← Research library

Policy reviews for agent programs: a quarterly governance checklist

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Agent programs in marketing now touch budgets, customer data, and public claims, and the rules around them were mostly written at launch. Those rules age faster than the agents do. Governance is easy to set up and easy to neglect.

Permissions granted for a project stay in place after it ends. Autonomy levels set in spring are never revisited. Guardrail lists fall behind when products, prices, or policies change. Each gap is small on its own, and together they create real risk.

In our research, we treat governance as a recurring maintenance task with a fixed slot on the calendar. This post sets out the quarterly review we recommend, with a checklist you can copy.

Why governance decays

Governance drifts for ordinary reasons. Agents change faster than the documents that describe them. A team adds a tool to meet a deadline and means to scope it down later. An owner changes roles, and the agent keeps running under a name that no longer answers messages.

None of this is negligence. It is what happens when a program grows and the people running it are busy shipping campaigns. The fix is a short, regular review that compares what the rules say with what the agents actually do.

The comparison is the point. A policy that says "agents may only read customer records" is only as good as the permission settings that enforce it. The review checks the settings, the logs, and the people, not the policy text alone.

How to run the review

We recommend scheduling the review in the last month of each quarter and planning for about two hours for a program with several agents. Larger programs can split the work across owners and meet to discuss findings.

The review works best with a small, fixed group. Bring the program owner, one agent owner who rotates each quarter, and someone from legal or compliance for the policy section. Prepare the inputs in advance so the meeting is spent on judgment rather than on gathering data.

  • Agent inventory: the current list of agents, owners, and autonomy levels.
  • Permission export: the tools and data each agent can reach today.
  • Usage logs: which permissions each agent actually used this quarter.
  • Guardrail triggers: the log of blocked or escalated actions.
  • Incident notes: anything that went wrong and what was done about it.

Inventory and permissions

Start with what exists. A review that misses an agent misses everything that agent can do, so the inventory comes first.

Then compare each agent's granted permissions against what it used. Unused permissions are the cheapest risk to remove, and they tend to accumulate quietly around one-off projects and integrations. Our post on roles and permissions for a mixed human and agent team covers how to scope them in the first place.

  • Each active agent is listed with its owner, purpose, and autonomy level
  • Agents no longer in use have been retired and their access removed
  • New agents added this quarter went through the graduation process
  • Each agent's permissions match what it actually used this quarter
  • Temporary permissions have expired or been formally renewed
  • No agent runs on a person's credentials
  • High-risk permissions were approved by someone other than the agent owner

Autonomy and guardrails

Autonomy levels should follow evidence. An agent that has run cleanly for a quarter may be ready to move up a level, and one that has exceeded its failure budget should move down or be fixed. The autonomy ladder diagram shows the levels we use.

Guardrails need the same attention. Banned terms, required disclosures, and claim rules all depend on current products and policies. The trigger log is the richest input here: repeated triggers on the same rule suggest either a weak instruction or a rule that no longer fits the work.

  • Each agent's autonomy level is still justified by recent evidence
  • Agents that exceeded their failure budget have been moved down or fixed
  • Proposed promotions are documented with evidence
  • Banned terms, required disclosures, and claim rules reflect current products and policies
  • Spending and volume limits match current budgets
  • Guardrail trigger logs were reviewed for patterns

Records and policies

Records prove that governance happened. We recommend tracing three random customer-facing outputs end to end, from the request through each agent action and approval to the published result. If a trace breaks, the audit trail has a gap worth fixing before it matters. See audit trails that satisfy legal, brand, and finance.

Policies on disclosure, consent, and data use change with regulation and platform rules. The review should confirm that current practice matches current policy, and that consent rules are enforced in code rather than in instructions alone. Confirm the specifics of any regulatory requirement with counsel.

  • Three random outputs were traced end to end through the audit trail
  • Log retention follows policy
  • Incidents this quarter were reviewed and their follow-up actions completed
  • The disclosure policy reflects current practice and requirements
  • Consent and data use rules are enforced in code for each customer-facing agent
  • Relevant regulatory or platform changes have been reviewed with legal

People

Governance depends on people who know their role. Owners leave, reviewers change, and pause procedures get forgotten when nobody has used them in months. This section is short, and we suggest never skipping it.

  • Each agent has a current owner and a named backup
  • Everyone who can pause an agent knows how, and one person has tried it this quarter
  • Reviewers have been recalibrated if rubrics changed

After the review

Record what you found, what you changed, and what remains open, with an owner and a date for each open item. Share a short summary with leadership. Over several quarters, these records show that the program is governed deliberately, which matters for internal confidence and for any external questions.

We suggest tracking a few measures across reviews: open items carried from the last quarter, unused permissions removed, and the share of traced outputs with a complete trail. A rising carry-over count is an early sign that the review has become a formality.

We are still studying how much of this review can run continuously. Permission drift and stale owners are easy to detect automatically. Whether autonomy decisions can be partly automated without weakening human judgment is an open question we plan to test next.

How to apply this: schedule a two-hour governance review in the last month of each quarter. Prepare the inputs in advance, work through this checklist, and give each open item an owner and a date.

Related research

Human Oversight

Holiday content at scale without brand drift

Holiday volume tempts teams to loosen review, and the result is often brand drift: content that slowly stops sounding like you. A few safeguards prevent it.

5 min read

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.