending proposals. First proposal:
Acme Corp health dropped to 62. Escalate to VP? Confidence 82 percent.
Actions: View Context, Approve, Modify, Dismiss."
[User presses Tab]
"View Context button."
[User presses Tab]
"Approve button."
[User presses Enter]
"Proposal approved. Acme Corp escalation queued for execution."
Keyboard Navigation
| Key | Action |
|---|---|
Tab | Navigate between proposals and actions |
Enter | Activate button (Approve, Reject, View) |
Shift + D | Dismiss selected proposal |
Shift + A | Approve selected proposal |
Shift + M | Modify selected proposal |
Shift + V | View context of selected proposal |
Esc | Close detail panel / return to list |
Cmd/Ctrl + Shift + A | Open AI Workbench from human workbench |
Cmd/Ctrl + Shift + W | Return to human workbench |
Focus Management
- When switching to AI Workbench, focus lands on first proposal card.
- When a proposal is approved, focus moves to next proposal.
- When detail panel opens, focus traps inside panel until closed.
- When empty state, focus lands on "Go to Workbench" button.
TypeScript Interfaces
// AI Workbench Panel
interface AIWorkbenchProps {
activePanel: AIWorkbenchPanel;
tenantId: string;
userId: string;
onClose: () => void; // Returns to human workbench
}
type AIWorkbenchPanel =
| "proposals"
| "history"
| "memory"
| "workflows"
| "performance"
| "governance"
| "models";
// Proposal Queue
interface ProposalQueueProps {
proposals: AIProposal[];
filter: ProposalFilter;
onApprove: (proposalId: string) => void;
onReject: (proposalId: string) => void;
onModify: (proposalId: string, modification: string) => void;
onDismiss: (proposalId: string) => void;
onViewContext: (proposalId: string) => void;
onBatchApprove: (proposalIds: string[]) => void;
onBatchDismiss: (proposalIds: string[]) => void;
}
type ProposalFilter = "all" | "pending" | "approved" | "rejected" | "executed";
interface AIProposal {
id: string;
entityId: string;
entityName: string;
capability: string;
description: string;
confidence: number; // 0.0–1.0
source: "twin" | "workflow" | "user";
status: "pending" | "approved" | "rejected" | "executed" | "expired";
createdAt: Date;
expiresAt: Date;
context: Entity360; // Full context for human review
reasoningTrace: ReasoningTrace;
humanAction?: HumanAction; // What the human did
}
// AI Actions History
interface AIActionsHistoryProps {
actions: AIAction[];
filter: ActionFilter;
onViewDetails: (actionId: string) => void;
onViewEntity360: (entityId: string) => void;
}
type ActionFilter =
| "all"
| "proposal"
| "execution"
| "enrichment"
| "memory"
| "tool_call"
| "read_only";
interface AIAction {
id: string;
type: ActionType;
entityId: string;
entityName: string;
capability: string;
description: string;
confidence: number;
source: "twin" | "workflow" | "user";
status: "success" | "failure" | "pending";
outcome?: string; // What happened
createdAt: Date;
humanApprover?: string; // Who approved it
reasoningTrace?: ReasoningTrace;
}
type ActionType =
| "proposal"
| "execution"
| "enrichment"
| "memory_update"
| "tool_call"
| "read_only";
// AI Memory Explorer
interface AIMemoryExplorerProps {
memories: AIMemory[];
filter: MemoryFilter;
onReinforce: (memoryId: string) => void;
onChallenge: (memoryId: string) => void;
onDelete: (memoryId: string) => void;
onViewGraph: (entityId: string) => void;
}
type MemoryFilter = "all" | "personal" | "work" | "organization" | "ai_pattern";
interface AIMemory {
id: string;
entityId: string;
entityName: string;
content: string;
confidence: number;
source: "user_input" | "ai_pattern" | "external_data";
tier: "active" | "warm" | "cold" | "permanent";
repetitions: number;
decayDays: number; // Days until decay
createdAt: Date;
lastReinforcedAt: Date;
relationships: EntityRelationship[];
}
// AI Workflow Monitor
interface AIWorkflowMonitorProps {
workflows: AIWorkflow[];
filter: WorkflowFilter;
onPause: (workflowId: string) => void;
onStop: (workflowId: string) => void;
onRunNow: (workflowId: string) => void;
onEdit: (workflowId: string) => void;
onDelete: (workflowId: string) => void;
onViewDetails: (workflowId: string) => void;
}
type WorkflowFilter = "running" | "scheduled" | "completed" | "failed";
interface AIWorkflow {
id: string;
name: string;
description: string;
trigger: WorkflowTrigger;
status: "running" | "scheduled" | "completed" | "failed" | "paused";
progress?: { current: number; total: number }; // For running workflows
nextRunAt?: Date; // For scheduled workflows
lastRunAt?: Date;
lastRunOutcome?: "success" | "failure";
steps: WorkflowStep[];
createdBy: string;
createdAt: Date;
}
interface WorkflowTrigger {
type: "cron" | "event" | "manual" | "ai_proposed";
config: string; // Cron expression or event filter
}
interface WorkflowStep {
id: string;
name: string;
status: "pending" | "running" | "completed" | "failed";
output?: string;
error?: string;
}
// AI Performance Dashboard
interface AIPerformanceDashboardProps {
metrics: AIPerformanceMetrics;
dateRange: DateRange;
onViewDetails: (capability: string) => void;
onInvestigate: (failureId: string) => void;
onRetrainModel: (capability: string) => void;
onAdjustThreshold: (capability: string) => void;
}
interface AIPerformanceMetrics {
totalProposals: number;
acceptedProposals: number;
rejectedProposals: number;
accuracy: number; // 0.0–1.0
acceptanceRate: number; // 0.0–1.0
trend: "up" | "down" | "stable";
capabilityBreakdown: CapabilityMetric[];
recentFailures: AIFailure[];
}
interface CapabilityMetric {
capability: string;
totalProposals: number;
acceptedProposals: number;
accuracy: number;
trend: "up" | "down" | "stable";
}
interface AIFailure {
id: string;
capability: string;
description: string;
expected: string;
actual: string;
confidence: number;
createdAt: Date;
}
// AI Governance Settings
interface AIGovernanceSettingsProps {
settings: AIGovernanceConfig;
onSave: (settings: AIGovernanceConfig) => void;
onPreview: (settings: AIGovernanceConfig) => void;
}
interface AIGovernanceConfig {
autoApproveThreshold: number; // 0.0–1.0
queueThreshold: number; // 0.0–1.0
discardThreshold: number; // 0.0–1.0
scope: AIAllowedScope;
perCapability: Record<string, CapabilityGovernance>;
}
interface AIAllowedScope {
readEntity: boolean;
proposeAction: boolean;
sendNotification: boolean;
executeWithoutApproval: boolean;
enrichData: boolean;
accessSensitiveData: boolean;
deleteEntity: boolean;
}
interface CapabilityGovernance {
threshold: number;
autoApprove: boolean;
enabled: boolean;
}
// AI Model Management
interface AIModelManagementProps {
defaultModel: AIModelConfig;
perCapabilityModels: Record<string, AIModelConfig>;
usage: AIModelUsage;
onChangeDefaultModel: (model: AIModelConfig) => void;
onReassignModel: (capability: string, model: AIModelConfig) => void;
onTestModel: (model: AIModelConfig, sample: string) => void;
}
interface AIModelConfig {
model: string; // e.g., "gpt-4o-mini"
provider: string; // e.g., "openrouter"
costPerMillionTokens: number;
}
interface AIModelUsage {
totalTokens: number;
totalCost: number;
breakdown: Record<string, number>; // capability → cost
}
Appendix: Difference from Audit Logs
| Aspect | AI Workbench | Audit Logs |
|---|---|---|
| Purpose | Human manages AI operations | Compliance and legal record |
| Audience | Human operators, AI managers | Auditors, compliance officers, legal |
| Content | Proposals, memory, workflows, performance | All system events, data access, changes |
| Interaction | Active — human takes action | Passive — human reviews |
| UI | Rich interactive panels | Simple timeline/filter table |
| Actions | Approve, reject, modify, delete, adjust | Read, export, search |
| Mutability | Human can change AI behavior | Immutable, append-only |
| Scope | AI operations only | All operations (human + AI) |
| Location | Secondary tab in app | Admin section or external tool |
| Default view | Proposals Queue | Full event timeline |
| Human control | High — human adjusts AI settings | None — read-only compliance |
The AI Workbench is the human's control panel for the AI.
The Audit Logs are the legal record of everything that happened.
The Human-First AI Workbench Philosophy
┌─────────────────────────────────────────────────────────────────┐
│ THE HUMAN-FIRST AI WORKBENCH │
│ │
│ The AI Workbench is not where the AI lives. │
│ It is where the human controls the AI. │
│ │
│ The AI lives in the Capability Layer, watching the Spine. │
│ The human lives in the Workbench, doing their work. │
│ When the human wants to manage the AI, they open the AI │
│ Workbench — a separate, organized, human-controlled surface. │
│ │
│ The AI does not own this interface. │
│ The human does. │
│ The AI does not use this interface. │
│ The human uses it to manage the AI. │
│ │
│ Every AI proposal is a request, not a command. │
│ Every AI memory is a suggestion, not a fact. │
│ Every AI workflow is a tool, not an autonomous agent. │
│ Every AI setting is a human decision, not a system default. │
│ │
│ The human is the operator. │
│ The AI is the tool. │
│ The AI Workbench is the human's control panel. │
│ The Workbench is the human's workspace. │
│ │
│ Two separate surfaces. │
│ One human protagonist. │
│ One AI supporting character. │
│ │
│ That is the human-first AI Workbench. │
└─────────────────────────────────────────────────────────────────┘
END OF DOCUMENT
IntegrateWise AI Workbench Specification v1.0.0. Human-first. AI-managed. Separated. Organized.