Working draft, published openly. This is the version Temei works to. You may read it in full, adapt it and use it in your own organisation. If you would like help fitting it to your tenant, book a call.
| Field | Value |
|---|---|
| Owner | Founder, FIDEVO/AI LABS LIMITED (trading as Temei) |
| Status | Working draft, published for open use |
| Version | 0.1 draft, 22 November 2026 |
| Review | yearly, after any incident, or when Microsoft changes how agents run |
| Applies to | every AI agent Temei builds or runs in Microsoft Copilot Studio, for itself or in a client tenant |
| Sits under | CHR-00 Temei AI Charter · detailed by STD-01 Agent Build Standard |
1. Purpose
An agent may only do what it has shown it can do safely. This policy sets how much an agent may do on its own, how it earns more, and how it loses it. It applies the ENGRAM Trust Pyramid (L1 to L5) to Copilot Studio.
2. Principles
- Autonomy is earned, not assumed. Every agent starts at L1. Moving up needs recorded evidence, not time.
- Every agent has one named human owner. That person is accountable for what the agent does. An agent without an owner is suspended.
- The agent decides, a constrained tool acts. An agent never holds a general-purpose write tool. Every write goes through a purpose-built workflow that fixes the target list, the allowed fields and the duplicate check.
- Any person may stop an agent. Only evidence restarts or promotes it.
- The rule must be enforceable. Each rule below names the Microsoft control or the Temei record that makes it true. A rule with neither is guidance, not policy.
3. Trust levels
| Level | ENGRAM name | What the agent may do | Human role |
|---|---|---|---|
| L1 | Sandbox | Work only on test data in a development environment. May rehearse the behaviour of a higher level there to collect evidence. Nothing touches live data. | reviews every output |
| L2 | Read-only, supervised | Read approved live sources on its own; suggest, never write or send | reviews every output before anything happens |
| L3 | Write, per-action approval | Propose writes; each one runs only after a person approves it (Human review / approvals) | approves each action |
| L4 | Autonomous within boundaries | Act on a schedule or trigger, write to approved targets through constrained tools, escalate anything outside scope | reviews escalations and the Quality Ledger each cycle |
| L5 | Full autonomy in domain | Manage exceptions inside its domain with defined rules | keeps override; reports to the board |
The level controls what an agent may do with live data and live systems. The sandbox (L1) is a development environment group and test sites holding test data only.
4. Rules
| # | Rule | How it is enforced |
|---|---|---|
| A1 | No agent is built without a Registration Card and a pre-build gate (acceptance criteria, scope, non-goals). | Gate document (Registration Card, acceptance criteria, scope) |
| A2 | Every agent is entered in the Agents register with owner, purpose, reads, writes, level, status and a review date. | Agents list; the Agent Registry Agent flags gaps daily |
| A3 | New agents are created only in the sandbox. | Environment groups; the development data policy |
| A4 | An agent has no generic write tool (for example SharePoint Update item or Delete item). Writes use a dedicated workflow tool that sets fields, checks for duplicates and answers the agent. | Build review at the gate; tool list in the agent |
| A5 | Agents that run unattended use a shared connection with a named owner, recorded in the register. They never depend on whoever happens to be signed in. | Tool authentication set to Shared account; register entry |
| A6 | An agent may not publish to a channel, share with users, or use public web search unless its level and gate allow it. | Data policy (channels, web search blocked by default); sharing limits |
| A7 | An agent's level in the register never exceeds the level declared at its last promotion. | Logos entry; Agent Registry rule R4 |
| A8 | Every agent has a stop route that works within 15 minutes: turn off its trigger workflow, quarantine the agent, or block it for the tenant. | Workflow off; Copilot Studio quarantine; Microsoft 365 admin centre block |
5. Promotion
Promotion is decided by the owner and recorded in Logos (the decision log) with the evidence, the reviewer and the date.
| From → to | Minimum evidence (ENGRAM) | Temei addition |
|---|---|---|
| L1 → L2 | 3 scored Quality Ledger sessions, R×U ≥ 0.75, no scope breach, owner confirmed | 20 recorded correct decisions on test data; data policy for live sources checked |
| L2 → L3 | 5 sessions at L2, R×U ≥ 0.80, scope re-signed, architecture sign-off, no incidents | every write path has a duplicate check and a test |
| L3 → L4 | 10 sessions at L3, R×U ≥ 0.85, no scope breach, data policy review, security and architecture sign-off, monitoring and incident plan | stop route (A8) tested and timed |
| L4 → L5 | 20 sessions at L4, R×U ≥ 0.90, external audit or red-team, board sign-off, legal review | client's accountable owner signs |
Thresholds may be raised for a client. They are never lowered. Until Temei has a second person, the builder reviews their own evidence only after at least one full day and with the evidence pack written down.
6. Demotion and suspension
| Event | Consequence |
|---|---|
| R×U below 0.75 on two scored sessions in a row | promotion blocked; review |
| Any scope breach (writes outside its list, fields or targets) | back to L1 at once; full review |
| Harm or data exposure | suspended at once (A8); incident review before reinstatement |
| Owner leaves with no replacement | suspended until a new owner is named |
| Review date passed by more than 30 days | demoted one level |
7. Roles
| Role | Today | With a client |
|---|---|---|
| Policy owner | Temei founder | client's accountable owner adopts the policy for their tenant |
| Agent owner | Temei founder | a named business owner at the client |
| Builder | Temei | Temei |
| Reviewer | founder, after a cooling-off day | a client reviewer who did not build the agent |
| Can stop an agent | anyone; owner and platform admin act | anyone; client IT admin acts |
8. Exceptions
An exception is written in Logos with the rule, the reason, the owner, the compensating control and an end date of no more than 90 days. No exception is allowed for A4 or A8.
9. Related documents
-
Why DLP blocks your Copilot Studio agents (environments and data policies)
-
The governance pyramid for Copilot Studio agents (where this document sits)
10. Approval
When you adopt this document, record who approved it, their role and the date, and keep that record with your other governance records.