Copilot Now Executes, Verifies, and Governs Itself: What Changed in 2026

In 2026, Microsoft 365 Copilot went from suggesting text to executing complete tasks, verifying its own answers across multiple models, and understanding context across the entire organization. Anyone still evaluating it as a writing assistant is measuring the wrong product.

The License Nobody Uses Well

An operations manager opens Copilot, asks for a summary of Tuesday’s meeting, and closes it. That’s all they do all month. The license keeps getting billed.

That scene is common, and it explains why many IT committees hesitate when it’s time to renew. But the Copilot they evaluated in 2024 no longer exists. What exists today is a platform that delegates work, compares models, connects to business systems, and gets managed like any other critical asset.

The scale changed too. In January 2026, Microsoft reported 15 million paid Copilot licenses, close to 3.3% of its 450 million commercial Microsoft 365 users. Adoption is still low. The advantage for whoever moves first, high.

Figure	What it means
15 million	Paid Copilot licenses as of January 2026: the market is still in an early stage
May 1, 2026	Agent 365 reached general availability, at USD 15 per user per month or included in Microsoft 365 E7 (USD 99)
July 24, 2026	Claude Opus 5 began rolling out in Word, Excel, PowerPoint, Chat, Cowork, and Copilot Studio

This article walks through the four changes with the biggest impact on IT leaders, and what’s worth doing about each one.

From Answering to Executing: Copilot Cowork and Agent Mode in Office

The deepest change isn’t a new model. It’s that Copilot now finishes work.

Copilot Cowork

You delegate an outcome, not an instruction.

• The user describes what they need, and Cowork executes it using the organization’s emails, meetings, messages, files, and data.
• The task becomes a plan that moves forward in the background and reports checkpoints.
• Since August, you can choose the effort level: “Faster” for simple tasks with lower credit consumption, “Smarter” for complex reasoning, slower and costlier.
• The /cost command shows the percentage of the monthly credit limit still available and when it resets. A small detail finance cares about a lot.

Copilot Inside Word, Excel, and PowerPoint

The whole document, not the stray paragraph.

• In Excel, Copilot now executes Python code for advanced analysis, automation, and data transformation, with results delivered inside the workbook.
• A new design skill applies palette, chart style, and structure in a single pass. Raw data goes in; a readable report comes out.
• Copilot Chat and Cowork can reason over Power BI reports and semantic models.

Copilot’s value is no longer measured in minutes saved drafting. It’s measured in tasks nobody had to push through step by step.

From One Model to Several: Choice, Opus, and Cross-Verification

For two years, which model Copilot used was Microsoft’s decision. Not anymore.

Today, models from OpenAI, Anthropic, and Microsoft’s own coexist. The Anthropic models are added as alternatives, not as a replacement for OpenAI.

Claude Opus: the specialist in long-form work

Opus 5 is built for complex, multi-step work and improves on Opus 4.8 in agentic code, professional work, and extended reasoning. In practice, it shows up in longer Word drafts, sturdier Excel models, and presentations that arrive nearly finished.

Researcher with Critique and Council: models that review each other

Here’s the improvement that should matter most to whoever signs off on decisions based on reports.

• Critique: one model drafts the research and another independently reviews it for accuracy, completeness, and source quality.
• Council: two models produce full reports in parallel, and a third points out where they agree, where they differ, and what each one contributes.

According to Microsoft’s own tests on the DRACO benchmark, Researcher with Critique scored 57.4 points, 13.8% higher than the best previous result. It’s a vendor-reported measurement, but the logic holds: two models that contradict each other reveal exactly where a human needs to look more closely.

From Guessing to Understanding: Work IQ and Organizational Context

A frontier model with no context is a brilliant consultant on their first day: they reason well, but they don’t know who decides, which project is at risk, or what was agreed on Thursday.

Work IQ fixes that. It’s the layer that connects Copilot and agents to the organization’s shared context in real time, organized across three levels: data, memory, and inference.

What matters for IT is that it isn’t limited to Microsoft 365:

• Copilot Connectors pull in data from external systems and business applications, with hundreds of ready-made connectors or the option to build your own.
• Copilot can now create and query Planner tasks directly.
• Service Agent, now generally available, brings Copilot the context of cases, customers, and knowledge bases, backed by Microsoft Graph and Dataverse through Work IQ.

Before: the user uploaded files, explained the project’s history, and rewrote the prompt three times.

Now: Copilot already knows where to look, with that exact user’s permissions.

From Loose Agents to Governed Agents: Agent 365

When any department can create agents, the question stops being “what can they do?” and becomes “how many are there, and who controls them?”

Microsoft answered with Agent 365. It gives real-time visibility into the agent ecosystem to catch performance or risk signals before they hit the business.

• It extends Microsoft Entra’s network controls to Copilot Studio agents and to local agents on users’ devices, to detect unauthorized use and block prompt-based attacks.
• It can import agents from AWS Bedrock and Google Gemini Enterprise into its inventory, in preview.
• Admins can generate custom reports by exporting data from the Copilot and Agent 365 dashboards.

There’s a useful nuance for architecture: agents built in Copilot Studio don’t auto-update to the new Opus models on their own. That gives you room to test before an agent’s behavior changes in production.

What No Upgrade Solves for You

Every Work IQ response respects the user’s identity, permissions, and compliance policies. That sounds reassuring. But it means Copilot respects the permissions you have today, including the bad ones.

If a SharePoint site with salary data was left open to “the whole organization” four years ago, a Copilot capable of executing tasks, cross-referencing sources, and reading external systems is going to find it before your auditor does.

Every new Copilot capability multiplies what already exists in your tenant. If there’s order, it multiplies value. If there’s disorder, it multiplies exposure.

There’s a second front for anyone operating in Europe or under strict regulation: Anthropic’s models sit outside the EU Data Boundary and its local-processing commitments. For most commercial tenants, including Latin American ones, they’re enabled by default. It’s worth reviewing that decision instead of inheriting it.

How to Capture the Value in 90 Days

1. Review your actual configuration. Enabled models, Anthropic controls, and active Frontier program features. Don’t assume they’re the same across every experience.
2. Clean up over-shared content. Prioritize sites with financial, legal, and HR information. It’s the biggest risk, and the cheapest one to fix today.
3. Pick three processes for Cowork. Repetitive, multi-step tasks: monthly closes, project reports, committee prep. Measure hours before and after.
4. Connect a business system. An ERP, CRM, or help desk via Copilot Connectors. That’s where the value the board can actually quantify shows up.
5. Govern agents like software. Inventory, owners, testing before switching models, and tracking credit consumption.

Reasonable Objections

“Copilot adoption at our company is low. More features won’t fix it.” More features, no. Features that finish work, yes. Low adoption usually comes from a Copilot used as a search engine with better writing. When a user delegates a full report and gets it back ready to review, usage shifts into a different category.

“Autonomous agents are an operational risk we can’t take on.” That’s a legitimate concern, and it’s exactly why Agent 365 exists. The real risk isn’t governed agents — it’s the ones already being created in your organization with no inventory and no owner.

“Costs are going to become unpredictable.” Partially true: the credit model demands discipline. But today there are concrete levers, like choosing Cowork’s effort level, checking consumption with /cost, and analyzing usage from the dashboards. What’s unpredictable is not measuring.

The Advantage Is No Longer the License

Frontier models, autonomous execution, and multi-model verification are available to any company paying for Copilot. That no longer differentiates anyone.

What does differentiate you is having an orderly tenant, business systems connected, agents under control, and measured use cases. That advantage gets built; it doesn’t arrive with the next update.

Copilot already changed. The question is whether your organization changed with it.

Sources

• Neowin — Here are all the new features added to Microsoft 365 Copilot in August 2026
• Microsoft Tech Community — What’s New in Microsoft Copilot, August 2026
• Microsoft Learn — Release Notes for Microsoft 365 Copilot
• A Guide to Cloud & AI — What’s New in Microsoft 365 Copilot: August 2026
• AlternativeTo — Microsoft gets its own Claude Cowork version
• Microsoft Tech Community — Available today: Anthropic Claude Opus 5 in Microsoft 365 Copilot
• iTechGuides — Anthropic’s Claude Models in Microsoft 365 Copilot: What’s Available in 2026
• Office Watch — Copilot Researcher: Critique and Council Explained
• How2Shout — Microsoft Copilot Now Uses Claude to Review GPT’s Work
• Microsoft Learn — Work IQ in Microsoft Copilot Studio
• Microsoft Tech Community — A closer look at Work IQ
• Microsoft — Work IQ
• Microsoft Dynamics 365 Blog — Service Agent general availability
• 4sysops — What is Microsoft Agent 365?
• Microsoft Learn — Microsoft Agent 365 overview
• Microsoft Security Blog — Microsoft Agent 365, now generally available
• WinBuzzer — Microsoft Agent 365 Hits General Availability
• Rob Quickenden — Claude Opus 4.8 now available in Microsoft 365 Copilot
• Microsoft Docs — Anthropic as a Microsoft subprocessor
• Tom Arbuthnot (X) — Anthropic models enabled by default in Microsoft 365 Copilot