AI governance

Your AI Agents Are Outgrowing Their Controls: 5 Governance Steps Before You Deploy

8 minute readFresh Mango AI
Business leaders and technology professionals reviewing an AI agent governance plan in a bright meeting room

AI agents are moving from pilots into production, often faster than the organisation's controls are developing.

That matters because an agent is not always just a chat interface. Some agents answer questions from approved sources. Others can call tools, update records, route requests or trigger workflows. The second type can create operational consequences, not just an inaccurate answer.

A human employee's access is usually reviewed when they join, change roles or leave. An agent can keep running unattended while its access, integrations or purpose quietly expand. If nobody owns that lifecycle, the organisation may not know what the agent can reach or whether it is still needed.

The governance response is practical, not prohibitive. Treat every production agent as a managed business capability with an identity, purpose, owner, controlled permissions, evidence and an exit plan.

First, distinguish an agent that answers from one that acts

An agent that answers from documents can still expose confidential or outdated information. “It only reads documents” is not a complete risk assessment: information may be sensitive, access may be too broad, and an answer may influence a customer, employee or business decision.

An agent that acts adds further questions. Can it create or amend records? Send messages? Approve requests? Delete content? Trigger a downstream process? Each action needs an appropriate authorisation boundary and, where the impact warrants it, a human approval step.

The relevant control depends on what the agent can do, what it can access, who uses it and what happens if it makes an error. Start there.

1. Give every production agent its own identity and lifecycle

Purpose: make agent activity attributable and access revocable.

Why it matters

If an agent operates through a shared service account, a developer's credentials or a departed employee's account, it becomes difficult to distinguish the agent's activity from a person's. Disabling the wrong account may also interrupt other work, while leaving the agent's access in place.

Where the platform supports it, Microsoft Entra Agent ID can provide agents with their own lifecycle-managed identities. This allows identity controls, including Conditional Access, custom security attributes and scoped role-based access control (RBAC), to be applied to the agent identity rather than relying solely on the identity of whoever built it. Check the platform's capabilities and current configuration; not every agent product uses the same identity model.

Microsoft's least-privilege guidance for AI agents (opens in a new tab) recommends dedicated identities, accountable owners, scoped access and tested revocation.

What good looks like

  • Each production agent has a unique, documented identity.
  • A named human sponsor is accountable for its business lifecycle.
  • Technical ownership and incident response responsibility are assigned.
  • Credentials, tokens and permissions can be revoked through a tested process.
  • No agent depends on a former employee's account or an unexplained shared identity.

Measurable outcome

Track the proportion of production agents with a unique identity, an accountable sponsor and a tested disablement process. Record how long it takes to suspend an agent and revoke its access.

2. Scope the knowledge and tools the agent can reach

Purpose: limit retrieval and actions to the minimum required for the approved task.

Technology and operations professionals reviewing an AI agent access checklist at a laptop

Why it matters

An agent's effective access is not defined only by what appears in its user interface. It also depends on its identity, connected tools, data sources, downstream systems and permissions at each point in a workflow.

For a read-only agent, broad access can still expose confidential information or combine information from sources that should remain separate. For an agent with write access, excessive permissions can allow unintended changes. Microsoft's guidance on securing and governing Microsoft 365 Copilot agents (opens in a new tab) covers controls for grounding data and interactions.

For SharePoint agents, Microsoft explains that users' responses depend on their permissions to the agent's underlying data sources in its guide to managing access to agents in SharePoint (opens in a new tab).

What good looks like

  • Approved knowledge sources have owners and current permissions.
  • Retrieval is restricted to the sites, collections or records needed for the use case.
  • Tools and actions are allowlisted; unreviewed integrations are not enabled by default.
  • High-impact permissions are separate from routine read access.
  • Permissions are checked end to end, including downstream systems.

Measurable outcome

Record which sources and tools each agent can reach, the percentage of access reviewed, and any broad or unnecessary permissions removed. Test representative user access before release.

3. Put a named owner and approved use case behind every agent

Purpose: establish who is accountable for the agent's purpose, scope and continued use.

AI governance workshop participants discussing policy and controls around a boardroom table

Why it matters

Agents can appear through departmental experiments, embedded platform features and custom builds. Without a register and approval route, teams may struggle to answer basic questions: what is this agent for, who approved it, which information does it use, and who will respond if it causes a problem?

A business owner should be accountable for whether the agent remains useful and appropriate. A technical owner should manage configuration, access and incident response. These roles may sit with the same person in a small business, but the responsibilities still need to be explicit.

What good looks like

Maintain an agent register with, at minimum:

  • Name, purpose, business owner and technical contact
  • Approved users, data sources, tools and permissions
  • Risk level and required approvals
  • Deployment environment, review date and retirement criteria
  • Dependencies, escalation route and known limitations

Approval should happen before production access is granted, not after an agent becomes part of a business process.

Measurable outcome

Track the percentage of production agents in the register with an approved use case, named owners, a review date and documented data sources. Investigate unregistered agents rather than assuming that low visibility means low risk.

4. Set oversight for the actions an agent takes

Purpose: make important actions visible, reviewable and recoverable.

Why it matters

An agent can generate an answer without changing a system. It can also create a purchase request, update a customer record or send a message. Those are different risk profiles.

Decide which actions can run automatically, which require confirmation and which the agent must not perform. For actions that could affect customers, employees, money, legal commitments or operational continuity, approval and a clear escalation path are usually appropriate. Logging should show what the agent did, which identity it used, what resource it changed and whether a person approved the action.

What good looks like

  • Low-impact, reversible actions are clearly distinguished from high-impact actions.
  • Approval gates apply to defined sensitive or consequential actions.
  • Logs record tool calls and outcomes, not only the agent's final response.
  • Owners know how to stop the agent and reverse or correct a change where possible.
  • Testing includes failed, ambiguous and out-of-scope requests.

Measurable outcome

Measure the proportion of material actions covered by logging, approval and recovery procedures. Test a stop or rollback route before release and record any actions that cannot be reversed.

5. Review the agent estate, and retire agents deliberately

Purpose: prevent access, integrations and unattended workloads from persisting after their original need has passed.

Leadership team reviewing business performance and governance information together

Why it matters

A pilot may end while its agent, permissions and credentials remain active. A team may change, source content may move or a connected system may add new capabilities. An agent that once had a narrow purpose can gradually become an unmanaged route into business information or processes.

Review the estate on a defined cadence: monthly during a pilot and, once stable, often quarterly, plus after material changes to the agent, its data sources or its tools.

What good looks like

  • Owners confirm that each agent is still needed and its scope remains appropriate.
  • Usage, incidents, access changes and business value are reviewed together.
  • Material changes trigger a reassessment before new capabilities go live.
  • Retirement includes disabling the agent, revoking tokens and credentials, removing permissions and connectors, and preserving required records.
  • The register records the retirement decision and completion.

Measurable outcome

Track review completion, unresolved findings, access removed, inactive agents retired and the time between a retirement decision and full revocation. Measure outcomes against the original use case, not simply the number of interactions.

Treat Restricted Content Discovery as a bridge, not a destination

Microsoft blocked new enablement of Restricted SharePoint Search from 31 July 2026 and identifies it as a short-term measure, not a long-term security boundary. Organisations using it should plan their next steps rather than build a new governance model around it. See Microsoft's Restricted SharePoint Search guidance (opens in a new tab).

Microsoft describes Restricted Content Discovery (opens in a new tab) (RCD) as a current interim, site-level control. It can limit selected SharePoint content from appearing in organisation-wide search and Copilot responses while permissions are reviewed.

RCD does not change permissions or prevent people who already have access from opening content directly. It applies to SharePoint sites, not OneDrive, and changes can take time to propagate. Treat it as a temporary discovery control while you review and right-size access, not as proof that information is secure. Where broader access restriction is needed, assess the appropriate permissions and data controls for your environment.

What to do before deployment

Use a controlled sequence:

  1. Inventory: identify agents in use or planned, their owners, identities, data sources and connected tools.
  2. Assess: define each approved use case and the consequences of an error or unauthorised action.
  3. Control: scope permissions, set approval and logging requirements, and test suspension and recovery.
  4. Pilot: deploy to a defined group with training, baseline measures and a route to report issues.
  5. Review: confirm controls and outcomes before extending access; retire agents that no longer have a justified purpose.

For organisations building a custom assistant, governance should shape the design and knowledge sources from the start. Fresh Mango AI's AI Governance & Security Project, Custom AI Agent / Assistant Project and Quarterly AI Success Review address those stages in different ways.

A practical next step

A 30-minute discovery session can help clarify which agents you have, what they can access and which controls to address before production.

There is no obligation, and if a different next step is more suitable, we will say so.

Common questions

Frequently asked questions

How is an AI agent different from a chatbot?

A chatbot may mainly answer questions. An agent may also retrieve information, call connected tools or take actions in business systems. The governance required depends on its access and actions, not its label.

Is an agent safe if it only reads documents?

Not necessarily. Read access can expose sensitive or outdated information, and the agent's response may influence a consequential decision. Review the agent's identity, permissions, knowledge sources and intended use.

Does Restricted Content Discovery change SharePoint permissions?

No. Restricted Content Discovery limits discoverability in specified experiences; it does not change existing permissions. Users who already have access can still access that content directly.

Why does retiring an agent need a formal process?

An unused agent may retain an identity, credentials, connectors or data access. Retirement should include disabling the agent and revoking its access, while retaining any records the organisation needs.

Sources and methodology

This article draws on Microsoft documentation listed below.

Microsoft capabilities, availability and licensing can change. Confirm the current configuration and requirements for your tenant and platform before implementation. Recommendations in this article are general governance guidance, not a substitute for a review of your specific environment.

Your next step

Not Sure Where to Start?

Whether you're exploring AI for the first time or looking to scale existing AI initiatives, our specialists can help you identify practical opportunities and avoid common pitfalls.

Fresh Mango AI specialists reviewing an AI adoption plan with a business client
Book Free Discovery Session