AI Governance Report Q3 2026

Executive Summary

Quarter summary

Agent-assisted processes handled 18,472 actions in Q3 2026. 71.8% executed autonomously within ratified policy thresholds — up from 65.2% in Q2, reflecting threshold adjustments ratified after Q2's outcome review. Every threshold change is versioned and on record. The 28.2% of actions that escalated reached a named human reviewer with full context. No action bypassed a governance gate. Every decision is logged with the reviewer's identity, timestamp, engagement evidence, and the policy version in force at decision time.

1. What are our AI agents doing?

Active governed processes

Agent action breakdown by execType

2. Who approved what, and when?

Decision summary — Q3 2026

Every approval in Q3 carried a full audit record: reviewer identity, timestamp, context presented, engagement time, and policy version in force. No decision settled without a complete record.

Sample decision record — representative example

decision_id:   DEC-20260912-04721
run_id:        MEC-20260912T044343Z
process:       Month-End Close
task:          NAV Controller Sign-Off
actor:         agent.recon.v3 (proposed) → r.okafor (decided)
policy:        NAV_SIGNOFF v4 (effective 2026-08-01)
engagement:    47 seconds · sections viewed: 4/4
outcome:       APPROVED at 2026-09-12T06:14:22Z
record_hash:   sha256:a4f2c8...

The engagement field confirms the reviewer actively engaged with the context — 47 seconds across 4 of 4 sections — before the approve button became available. This is the evidence of comprehension required by AI governance frameworks including EU AI Act Article 14.

3. What policies govern our agents?

Active policies — Q3 2026

Policy changes in Q3

4. What happened when things went wrong?

Incidents and SLA breaches — Q3 2026

Q3 SLA performance

5. How are we improving over time?

Autonomous action rate trend

The autonomous rate has grown from 58.4% to 71.8% over three quarters — not by lowering governance standards, but by raising thresholds where outcome data demonstrated the agent was consistently right. Each change is ratified by a named authority and versioned. Human review is becoming more targeted, not less rigorous.

Q3 improvement actions ratified

• INVOICE_AUTO threshold raised from $250 to $500 — ratified by Finance Controller on 2026-07-15. Basis: Q2 data showing 87% of $250-500 escalations approved without modification.
• COUNTERPARTY_LIMIT threshold raised from $1M to $2M — ratified by Risk Manager on 2026-09-01. Basis: Q2 data showing 91% approval rate at unchanged terms in that band.
• VENDOR_ONBOARD confidence threshold lowered from 0.85 to 0.80 — ratified by COO on 2026-07-01. Basis: Q2 data showing no quality difference between 0.80-0.85 confidence band outcomes.

Pending improvements — Q4 consideration

• INVOICE_AUTO threshold review — current 82% escalation approval rate suggests further upward adjustment may be warranted. Recommendation: review after 30 additional days of data at current threshold.
• Backup reviewer coverage — Q3 had 2 SLA breaches caused by reviewer unavailability. Recommendation: require backup reviewer assignment on all AGENT_REVIEW policies before Q4.

Board attestation

This report documents that during Q3 2026:

• All AI agent actions were governed by versioned, effective-dated policies ratified by named human authorities.
• Every consequential AI output that exceeded a governance threshold was reviewed by a named human before being acted upon.
• The review record includes comprehension evidence — the reviewer engaged with the context before the approve action became available.
• All policy changes were ratified by named authorities and are permanently versioned.
• No agent action bypassed a governance gate.
• The complete audit trail is available for regulatory examination on request.