By Errin O'Connor — Founder & Chief AI Architect, EPC Group · Microsoft Solutions Partner across all six designations · G2 Leader in BI consulting
Allie K. Miller — former AWS and IBM AI leadership, and one of the few voices in this industry whose frameworks actually get implemented rather than just reposted — stood on a stage in New York this year and said the thing most executives are still resisting: AI is not a tool. We would love to treat it like a tool, she noted, because tools make us feel secure. But we have entered the era of autonomous teammates, and the human role is shifting from writing prompts to operating like a chief operating officer of AI staff.
She is right, and the psychological reframe she is asking for is the hard part — most organizations are still budgeting, governing, and securing AI like it's Visio with opinions. But the moment you accept her frame, a very old and very boring discipline comes due, and this is the part the keynote circuit never gets to: companies have run teams of workers for a century, and every functioning company answers the same five questions about every worker. Nobody is answering them about the AI ones.
The five questions every org chart answers
Think about what your company knows about any human employee, on day one, without anyone considering it exotic. Who are you — an identity, credentials, a badge. What is your job — a role, a scope, a boundary you don't cross. What do you cost — a salary line someone approved and someone monitors. How are you doing — a review, against criteria, on a cadence. How do we part ways — offboarding, access revocation, a documented exit. Five questions. A century of management science. The entire apparatus we call an employment file.
Now run Miller's autonomous teammates through the same five. Who are they? In most tenants — nobody; they run under the human's identity. What is their job? Undefined; scope is whatever the inherited permissions allow. What do they cost? Unknown until the consumption bill lands — and per-token pricing means your AI staff now has a payroll line that nobody is managing; Gartner predicts AI coding costs will pass the average developer's salary by 2028. How are they doing? Nobody reviews agent output systematically; we spot-check when something feels off. How do we part ways? Ask an enterprise to terminate one specific agent in under an hour and watch the meeting calendar fill up. We adopted the teammates and skipped the employment file. That is the entire risk posture of enterprise AI in 2026, stated in one sentence.
The Non-Human Org Chart
The framework I build with clients is deliberately unglamorous, because the power is in the unglamour. We call it the Non-Human Org Chart, and it applies the five questions with the Microsoft stack the client already owns. Identity: every agent gets an Entra workload identity — its own name, credential, and conditional-access policy; identity inheritance ends as a matter of written policy. Job description: scope defined in permissions and Purview boundaries — which data, which tools, which actions, at what thresholds; one sentence per agent, signed by a named human manager. Payroll: token budgets with monitoring and escalation, routed deliberately — small models for routine work, frontier models where the problem earns them; that routing is cost control, not ideology. Performance review: logged output sampled on a cadence against written quality criteria — the same evidence layer that satisfies an auditor doubles as the review file. Termination: deprovisioning tested, timed, and owned by a name, not a committee.
Working in this industry since 1997 has taught me that the enterprises that survive platform shifts are never the ones with the most exciting adoption story. They are the ones that made the new thing boring fastest — governed, inventoried, reviewed, and revocable. I watched it with SharePoint sites, with Power BI workspaces, with Azure subscriptions. Miller's teammates are the same movie with higher stakes: this time the ungoverned thing doesn't just store your data. It acts on it.
What I tell clients to do
One. Accept Miller's frame fully — then act on its implication. If they're teammates, open the employment files. This week.
Two. Build the Non-Human Org Chart for your top ten agents: five questions, five columns, ten rows. You will not enjoy the blanks. That is the point.
Three. Assign every agent a named human manager whose own performance review includes the agent's conduct. Accountability that isn't attached to a person isn't accountability.
Four. Put the token payroll under the same discipline as the human one — budgets, thresholds, monthly review. Nobody runs headcount on vibes; stop running agent spend on them.
The five-column register
Ship the Non-Human Org Chart as a living register (Dataverse or SharePoint list): AgentID (Entra object) · JobSentence · Sponsor (the named human accountable for the agent; agent conduct appears in their review) · TokenBudget/period plus alert thresholds plus model-routing tier · ReviewCadence plus last-review score · KillSwitchOwner plus last-tested time-to-kill. Board report: top-ten agents by spend and by risk, quarterly, next to headcount.
The register as Microsoft Entra records it
Most of the register already has a home in Microsoft Entra Agent ID, so the spreadsheet only has to hold what Entra doesn't. Start from the tenant list in the Entra admin center (Entra ID, then Agents, then Agent identities). It shows agent identity objects and agents that run on an ordinary service principal; the Uses agent identity column tells them apart. Agents listed in the Agent Registry with no identity sit on a separate list, and they belong in the register too.
| Register column | Where Entra records it | What to check |
|---|---|---|
| Agent ID | Object ID and status (Active or Disabled) on the agent identity; the Blueprint App ID links it to the template it was created from | An agent with no entry here is running under someone else's identity |
| Accountable person | Sponsor. At least one is required when an agent identity or blueprint is created. A sponsor can disable or soft-delete the agent but can't re-enable it, and if a sponsor leaves, sponsorship passes to their manager | Name a sponsor group (dynamic or Microsoft 365) so sponsorship survives a departure |
| Technical owner | Owner. Optional; manages credentials and configuration, and can re-enable, restore or hard-delete the agent | Two owners for any agent that touches regulated data |
| Reporting line | Manager. Can request access packages for the agent's user account; can't change, disable or delete the agent | Don't record the manager as the kill-switch owner |
| Expiry | Lifecycle policy (preview): periodic reconfirmation every 30 to 730 days, a minimum sponsor count, and inactivity (no tokens issued for the set period). The action is disable, soft delete, or disable then delete after a grace period | Without a policy, the expiry column is a dated reminder that someone owns |
| Evidence | Audit log: agent identity changes appear as service principal events with agentType agenticAppInstance. Sign-in logs filter by Agent type and Is Agent | Keep the audit entry ID for every disable test |
Licensing decides how much of this runs on its own. Microsoft lists Agent ID as available to all Microsoft Entra customers, but lifecycle policies are in preview and need Microsoft 365 E7, or a Microsoft Agent 365 license with at least Microsoft Entra P1 or Microsoft 365 E3.
The agent identity register (CSV) carries those columns plus the ones Entra doesn't hold: the job sentence, data scope, token budget, review cadence and the last disable test. Its rows are examples, not client data. Three of them show gaps worth looking for: a sponsor inherited when the original sponsor left, an agent on a plain service principal, and an agent with no identity at all.
The decommissioning test
Disabling an agent identity blocks access to it and stops Entra issuing it tokens. An owner or admin can re-enable it, which restores access, so disable is a pause, not a retirement. Disabling a blueprint blocks every agent identity created from it and stops new ones being created. The test proves the pause works and times it. Run it on one test agent per platform:
- Run the agent's normal task once and note the time.
- The sponsor disables it in My Account (Manage agents, then Disable agent). Note the time.
- Confirm the status shows Disabled in the Entra admin center.
- Find the audit entry for the change and keep its ID.
- Re-run the task every few minutes until it fails. Entra issues a disabled agent no new tokens; this step measures whether anything issued before the disable keeps working, and for how long.
- Check the sign-in logs (Agent type: Agent Identity) for any successful sign-in after the disable time.
- Confirm the sponsor can't re-enable it; only an owner or admin can.
- Write the time-to-kill, from the disable to the first failure, into the register.
The test passes when the status is Disabled, the audit entry exists, no sign-in succeeds after the disable, the task fails, and the time-to-kill is written down. Retirement is the later step: remove the agent's access, then delete it. Owners can restore a soft-deleted agent identity or hard-delete it. The decommissioning test record (CSV) lists each step with its expected result and where the evidence lives, with blank columns for your times.
Sources, read 2026-11-02 (Microsoft Learn): What is Microsoft Entra Agent ID; Administrative relationships in Microsoft Entra Agent ID; Manage agent identities in your organization; View and filter agent identities in your tenant; Manage agents in end user experience; Entra Agent ID lifecycle policies (preview); Microsoft Entra Agent ID logs. Links are in the sources list below.
Where I land
Miller gave the market the right metaphor, and I want to finish it for her. If every employee is becoming a COO of AI teammates, then every enterprise is becoming a company whose workforce is partly non-human — and a company that cannot name, scope, pay, review, or terminate part of its workforce is not transformed. It is exposed. The org chart is the oldest governance technology we have. It is time it learned to hold entries that don't eat lunch — and EPC Group's AI Governance practice builds exactly that.
The data behind this (sources and verification)
- Allie K. Miller — Activate 2026 keynote — AI is not a tool; autonomous teammates; human role shifting from prompt-writer to COO of AI staff.
- Gartner press release, June 24, 2026 — Predicts AI coding costs will surpass the average developer salary by 2028 as token consumption surges.
- Gartner press release, October 21, 2024 (Top 10 Strategic Technology Trends for 2025) — At least 15% of day-to-day work decisions made autonomously through agentic AI by 2028, up from 0% in 2024.
- Microsoft Learn — What is Microsoft Entra Agent ID? — Available to all Microsoft Entra customers; agent identity blueprints as templates; third-party agents via workload identity federation (updated 2026-08-13).
- Microsoft Learn — Administrative relationships in Microsoft Entra Agent ID — Owners, sponsors and managers: what each can do; a sponsor is required on create; sponsors cannot re-enable (updated 2026-09-03).
- Microsoft Learn — Manage agent identities in your organization — Disable at agent, blueprint or tenant scope; automatic sponsor transfer to the manager; sign-in and audit logs (updated 2026-06-17).
- Microsoft Learn — View and filter agent identities in your tenant — The tenant list includes agent identity objects and agents using a service principal (updated 2026-06-17).
- Microsoft Learn — Manage agents in end user experience — Disable agent in My Account blocks access and token issuance (updated 2026-09-03).
- Microsoft Learn — Entra Agent ID lifecycle policies (preview) — Reconfirmation every 30-730 days, minimum sponsors, inactivity; disable, delete, or disable then delete; licensing (updated 2026-10-28).
- Microsoft Learn — Microsoft Entra Agent ID logs — Agent identity changes logged as service principal events with agentType agenticAppInstance (updated 2026-06-17).
Third-party figures above are attributed to their named sources as of the Last verified date. EPC Group audit figures are directional findings from client engagements.
Frequently asked questions
Do AI agents really need performance reviews?
Yes — sampled output against written criteria is both quality control and the evidence file auditors and carriers ask for. Without a review cadence, the only signal is a user complaint or a compliance incident.
How do we budget tokens like payroll?
Budgets per agent or workload, monitoring thresholds, escalation paths, and deliberate routing: small models for routine work, frontier models where the problem earns them. The routing decision is cost control, not ideology.
Who is the agent's manager?
A named human whose own performance review includes the agent's conduct — accountability unattached to a person is not accountability. In Microsoft Entra Agent ID that person is the sponsor: at least one sponsor is required when an agent identity is created, and a sponsor can disable the agent. Entra's separate manager field records the reporting line and can't change or disable the agent.
How do we prove we can switch an agent off?
Run a decommissioning test. The sponsor disables a test agent; you confirm the Disabled status and the audit entry, re-run its task until it fails, check the sign-in logs for anything that succeeded afterwards, and record the time from the disable to the first failure. Disabling stops Microsoft Entra issuing the agent tokens, and an owner or admin can re-enable it, so retirement is a separate step.
What goes in an agent's job description?
One sentence of purpose plus enforced boundaries: which data classes, which tools, which action thresholds — encoded in Microsoft Purview permissions, not left to prompts. The sentence is signed by the named manager.
How is this different from service-account management?
It extends the same discipline: identity plus scope plus logging. The additions are autonomy thresholds, token payroll, and review cadence — things service accounts never needed because they didn't act on your data autonomously.
Ready to act on this?
Start with the practice most relevant to your estate, or reach out directly for a senior-architect conversation.
📧 contact@epcgroup.net · 📞 (888) 381-9725
Multiple models. One truth.
