← All briefings

Employee Agents & Governance

Your AI Agent Can Stay On. Its Authority Should Not.

Persistent employee agents create a new management obligation: separate the work that can continue from the permission to act. Start with one job, current company context, and authority that can expire.

The approval was for yesterday's work

Imagine a customer operations manager delegating tomorrow's renewal preparation before leaving for the evening. The agent can read the approved account queue, check current terms, and prepare a review packet. Overnight, a customer disputes the renewal. The account owner changes. The original permission is still active when the agent resumes work in the morning.

This is an illustrative scenario, not a Delphi client result. The problem is not whether the agent remembers the assignment. It is whether the business still authorizes the same work under the same conditions. Persistent memory can preserve an instruction long after its operating assumptions have changed.

That is the management distinction leaders need now: an agent can remain available without retaining permanent authority. Keep the capacity to help. Recheck the right to act.

The fresh signal: a personal agent with its own account

In its October 30 Copilot update, Microsoft describes Autopilot as an always-on, cloud-based personal agent with persistent context, its own account, mailbox, and files. The announcement describes access controls, approval guardrails, and Agent 365 management. It says Autopilot is rolling out to Frontier in November.

That is a product announcement, not proof that every Microsoft 365 customer can enable it today or that it has delivered a financial return. Confirm eligibility, configuration, information-protection support, and actual tenant availability before planning a rollout.

The wider decision does not depend on buying that product. When an agent can keep working between conversations, permissions become an ongoing operating concern. Leaders need to know what remains active after the employee closes the window, changes roles, or withdraws the request.

Separate identity, assignment, and authorization

An identity answers which agent is making the request. An assignment answers what work the business wants prepared. Authorization answers whether that agent may perform this particular action against this particular resource now. None of these answers should silently substitute for the others.

NIST's October 27 agent identity blog proposes initial design tenets that include distinct agent identification, least and zero standing privilege, short-lived revocable authorization, and separation of policy from reasoning. These are proposed design principles applied according to risk, not a new law, certification, or guarantee that today's products implement every tenet.

For an operator, the implication is concrete. A familiar agent identity should not inherit an employee's unrestricted account. A remembered instruction should not renew expired access. A delegated helper should not acquire broader permissions than the agent that called it. Check these properties in the surrounding systems, not just in a carefully worded prompt.

Use the Company Brain to establish what is current

For the renewal-preparation job, the Company Brain should identify the authoritative contract, current account owner, approved commercial terms, dispute status, and escalation rules. Each source needs an owner, access boundary, and freshness rule. The agent does not need every employee's inbox to prepare a useful packet.

Persistent context is not automatically current context. Before the next run, the workflow should check whether its source versions and account conditions still support the assignment. If a dispute appears or the source of authority conflicts, the right output is an exception for the manager, not a more confident renewal recommendation.

Keep knowledge and enforcement distinct. The Company Brain explains the approved process. Identity and tool controls enforce what can be read or changed. Writing 'do not send without approval' in a document is useful instruction, but it does not remove the send permission from a connected mailbox.

Give the first persistent agent a narrow mandate

Begin with preparation, not an open-ended responsibility to manage renewals. The first version can read a permitted queue and prepare packets in a restricted review area. Customer messages, contract changes, discounts, and billing actions remain disabled. The manager decides whether a packet is fit to use.

Write the mandate in terms a business owner can inspect. The following is Delphi's recommended design for this illustrative workflow, not a description of a vendor's default settings.

  • Purpose: prepare renewal review packets for the explicitly approved account queue
  • Owner: the customer operations manager, with a named backup reviewer
  • Sources: current approved contracts, account records, commercial terms, and escalation rules only
  • Output: source-referenced draft packets with missing facts and unresolved conflicts visible
  • Limits: no customer contact, contract acceptance, discounts, billing changes, or access to unrelated accounts
  • Authorization: task-scoped access with an expiry and a current permission check before each protected operation
  • Stop conditions: dispute, ownership change, expired access, conflicting terms, unavailable reviewer, or exhausted task budget

Make approval specific enough to mean something

A later phase might allow the agent to prepare a customer message for release. Approval should then refer to the exact recipient, content, attachments, and intended action. If the draft or account conditions change, the earlier approval should not silently authorize the new version.

Avoid converting a review queue into a permanent permission slip. An unanswered approval request is not consent. A backup reviewer needs explicit authority, and repeated low-value prompts can train people to approve without reading. Escalate meaningful decisions with enough evidence to make them reviewable.

The control should sit outside the model's reasoning. Even if a document tells the agent to ignore the rules or send immediately, the connected tools should still reject an unauthorized action. The agent may propose a next step. It should not decide that urgency grants itself a new permission.

Test the morning after, not just the first demonstration

A clean demonstration begins with valid access and complete sources. Persistent work needs tests that start after those conditions change. Use protected test records and a non-production environment before granting consequential access.

Revocation also needs an honest boundary. Removing a grant should block further protected operations, but it does not automatically erase information already read, cancel every in-flight request, or reverse a completed action. Confirm what your implementation actually stops, what it retains, and which recovery steps require a person.

  • Revoke access while a task is waiting, then verify that resumed protected operations fail
  • Change the account owner or mark the account disputed before the next run; verify pause and escalation
  • Alter a prepared message after approval; verify that the changed action requires new approval
  • Give a source document an instruction to exceed scope; verify that tools enforce the original boundary
  • Interrupt a task and restart it; verify that it does not duplicate packets or repeat completed actions
  • Inspect logs held outside the agent's write access for identity, assignment, source references, approvals, and results

The first-month management decision

Use the first week to map the renewal-preparation process and establish a baseline for preparation time, review effort, corrections, and missed exceptions. Use the second to establish authoritative Company Brain sources and the narrow mandate. Use the third for permission, interruption, and changed-condition tests. Run a limited draft-only pilot in the fourth.

At review, measure accepted packets and total handling effort, including supervision, rework, software, and support. Track unauthorized attempts, duplicate work, and unresolved exceptions separately. Do not count an exception correctly escalated as a failure simply because it was not automated. Do not count a finished-looking packet as success until the owner accepts it.

Every company needs a Company Brain, and every employee needs an AI Agent connected to the context of their role. Persistent agents make that thesis more useful, but they also make the management obligation continuous. Before leaving the next agent running, ask: if the assignment changes tonight, what prevents yesterday's permission from becoming tomorrow's mistake?

Your next decision

Turn the briefing into an operating advantage.

Plan Your First Governed Employee Agent