Agent Autonomy Model
Autonomy is default; intervention is selective and risk-driven
Agent trust is not a soft metric — it directly determines whether reason tags are honest, whether overrides are reported, and whether the learning loop gets clean signal. Introducing a deviation classification system will trigger surveillance fear if rolled out without deliberate change management.
Autonomy Bands
Three tiers of agent freedom — friction concentrated only where risk justifies it.
Corrective overrides flow with minimal friction. Tier L actions require zero extra clicks.
Medium-risk deviations require reason tags — structured dropdown adding ~5 seconds of handling time.
Targeted supervisor friction on risk-sensitive actions. Async approval — agent not blocked from other case work.
Anti-Surveillance Safeguards
Five guardrails that prevent the system from becoming a surveillance tool.
Show agents why a deviation was flagged — risk rationale visible in the copilot panel.
Inline appeal button for flagged decisions, logged for calibration review.
Use signals for coaching and quality improvement before any punitive actions.
Agent deviation records are accessible to the agent — right of access per privacy framework.
Statistical monitoring visible to ops managers only, not used for individual ranking without separate HR process.
Phased Adoption Model
Four phases that build trust before adding friction — shadow first, enforce last.
Shadow Mode
4 weeks before pilot- Copilot panel active with suggestions, no enforcement
- Deviation classifier runs in background — logged but not shown
- Collect baseline data and validate classifier accuracy
"New suggestion tool" — no mention of deviation tracking
Transparency Mode
First 2 weeks of pilot- Agents see the deviation status indicator (green/yellow/orange/red)
- Reason tags available but optional (encouraged, not required)
- Agents shown what would have been flagged, no friction applied
"We're testing a system that helps identify when our suggestions are wrong so we can improve them."
Soft Launch
Weeks 3–6 of pilot- Reason tags required for Tier M overrides (<5 seconds)
- Tier H cases flagged for review — agent not blocked
- Weekly feedback sessions with pilot agents (15 min)
"We're now using your input to catch risky situations earlier and to make sure good judgment gets recognized."
Active Mode
Weeks 7+ of pilot- Full selective friction: Tier H requires supervisor approval
- Appeal path available for misclassified flags
- Gamified reinforcement — badges for confirmed good judgment
- Monthly agent advisory panel (5 rotating agents)
"The system is live. Most overrides flow through with no friction. When flagged, here's why and how to appeal."
Communication Principles
How we frame the system to agents — trust is earned through language and action.
Lead with agent benefit
"This system exists to recognize good judgment, not to police you."
Be honest about tracking
Agents will discover it anyway. Transparency builds trust; secrecy destroys it.
Show early wins
Within 2 weeks of active mode, publish corrective overrides that improved suggestions — with agent credit.
Never punitive without process
Coaching first, always. HR escalation only after repeated, confirmed violations post-coaching.
Coaching Model
Escalation ladder — coaching through 1:1 sessions, not automated warnings.
Feedback Mechanisms
Four channels ensuring agent voice shapes how the system evolves.
"Did the copilot help today? Was any flagging unfair? What should we fix?"
Deeper discussion on UX, trust, and suggestion quality.
Reviews threshold changes, new rules, and gamification design before deployment.
Thumbs up/down with optional comment — feeds suggestion quality metrics.
Adoption Success Criteria
Five measurable gates that define pilot success.
agents using suggestions within 60 days
trust score by end of pilot
reason-tag completion for Tier M
appeal rate of Tier H flags
attrition attributed to copilot