What to show the board about your agents

When a board member, auditor, or regulator asks about your AI agents, you need more than reassurance. You need a working answer — specific, evidenced, and on record. This checklist covers the five questions they will ask and what evidence you need to answer each one.

Question 1 — What are your AI agents doing?

The board wants to know that you have a complete picture of agent activity — not just a list of tools, but a record of what agents are actually doing in production.

Evidence you should have ready

• A list of every active agent process by name, with the business function it supportse.g. "Invoice reconciliation — Finance Ops · Billing agent · Active since 2026-07-01"
• The execType for each agent task — AGENT, AGENT_SUPERVISED, or AGENT_REVIEW
• Number of agent actions taken in the last 30 days, by process
• The percentage of actions that ran autonomously vs required human review

What good looks like

“In the last 30 days, our agents completed 4,847 actions across 6 governed processes. 71% ran autonomously within defined thresholds. 29% escalated to a named reviewer.”

Question 2 — Who approved what, and when?

This is the accountability question. The board needs to know that every consequential agent action has a named human behind it — not just that someone approved it, but who, when, and based on what information.

Evidence you should have ready

• Audit trail export for any process, any date range, on demand
• For each approval: reviewer identity, timestamp, what was presented to them, time spent reviewing
• For each rejection: reviewer identity, reason, and what happened next
• Escalation chain for any decision that was delegated

What good looks like

“Here is the approval record for the October month-end close. Every decision is logged with the reviewer's name, the timestamp, the agent's output, and the confidence score at the time of review. The CFO signed off on the final submission at 09:14 on November 3rd.”

Question 3 — What policies govern your agents?

The board wants to know that agent behaviour is not arbitrary — that there are defined rules, that those rules are documented, and that they can be shown to a regulator.

Evidence you should have ready

• Current policy version for each process, with effective date
• Policy change history — what changed, who changed it, when, and why
• The threshold for each policy — what auto-approves and what escalates
• Named reviewers and backup reviewers for each escalation path

What good looks like

“Our refund approval policy is currently at version 3. It was updated on September 15th — the threshold was raised from $500 to $750 based on 30 days of escalation data showing 83% of escalations in the $500-750 range were approved without modification. The change was ratified by the Finance Controller.”

Question 4 — What happens when something goes wrong?

This is the resilience question. The board wants to know that the system does not just run — it recovers, it escalates appropriately, and it does not let a bad agent action propagate downstream.

Evidence you should have ready

• The escalation path for every process — what happens at each failure mode
• SLA breach history — when did processes miss their windows, what fired, what happened
• A record of any blocked actions — agent attempted an irreversible action and was stopped
• MTTD and MTTR for the last quarter — how fast were issues detected and resolved

What good looks like

“In Q3, we had 14 SLA breaches across all processes. 12 were detected and escalated automatically before any human noticed. The average time to detection was 4 minutes. The two that required manual intervention were resolved within 22 minutes of the escalation reaching the ops lead.”

Question 5 — How are you improving over time?

The board wants to know that the governance system learns — that it is not static, that thresholds adjust based on evidence, and that human oversight becomes more targeted over time, not more burdensome.

Evidence you should have ready

• Threshold change history with rationale — every adjustment and the data that drove it
• Trend: is the autonomous action rate increasing as trust is earned?
• Trend: is the escalation approval rate stable, rising, or falling?
• Reviewer feedback summary — where are humans adding the most value?

What good looks like

“In July, 58% of agent actions ran autonomously. By September, that was 71% — not because we lowered our standards, but because the agent demonstrated consistent performance and we raised thresholds with evidence. Every change is on record. The human review that remains is more meaningful, not less.”

Quick reference — evidence map