That reading is an implication, not a Microsoft finding. Microsoft did not announce a CX policy, a customer-service agent, or any customer-outcome result. What it did say is concrete and useful for admins, and it comes with some limits.
What Microsoft actually said
Here are the figures from the call, with what each one does and doesn't show.
- Agent registry: Over 50 million agents are now in the registry, up over 30 percent quarter over quarter. This is a registry count. It doesn't show how many agents are active, in production or facing customers.
- Audit volume: Customers have used Purview to audit over 7.6 billion agent activities to date, up 5x quarter over quarter. That is a cumulative activity count. It isn't a count of agents or customer cases, and it says nothing about whether the actions were correct.
- Automated tasks: Nadella said automated tasks have nearly doubled since July. The call passage gave no absolute number, no measurement method, and no sign that the figure is about customer service.
- Usage-based billing: Nadella said the number of Copilot customers consuming credits through usage-based billing "grew over 75% over the quarter." That measures customers consuming credits. It isn't a 75% rise in spend.
- Dynamics 365: Microsoft introduced 30 new Copilot skills across Dynamics 365. They give access to CRM and ERP capabilities in the flow of work. Microsoft didn't call them CX skills.
- Healthcare: Microsoft said it automated 30 million patient encounters this quarter, up 70% year over year. This is a company-reported figure, not a count of autonomous decisions.
Corroboration is broad. The Next Web reported more than 50 million agents in Agent 365, up 30% in a quarter. Several transcript publishers carry the same line.
"Separate intelligence from authority"
The key governance line from the call is that Microsoft's approach is to separate intelligence from authority, enforcing identity, permissions, monitoring and containment outside the model itself. In practice, the model proposes and a separate layer decides what is allowed.
Nadella also said Microsoft expanded its Perception security offering last week. The additions were:
- Agent discovery across local and cloud environments.
- Dynamic permissions meant to enforce least privilege for each task.
- Continuous risk assessment.
- Teams of "guardian agents" that watch agent activity to help prevent unsafe actions.
These are Microsoft's descriptions of its own products. They don't show that every customer has the controls switched on, and they don't guarantee safe outcomes.
The Windows angle
This isn't only a cloud story. Microsoft said it introduced Microsoft Execution Containers, a policy-driven execution layer that governs and contains agents across Windows and other platforms, integrated with Agent 365 for governance and observability. Agents like OpenAI Codex, Anthropic Claude Code, and Meta Muse are adopting MXC, and customers like Meta are starting to deploy it across their enterprise environments. The call gave no deployment detail beyond that, such as supported Windows versions or policy options.
What Agent 365 covers, and where it stops
Microsoft's product pages describe Agent 365 as a control plane that splits responsibilities across four admin groups:
- Defender admins handle security posture and threat protection.
- Entra admins handle agent identities and access.
- Purview admins handle the data that agents use and create.
- Microsoft 365 admins handle inventory, onboarding, policy templates and analytics.
Microsoft says agents published through Microsoft 365 channels and registered with an Entra Agent ID appear in the inventory automatically. Agents built elsewhere need extra steps. So "50 million in the registry" doesn't prove that one inventory captures every agent in every enterprise system.
Microsoft Learn documents how external platforms connect. Supported platforms are Amazon Bedrock, Google Vertex AI, Salesforce Agentforce, Databricks Genie, Anthropic Claude Managed Agents, Oracle Generative AI Agents and Snowflake Cortex. The documented limits matter for anyone planning around it:
- Manual sync: Synchronization is manual. There is no scheduled sync.
- Deletions: An agent removed from the source platform isn't guaranteed to disappear from the Microsoft registry. Admins may need to remove it by hand.
- Observability gaps: Oracle and Snowflake observability were marked "Coming soon" in the documentation.
- Frontier preview: Observability collection is a Frontier preview feature, so availability and capabilities may change.
- Provider charges: Pulling telemetry can query the provider's logging or data services, such as Google BigQuery or Amazon S3. That can incur charges from the provider.
- Default setting: New connections have the "Collect agent observability data" setting enabled by default.
- Read-only registry sync: Sync gives visibility and governance only. Organizations that build and host their own agents need the Agent 365 SDK for code-level capabilities.
- Credentials: Every currently supported platform needs a key-based credential. Keyless options such as workload identity federation aren't supported yet.
Audit trails: necessary, not sufficient
Purview's audit documentation lists Agent 365 activities that are logged when an agent is invoked or performs actions or tool calls. The events include Executed AI tool, Invoked AI agent, Made AI inference call and Applied AI guardrail. Audit logging is on by default for Microsoft 365 organizations.
These give admins specific event types to search after a disputed outcome. A log entry shows that an action happened. It doesn't show that the decision was right, that a customer wasn't harmed, or that a human approved it.
Consumption pricing and the budget question
Nadella said Microsoft's model is expanding from per-seat subscriptions to include consumption. Benzinga's coverage notes the shift toward usage-based billing. Microsoft's list pricing for Agent 365 on its product page is $15 per user per month on an annual commitment. It is also included in Microsoft 365 E7, listed at $99 per user per month.
Spend that follows usage is harder to forecast than seat counts. On the call, Nadella said Microsoft doesn't want customers to overbuy one option and end up "token maxing." He also said consumption would move faster in coding than in core Microsoft 365, where Amy Hood expected a slower shift.
A practical review checklist for CX and IT teams
Microsoft doesn't prescribe this list. It is a governance checklist built from the capabilities and limits above. Before an agent executes work that touches customers, check that you can answer these questions:
- Who is the accountable human owner of this agent?
- Which systems, data and actions is it permitted to use?
- Is access least-privilege, and what thresholds trigger escalation to a person?
- Are tool calls and actions logged, and can you search them in Purview?
- Is the agent actually in the registry? For third-party platforms, has someone run a sync lately?
- Who can stop, reverse or review an action after a customer dispute?
- What will it cost to collect telemetry, on both the Microsoft side and the provider side?
Bottom line
The earnings call confirms that Microsoft is pushing agents toward governed execution, with identity, permissions, audit and containment kept outside the model. It doesn't prove that automation is improving customer outcomes. The call named no CX deployment and gave no service-quality result. The Unilever dispute-handling example Nadella cited is a business-process case, not a Microsoft CX result. The CX accountability question is real, but the call raises it without answering it. Teams should answer it before agents get write access to customer records.
References
- Microsoft’s AI Agents Put CX Accountability on Notice - CX Today CX Today · 2026-10-29T12:31:00.501000+00:00
- Microsoft Wants to Charge More as AI Agents Do More Work, While Google Sees a New Advertising Gold Mine — - Benzinga benzinga.com
- Microsoft Agent 365: The Control Plane for Agents microsoft.com