Consumption became a project cost

Microsoft's June update described a cost dashboard for usage-based Copilot experiences, with limits and monitoring across users, groups and agents. Group reporting and more detailed agent metrics were scheduled in stages. Project leaders could connect agent consumption with deliverables while distinguishing released features from previews and later rollout.

Purview controls extended to Copilot Cowork, including sensitivity-label inheritance, audit logging, eDiscovery and lifecycle management. These common controls supported collaborative content, while project owners still had to decide which sources and actions an agent actually needed.

The decision became a governed object

Gartner described decision governance on 16 June as a way to make agent decisions explainable, auditable and connected with outcomes. It also identified real-time data and semantics as foundations for agent work. An agent prioritising project variances therefore needed to show the rule, data version and reason for choosing an action.

At the same conference, Gartner warned about agentwashing, limited integration maturity and immature lifecycle tools. OpenAI announced changes to Agent Builder evaluation tooling on 3 June, reminding teams that test infrastructure also has a lifecycle. Evaluations needed exporting, versioning and ownership as project artefacts.

Cost, quality and action were reviewed together

A governance card could record agent version, data sources, permitted decision, consumption limit, evaluation set, human approver and actual result. A weekly review of cost per completed task, incorrect decisions and overrides showed whether automation scaled responsibly. The collaboration team then shared one basis for economics, analysis and accountability.

Controls in practice

A consumption limit also needed to address the agent loop. An agent repeatedly searching the same sources or retrying a failed action could consume substantial capacity without completing work. Telemetry therefore needed steps, tool calls, retries and stop reasons per completed deliverable rather than only aggregate tokens or credits.

Gartner's trend on agentic data streaming made time significant. A live signal might be necessary for an operational choice, while board evidence required stability and reconciliation. Project owners needed to define required freshness, how long a decision remained valid and when an agent should stop because data was incomplete.

Sources

Microsoft 365 Copilot Blog: What’s new in Microsoft 365 Copilot – June 2026, 30 June 2026.

Gartner: Top trends for data and analytics, 16 June 2026.

Gartner: Data & Analytics Summit 2026 highlights, 16 June 2026.

OpenAI: Agent Builder evals retirement notice, 3 June 2026.

For discussion

Which agent decision can the project explain, audit and cost from beginning to end?