Direct answer: this month, enterprise Copilot teams should make three decisions. Define where human approval is mandatory for agent actions, add cost evidence to the agent test gate, and move product-change monitoring to Microsoft’s consolidated AI at Work roadmap. Everything else can remain in the watchlist unless it changes a live use case, control or budget.
This briefing is valid as of 24 September 2026. It is deliberately selective. A steering group does not need a catalogue of announcements. It needs a defensible record of what changed, why it matters and who must act.
The four-action filter
Classify every material update as one of four actions:
- Adopt: the benefit and controls are understood, and the change can enter the production standard.
- Test: the change could improve an approved use case, but evidence is still missing.
- Govern: the change alters permissions, data handling, human accountability, cost exposure or an existing policy.
- Watch: there is no current decision because availability, relevance or evidence is insufficient.
Before classification, apply a relevance gate. An update belongs in the briefing only if it affects at least one live workflow, a named control, a committed roadmap item or a material cost assumption. This prevents novelty from consuming steering capacity.
Decision 1: govern human approval for consequential tool calls
Microsoft’s September roadmap describes a Copilot Studio control that lets makers require human approval before an agent executes a specific tool. The request can be approved, approved for the session or denied, and is presented in the channel where the agent is used. Microsoft lists the capability for general availability rollout beginning in September 2026. Because roadmap timing can move and rollout is gradual, verify it in the tenant before relying on it.
Why this requires a decision: a product toggle does not determine which actions require a person. The enterprise must define the boundary. Sending a draft to an internal reviewer is different from sending it to a customer. Reading a ticket is different from closing it. Preparing a payment is different from releasing it.
Use three approval classes:
- No approval: read-only or reversible actions with low consequence and adequate monitoring.
- Conditional approval: actions allowed autonomously only below a defined value, audience or risk threshold.
- Mandatory approval: external communication, financial commitment, access change, deletion, legal effect or other difficult-to-reverse action.
The owner is the business process owner, supported by Security and the platform team. The output is not a generic AI policy. It is a tool-level approval matrix linked to agent inventory and test cases.
Decision 2: make cost evidence part of agent readiness
The same roadmap lists cost visibility in Copilot Studio preview chat, history and agent evaluations for September 2026 rollout. This should change the release gate. A team that can see test and evaluation consumption should no longer approve an agent on functional quality alone.
For every production candidate, record:
- cost per representative completed task, not merely cost per conversation;
- retry and escalation rate;
- evaluation cost and test frequency;
- expected monthly volume, peak load and growth assumption;
- the business alternative, including manual handling and existing automation;
- a budget owner and an alert threshold.
Do not extrapolate from a polished five-prompt demo. Run a representative workload with messy inputs, failures and human handoffs. The decision can then be adopt if quality and unit economics both pass, test if the evidence is weak, or govern if spending controls and ownership are absent.
Decision 3: change the monitoring source, not just the bookmark
Microsoft states that traditional Dynamics 365 and Power Platform release plans stop receiving new publications from September 2026, with new capabilities moving to the AI at Work roadmap. Existing plans remain for historical reference. Enterprises that monitor old release-plan pages risk missing changes or maintaining duplicate reviews.
Assign one owner to update the evidence map used by architecture, adoption and governance teams. The map should distinguish:
- release notes, which describe shipped changes;
- the AI at Work roadmap, which describes planned timing that can move;
- Message Center and tenant administration, which provide tenant-specific change information;
- internal validation, which proves that the capability exists and behaves acceptably in the organization’s environment.
This is a govern decision because it changes the operating cadence and source of record. It is not a reason to approve every roadmap item.
A worked steering example
Consider a service agent that reads support tickets, drafts replies and can close resolved cases. The September briefing produces three actions.
First, the process owner marks ticket reading and draft generation as no-approval actions, customer sending as conditional on confidence and category, and case closure as mandatory approval until false-closure evidence is acceptable. Second, the delivery lead adds cost per correctly resolved case, escalation rate and evaluation cost to the release gate. Third, the platform owner updates the monthly monitoring checklist to use the AI at Work roadmap and records tenant verification separately.
The result is a decision register, not a news digest:
| Decision | Owner | Evidence required | Deadline |
|---|---|---|---|
| Approve tool-level human approval matrix | Service owner | Risk classification and end-to-end tests | Before enabling write actions |
| Approve agent cost envelope | Finance and product owner | Representative workload and unit cost | Before production scale |
| Replace obsolete roadmap monitoring | Platform owner | Updated source map and review cadence | This month |
Limits and trade-offs
Human approval can reduce harm, but excessive approval creates queues and trains users to click through. Cost telemetry improves accountability, but early measurements can mislead when test volume is small. Roadmaps provide planning signal, but they are not contractual availability dates. Tenant-specific validation remains necessary.
The discipline is therefore simple: use product announcements to trigger investigation, not automatic adoption. Each investigation must end with an owner, evidence threshold and review date.
What to do in the next ten working days
- Review every production agent tool and assign an approval class.
- Add unit cost, failure cost and evaluation consumption to the production-readiness template.
- Replace retired release-plan monitoring with the AI at Work roadmap while preserving release notes and tenant messages as separate evidence types.
- Test the relevant capabilities in a sandbox with representative scenarios and customizations.
- Close the briefing with a signed decision register and carry only unresolved items into next month.
Amplified Pi uses this cadence to connect product change to a stable enterprise AI operating model. The briefing is successful when the steering group makes fewer, better decisions, not when it contains more announcements.