Nobody has the list
Agents live in ChatGPT, Claude, Copilot Studio and a few side projects. No one holds the full list, so no one can say which touch customer data.
Governance & control
Most companies with an AI policy still cannot say which agents are running, who owns each one, and what each one can touch. A policy is written once. Governance is a registry, a few rules and a one-hour review every month that keep those answers true.
The gap between a document and control
Agents live in ChatGPT, Claude, Copilot Studio and a few side projects. No one holds the full list, so no one can say which touch customer data.
A tool was reviewed at launch. Since then it gained a connector, a newer model and write access to the CRM, and nobody looked again.
The newsletter drafter and the refund agent go through the same process. Either light work drowns in paperwork or risky work slips through.
An agent sends the wrong email or posts a wrong number. There is no log to read and no one whose job it is to fix it.
Where the line sits
Your leadership decides how much risk is acceptable and who approves what. We build the registry, gates, logs and review rhythm so those decisions still hold in six months. The document itself is the AI usage policy; this page is about running it.
What we don’t touch: legal interpretation stays with your counsel, certification with your auditor.
How it gets built
We list every agent, custom GPT, Claude Project, Copilot Studio agent and connector in use, and give each a named owner. No owner, no agent.
In practice. Admin exports from ChatGPT, Claude and Microsoft 365 plus a short team-lead survey → AI groups them by system, data and actions, and flags duplicates → a draft registry in a shared sheet or Jira → each team lead confirms their own rows.
Tier 1 drafts for a person. Tier 2 reads internal systems. Tier 3 writes to systems, moves money, contacts customers or affects decisions about people, and every irreversible Tier 3 action waits for a human.
In practice. An agent prepares a credit in Zendesk or a payment batch in NetSuite → AI attaches the source records → the request waits in Slack or Teams → a named approver accepts or rejects it, and both are logged.
Tier 2 and 3 agents log what they read, what they did and who approved it. Each vendor and connector gets a one-page review before it is connected. Incidents get a short playbook.
In practice. A new tool or connector request lands in Jira → AI pulls the vendor’s security page, data terms and training-data position into a checklist → a one-page vendor review → IT or security signs off.
One hour a month with the agent owners. What is in use, what broke, what to retire. Most months it is short, which is the point.
In practice. Admin-console usage data and the incident log → AI prepares a one-page summary of new, unused and failing agents → a list of decisions → leadership approves retirements and new Tier 3 requests.
Agents reach internal systems through MCP connectors scoped to the task, so access is narrow and every call can be logged. More on our security page.
The template
One row per agent or use case. A row nobody can fill in is the first finding.
| Field | Why it matters | Example |
|---|---|---|
| Name and purpose | You cannot govern what you cannot name | Inbound lead triage |
| Owner | One person answers for it | Head of sales operations |
| Platform | Admin controls differ by tool | ChatGPT workspace agent |
| Data it reads | Defines the exposure | HubSpot contacts, Gmail threads |
| Actions it takes | Writing carries more risk than reading | Updates CRM fields, drafts emails |
| Risk tier | Sets the review path | Tier 2 |
| Human checkpoint | Where a person signs off | Rep approves before any send |
| Where the logs live | First place to look when it breaks | Workspace audit log |
| Vendor review | The provider was actually checked | Done, renewed yearly |
| Last and next review | Approvals go stale | Reviewed this month |
| Status | Retired agents keep their history | Live, pilot or retired |
Frameworks
The routine lines up with the three frameworks auditors and customers ask about most. Lining up is not the same as complying.
A voluntary US framework from NIST with four functions, Govern, Map, Measure and Manage. Registry and owners map to Govern, risk tiers to Map, the monthly metrics to Measure, approvals and incidents to Manage.
The international standard for an AI management system, which organizations can be certified against. This routine produces the records it expects, such as roles, risk assessment and periodic review. Certification is a separate audit.
A risk-based regulation. A few practices are banned, high-risk systems carry the heaviest duties, some carry transparency duties, and most everyday uses sit lowest. Obligations phase in over several years. Whether it applies to you is for your counsel; the registry gives them the facts.
The offer
Usually part of an AI implementation sprint: the registry, tiers and gates go in during the sprint, and the first monthly review runs before it ends. With AI Partner, every new agent our engineers build enters the registry the day it ships.
On our AI training and engineering enablement engagement, an AI Acceptable Use Policy and responsible-use guardrails sat next to measured adoption, and the governance satisfied both legal and the PE owners. Metrics follow our AI ROI measurement method.
When this comes up
If two of these sound familiar, it’s worth a call:
FAQ
The policy is the document: what data can go into which tool and how output is checked. Governance is the routine that keeps it true: the registry, tiers, approvals, logs and monthly review. Most companies arrive here once their agents outgrow anyone’s memory.
That is the real risk, which is why the path is tiered. Most uses are Tier 1 and need only a row in the registry. The heavy path is for agents that write to systems, move money, contact customers or affect people. Fast answers on low-risk requests stop people working around it.
No, you do. We bring the AI and the routine; your people bring the judgment about what is risky in your business. Your leadership sets the tiers and approval rules, and we make sure they are applied the same way every month.
No routine does that on its own. Legal interpretation stays with your counsel and certification with an accredited auditor. We build the operational evidence both will ask for: who owns what, how risk was assessed, what was approved, and what happened when something went wrong.
Very little. The registry describes systems and data categories, not the data itself. We work from admin views inside your tenant, logs stay where your platforms keep them, and agents reach internal systems only through scoped connectors.
One accountable executive, often the COO, CIO or head of operations, plus a named owner per agent. Security and legal join the monthly review when a Tier 3 request or an incident is on the agenda.
Where this leads
A short call, no deck. We will look at what is running, what is already covered, and say honestly whether we can add anything.