Microsoft rebuilt Copilot around agents and code — and switched to metered billing

From chatbot to platform in three tabs

According to Reuters, Copilot is no longer positioned as an assistant you talk to. It is now a workspace with three surfaces: Home for the usual conversational work, Code for building things, and Autopilot for delegating work to an agent that runs on its own.

The Office integration is the part most people will notice first. Word, Excel and PowerPoint are embedded directly inside the Copilot interface, so a document or a spreadsheet can be edited and discussed in the same window instead of being thrown back and forth between apps. That sounds like convenience. In practice it also means Copilot stops being an optional sidebar and becomes the place where the work happens — a meaningful shift in how Microsoft positions the product against Google's Workspace/Gemini bundle.

Code: a sandboxed IDE behind a chat box

The new Code tab lets users describe an app, a dashboard or an internal tool in natural language and get it built in a controlled environment. Microsoft's framing is "citizen development without the shadow IT" — keeping generated code separate from production systems until somebody approves it is the selling point. The exact underlying technology should be confirmed in Microsoft's official Copilot page before an organisation treats the environment as equivalent to an existing development platform.

As always with this category, the honest question is not "can it build a dashboard" but "what happens on the ninety-first iteration". Agentic coding tools burn tokens on retries, test loops and re-reads of the same file. In AI Arena, where we benchmark local models on a 16 GB RTX 5060 Ti alongside cloud APIs, our editorial observation is that a single chat turn and a single agent loop can differ by an order of magnitude in compute. That matters enormously when the compute is metered.

Autopilot is Scout with an identity card

According to the Reuters report cited above, the agent Microsoft first showed in June 2026 under the name Scout has been renamed Autopilot and given something more consequential than a new label: an IT-managed identity. Reuters also reported that Autopilot has persistent memory and delegated permissions, and is designed to be assigned work like a colleague — and to keep working when you close the laptop.

That is a genuinely different product category from a chat assistant, and it drags a very different compliance footprint behind it. An agent with persistent memory and delegated permissions needs to be assessed as part of the organisation's identity, access and data-governance model rather than treated simply as a text generator.

The real story is the meter

Microsoft is moving away from pure per-seat subscriptions toward a hybrid per-seat-plus-consumption model. The base per-seat licences remain, while use of Code, Autopilot and other metered capabilities may be billed on top, by usage. Microsoft's official Copilot page and the Reuters report linked above describe the hybrid mechanism, but the precise units, rates and billing controls need to be confirmed in the applicable commercial documentation.

Read that pairing carefully. The company is introducing variable costs alongside the existing licences, a shift that transfers more of the consumption risk to the customer. What we have so far describes the mechanism, not the rate — the per-unit price of the consumption layer is not something the published details pin down. For a European finance department, a variable line item with an unpublished rate is a line item that needs a hard cap before rollout, not after.

Why metering changes planning

There is no reliable model-by-model price comparison to make here until Microsoft publishes the relevant rates and the underlying model names and figures can be independently checked. That is important because agent loops can consume substantially more compute than a single chat turn, but it would be misleading to convert that observation into a precise Copilot cost estimate without verified billing units.

The practical trade remains clear even without an artificial price table. Managed integration and governance can be valuable to a European customer, but the organisation gives up some control over model selection and must understand how variable usage is measured. Procurement teams should therefore ask for the rate card, included allowances, overage rules, regional tax treatment and administrative spending controls before approving a rollout.

Article 50 and an AI agent acting for a company

Here is the European angle that deserves more attention than the demo videos. The EU AI Act's Article 50 transparency obligations are scheduled to apply from 2 August 2026, subject to the final applicable rules and the specific deployment context. The exact obligation depends on the type of interaction, the system's role and the deployment context. The European Commission's AI regulatory framework overview is the appropriate starting point, but organisations should obtain legal advice for their specific use case.

Now put Autopilot in that frame. If it communicates with a supplier, customer or job applicant, the interaction may trigger transparency requirements depending on what the recipient is told, the nature of the exchange and the way the system is deployed. The company using the system may be treated as the deployer, but that conclusion is context-dependent and should not be assumed automatically. Practical consequence: you need a documented disclosure policy for agent-sent correspondence, a record of what Autopilot's memory holds and how it is purged, and a lawful basis under GDPR for any personal data the agent stores. Legal and data-protection review should happen before external deployment. In Germany or France, employee-facing consumption telemetry is also the kind of feature that works councils will want to discuss before, not after, deployment.

One more note on the high-risk calendar: the Digital Omnibus (Regulation (EU) 2026/1744) amended the general implementation deadlines for Annex III high-risk systems to December 2027 and for Annex I embedded products to August 2028. Individual obligations depend on how a product is classified and deployed, so organisations should review their specific compliance plan rather than assuming every deadline moved in the same way. If your plan still assumes the original August 2026 date for all systems, it should be reviewed.

What I would check before switching Code and Autopilot on

Three things, in order. First, a hard consumption cap per team, not a global tenant cap — otherwise one enthusiastic engineer sets the budget for everyone. Second, written confirmation of where Code's sandbox and Autopilot's memory store data, and whether they sit inside your existing EU data boundary commitments; the question is contractual and worth asking in writing rather than assuming. Third, a documented owner for agent identity lifecycle: who creates the identity, who revokes it when a project ends, and what happens to its memory afterwards.

Teams that need European document processing inside their own pipeline should compare contractual data-residency terms directly with vendors. The point is not to avoid Microsoft. It is to stop treating agent platforms as a seat licence with a fixed number at the bottom of the invoice.