Agent governance and assurance

Know what is running, what it can reach, and who approved it.

An inventory of the agents in your environment, permission design in Entra and Purview, approval gates at the points that move money or commit the organization, and evidence an auditor will accept.

Boards, insurers and enterprise customers have stopped asking whether an AI policy exists. They are asking what was enforced, on which systems, on which dates, and who reviewed it.

Review your AI controls
The four failures

What actually goes wrong.

These are the findings that surface once an organization has more than a handful of AI workloads, and they tend to arrive in roughly this order.

Nobody has the list

Agents and assistants get built in four places by three teams. The first honest inventory is usually longer than the leadership team expected and includes at least one thing running on a personal account.

Permission inheritance

Retrieval surfaces a decade of over-permissioned sites in a way no access review ever did. The model is not the problem. It is the first tool efficient enough to find the problem.

Quality moves quietly

A prompt is edited, a source is re-indexed, a model version changes. Without a held-out test set and a regression run, the first signal is a user who stops trusting it and says nothing.

No evidence trail

The answer was right, and there is no record of which documents produced it, which tool it called, or who approved the action. That is a defensible system with an indefensible file.

What we put in place

Six controls we put in place.

  • An agent register with a named business owner for each entry
  • Identity and entitlement design in Microsoft Entra, with service identities scoped to the task
  • Sensitivity labelling and data loss controls through Microsoft Purview
  • Approval gates at every action that moves money, stock, master data or a customer commitment
  • An evaluation harness with a held-out set, run on every change to prompt, source or model version
  • Logged retrievals, tool calls and approvals, retained to your evidence standard

Purview and Entra are Microsoft controls that AIM implements. They govern what an identity may reach, which is most of what governs an agent. The evaluation harness and the agent register are ours, because no platform ships them for your processes.

Who does what

One accountable partner, two named practices.

AIM is the partner of record and holds the relationship, the architecture and the roadmap. Arancia is a separate organization with its own security practice and 24/7 operations, and it delivers the security work named below. We state the split because it matters to your procurement team, your auditor and your insurer.

AIM

Architecture and controls
Identity and access design, information protection through Purview and Entra, agent lifecycle and approval design, secure landing zones.
Assurance evidence
Evaluation results, change history, approval records and the operating model that keeps controls enforced between audits.

Arancia

Security operations
24/7 SOC and NOC, threat detection and monitoring, incident response and coordination.
Governance, risk and compliance
Control framework mapping, audit support, and offensive security testing where the environment warrants it.
Talk to both teams at once
Anthropic overview Deploy Build Operate
One partner. The full journey. No handoffs.

Start with the workload, not the licence.

A two-hour session with the people who own the process, the data and the risk. You leave with a shortlist of candidate workloads, the constraints that rule some of them out, and an honest view of which one is worth a pilot.

Book an AI value working session