Today, May 1, 2026, Microsoft 365 E7 hits general availability at $99 per user per month. It is the first new enterprise license tier Microsoft has introduced in eleven years — since E5 launched in 2015 — and the company is using it to make the largest single architectural bet of its 2026 enterprise AI cycle: that the control plane for AI agents lives where the identity and productivity stack already runs.
Agent 365, the most consequential component of E7, also goes GA today as a standalone SKU at $15 per user per month for organizations not ready to commit to the full bundle. That separation matters. Microsoft is selling E7 as the strategic suite. Agent 365, sold separately, is the tactical entry point — and it is where most of the early traction is going to come from. Tens of millions of agents have already registered in Agent 365 during the two-month preview period, and IDC's 1.3 billion-agents-by-2028 projection now has a procurement vehicle attached to it.
For enterprise AI engineers and the executives who are about to be handed a renewal quote with E7 on it, this is not a routine upgrade. It is a deliberate framing exercise from Microsoft about which layer of the agent stack you actually need to govern, and where you should be paying to do it. The answer Microsoft is selling — and the counter-argument Red Hat made at KubeCon Amsterdam yesterday — sit at very different layers of the same architecture diagram. Both can be right. But you cannot afford to standardize on both.
What Actually Ships Today
E7 is a bundle. The components have been sold separately — sometimes for years — but as of today they are buyable as a single SKU at a meaningful discount.
The four pieces:
Microsoft 365 E5 — the productivity, security, and compliance foundation. Word, Excel, PowerPoint, Teams, Defender, Purview, the works. Sold separately at roughly $60 per user per month.
Microsoft 365 Copilot — the AI assistant layer integrated into the Office apps and Teams. Sold separately at roughly $30 per user per month. Now multi-model via the Frontier programme: OpenAI models in default mode, Anthropic Claude available alongside through opt-in for tasks where Microsoft's routing prefers it.
Microsoft Entra Suite — Zero Trust identity, conditional access, and network access controls. Sold separately at roughly $12 per user per month.
Agent 365 — the new piece. The agent control plane. $15 per user per month standalone, included in E7.
Add the four à la carte and you get to roughly $117 per user per month. E7 lists at $99. The $18 bundle discount is meaningful at scale, but it is not the load-bearing reason Microsoft is shipping the SKU. The reason is that Microsoft wants the procurement decision for Agent 365 to be a line item on the same renewal as E5, not a separate AI committee evaluation that gets stuck.
That is a sales motion, not a technical decision. CIOs need to evaluate the technical merit of Agent 365 separately from the bundle math. The bundle math will look attractive in a procurement deck regardless.
Agent 365, Decoded
Agent 365 is, in Microsoft's framing, the control plane for AI agents. In practical terms, it provides four capabilities that did not previously have a unified Microsoft surface.
Agent Registry. A central inventory of every agent operating against an organization's Microsoft 365 tenant — including agents built in Copilot Studio, agents created by power users in Teams, and crucially, third-party agents from any vendor that publishes through Microsoft 365 channels and registers with an Entra Agent ID. The registry can also discover unsanctioned "shadow agents" that show up against tenant data without going through procurement.
Entra Agent ID. Each agent gets a first-class identity in Entra, separate from the human user it acts on behalf of. This is the load-bearing piece. Service accounts, the previous tool of choice for non-human identity, were never designed for entities that spawn sub-tasks, accumulate consent grants, and operate semi-autonomously. Entra Agent ID supports least-privilege access, conditional access policies, and lifecycle management as native primitives.
Purview integration. Agents inherit DLP, audit trails, and retention policies from the existing Purview footprint. The audit trail an EU AI Act regulator will eventually ask for can be produced from the same console that already produces the audit trail for human users.
Defender integration. Behavioral signals from agents — anomalous tool invocations, unusual data access patterns, unexpected sub-agent spawns — flow into the same SOC tooling the organization already operates.
The technical architecture is solid. The qualifier is that Agent 365 governs agents that operate against the Microsoft 365 surface. It does not, by itself, govern agents that operate purely against AWS, Salesforce, or homegrown infrastructure unless those agents are registered through Microsoft 365 channels. The vendor-published, Entra-registered agents get first-class treatment. Everything else is a discovery and approximation problem that Microsoft is solving but is not yet equivalent to native coverage.
What Agent 365 Is Not
Read the licensing fine print before you build the procurement model.
Agent 365 covers governance. It does not cover agent execution. Building and running an agent still requires Copilot Studio capacity or Microsoft Foundry credits, both of which are billed separately on consumption. The seat-based $15 per user per month buys you the registry, the identity primitives, the audit trail, and the security integrations. It does not buy you the right to run a single token of agent inference.
That distinction creates a hybrid billing model that procurement teams need to model carefully. Agent 365 appears on the M365 invoice as a per-seat line. Agent execution shows up on the Azure invoice as consumption. Different teams typically own the two invoices. Without explicit attribution work, the total cost of running governed agents at scale is not visible in either bill alone.
For AI executives, the practical implication is that the Agent 365 line in next year's budget should not be evaluated against the cost of doing nothing. It should be evaluated against the cost of running agents through any other governance layer — and against the consumption costs that will sit underneath whichever governance layer you pick.
The Frontier Programme Wedge
The other shift inside E7 is that Microsoft 365 Copilot, as of Wave 3, is genuinely multi-model. Anthropic's Claude is now available in mainline chat through the Frontier programme, alongside OpenAI's latest generation. Copilot's routing layer picks the model for each subtask based on capability and cost.
This is a strategically subtle move. Microsoft has spent seven years selling Copilot as the productized OpenAI experience. Bringing Claude into the same surface — and pricing it inside E7 rather than as an upcharge — accomplishes three things at once. It hedges Microsoft's dependency on OpenAI's roadmap. It blunts Anthropic's enterprise sales motion by making Claude available without a separate procurement event. And it positions Copilot as a model-neutral surface for the moment, three to five years out, when frontier model performance has commoditized enough that vendor choice matters more than capability ceiling.
For CIOs whose AI committees are currently running parallel evaluations of Claude and ChatGPT Enterprise, E7 changes the question from "which vendor do we standardize on" to "do we still need to standardize at all if Copilot routes to both?" That is a question with implications for at least two large procurement processes per enterprise.
The Control Plane Argument, Restated
Yesterday in Amsterdam, Red Hat made the case that Kubernetes is the actual control plane for enterprise AI — that the agentic stack runs on infrastructure your platform team already operates, and that AI-native control planes pitched by the cloud vendors will eventually collapse into Kubernetes-native primitives.
Microsoft, with E7 and Agent 365, is making a different argument. The control plane that matters, in Microsoft's framing, is the identity and governance layer where humans, agents, devices, and data already meet — the layer Microsoft has spent fifteen years building inside Entra and Purview, and which agents now plug into as first-class citizens.
Both arguments are partly right and they describe different layers of the same architecture diagram.
Red Hat's control plane is the workload layer. It answers questions about scheduling, isolation, multi-tenancy, sovereign deployment, and inference routing. It is the right answer for organizations whose platform team operates Kubernetes well and whose AI footprint is heterogeneous across clouds and on-prem.
Microsoft's control plane is the identity and governance layer. It answers questions about who an agent is, what it is allowed to access, what audit trail it produces, and how it shows up in the same SOC and compliance dashboards as the rest of the workforce. It is the right answer for organizations whose data and identity center of gravity is already inside Microsoft 365.
The mistake CIOs are about to make in 2026 — and the mistake every analyst note will reinforce — is treating these as competing answers to the same question. They are not. The Kubernetes layer cannot answer "is this agent allowed to read this email" without an external identity system telling it who the agent is. The Agent 365 layer cannot answer "where should this inference run" without an external workload layer telling it which fleet has capacity. The enterprises that get the architecture right will be the ones that buy a credible answer at each layer and refuse to let either vendor pretend it is the whole stack.
What CIOs Should Do This Quarter
For enterprise AI engineers, three immediate priorities:
Stand up Entra Agent ID for every agent in production, regardless of whether you have committed to Agent 365 yet. Identity is the precondition for everything else, and the work to assign agent IDs is reusable across whichever governance layer you eventually standardize on.
Push back on the bundle math. The E7 sticker discount is real, but $99 per user per month across a 50,000-seat enterprise is $5 million per month. The number you are actually evaluating is closer to $60 million annual incremental spend on top of existing M365 commitments. Bundle discounts do not justify that decision; agent governance value does. Make the value case explicitly.
Map Agent 365 coverage gaps before you sign. Identify which agents in your environment will not be governed natively — anything running purely against AWS, GCP, Salesforce, or homegrown infrastructure that does not register through Microsoft 365 channels — and decide whether you accept the discovery-and-approximation coverage Agent 365 provides, or whether you need a parallel governance layer for the non-Microsoft footprint.
For AI executives, the strategic priority is the same one that mattered yesterday and will matter tomorrow: do not let any single vendor convince you that their control plane is the control plane. The agent stack has at least three layers — workload, identity-and-governance, and runtime — and at most one of them is going to consolidate around a single vendor in 2026. Buying as if all three already have is the procurement mistake that makes 2027 expensive.
The Bottom Line
E7 is a meaningful product launch and Agent 365 is a credible governance layer. Both ship today. The bundle math will pencil out for organizations whose Microsoft 365 footprint is large enough to absorb the per-seat increment. For those organizations, the decision is straightforward: E7 if E5 is already the baseline, Agent 365 standalone if not.
The harder decision is architectural. Microsoft is using today's GA to draw a control-plane boundary that runs through identity and governance. Red Hat drew a different boundary yesterday that runs through workload and infrastructure. Google is drawing a third one through its Gemini Enterprise platform. AWS is drawing a fourth through Bedrock AgentCore. Salesforce, ServiceNow, and Snowflake are each drawing fifth, sixth, and seventh ones through their respective platforms.
Five of these will eventually be wrong. Two will turn out to be load-bearing. The CIOs who get the next two budget cycles right will be the ones who refuse to commit prematurely to a single answer — who buy E7 for the identity layer if it fits, who run llm-d on Kubernetes for the workload layer if their platform team can operate it, and who keep the question of which AI-native control plane wins open until the market gives them a less ambiguous signal than vendor announcements on launch days.
Today's signal from Microsoft is that the identity and governance layer is the layer the company believes will matter most. It is a credible bet. It is not the only one being placed.
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.
