Managed agents are here. Where are you?

In late 2023, Microsoft 365 Copilot launched as a $30/user/month add-on to E3 or E5. Two and a half years later, the licensing surface around "AI in your tenant" looks nothing like that. There is Copilot the productivity surface, Copilot Studio the agent builder, autonomous agents that consume Copilot Credits per run, message-pack pricing for chat workflows, agent runs counted against Power Platform, and the still-not-quite-clear distinction between Microsoft 365 Copilot Chat (free for E3/E5, gated to web context) and Microsoft 365 Copilot (paid, grounded in Microsoft Graph). Then comes the new licensing tier that Microsoft has been previewing — bundling agents, Copilot Studio capacity, and prebuilt agent templates into a higher-touch SKU that is being positioned as the default for any organization "serious about agents."

If you are a CIO or IT director walking into your next Microsoft EA renewal, you are being asked to sign on a tier whose architecture you have not yet validated. That is the procurement-architecture inversion driving most of the strategic mistakes in enterprise AI procurement today.

The licensing landscape, plainly

Strip the marketing. The current Microsoft AI licensing surface has six things you actually buy, and they stack:

• Microsoft 365 E3 / E5 — the existing per-user productivity license. E5 includes Defender, Purview, Audio Conferencing. Required as a base.
• Microsoft 365 Copilot ($30/user/month) — the add-on that grounds Copilot in your Microsoft Graph: SharePoint, Outlook, Teams chat, OneDrive. Without this, "Copilot" runs against generic web context only.
• Copilot Studio — the agent-building platform. Licensed via message packs (formerly capacity-based, now per-message under Copilot Credits). One agent run consumes a defined number of credits depending on what tools it invokes.
• Power Platform per-user / per-flow plans — required if your Copilot Studio agents call Power Automate flows, Dataverse, or custom connectors at meaningful scale.
• Azure AI Foundry — pay-as-you-go billing for any model not bundled in Copilot itself: GPT-4o on Azure, Phi-4, the Mistral and Llama families on AI Foundry, and any custom fine-tunes.
• The new agentic tier — bundling agents, Copilot Studio capacity, and selected agent templates into a higher SKU. Positioned as "agents-ready out of the box." In practice: a procurement compression of the four lines above.

Microsoft Learn has the canonical breakdown of the Copilot Studio licensing model. It is dense, it changes quarterly, and the per-Copilot-Credit unit costs are the single most important number you negotiate on at renewal.

The mistake we see most often: treating this as one licensing decision. It is six, and they interact. An agent built in Copilot Studio that calls a Power Automate flow against Dataverse and routes to Azure OpenAI for the heavy reasoning hits four billing meters every time it runs. Most teams do not know that until the first invoice lands in month two.

The procurement-architecture gap

Microsoft account teams, Anthropic enterprise sellers, AWS Bedrock specialists — they are all paid to close the SKU before the architecture is done. That is not a moral failing; it is how vendor sales works. The buyer's defense is to sequence the conversation correctly:

1. Identify the cohort first. Which 200, 500, or 2,000 users are going to use agents? What for? Until you can name three concrete workflows per cohort, you are not ready to license.
2. Map the data plane second. What sensitivity labels exist on the data those workflows touch? What DLP enforcement is in place? Who owns the SharePoint sites the agents will read? If the answer is "we'll figure that out later," your agents will leak before they ship.
3. Then — and only then — model the licensing. Per-cohort message volume × Copilot Credit consumption × frequency. That math tells you whether the new agentic tier is genuinely cheaper than stacking the components, or whether it is a procurement compression that costs you more.

We have walked into engagements where the licensing was already signed, the Copilot Studio capacity already provisioned, and the cohort and governance work were the things being figured out post-hoc. In every case the cost trajectory got reset within two quarters, and frequently the licensing tier got renegotiated mid-term because the agents being built did not match what had been licensed.

What "managed" actually means

The word managed is doing work in the phrase "managed agents." Strip it down: a managed agent is an agent definition that lives in Microsoft's (or Anthropic's, or AWS's) infrastructure, accessed through their SDK and their billing meter, with the conversation state, tool execution, and skill loading all handled server-side by the vendor.

The trade-offs are real and often underweighted:

• You do not own the runtime. When Microsoft adjusts how Copilot Studio messages are counted, your unit economics shift. We have seen this happen twice in eighteen months.
• You are coupled to the vendor's update cadence. A skill or prompt change Microsoft pushes to Copilot's underlying foundation model can change your agents' behavior overnight.
• Audit logging is what the vendor decides to expose. Microsoft 365 audit logs cover Copilot interactions reasonably well, but the granularity is theirs to tune.
• Exit cost compounds. Migrating an agent built in Copilot Studio to a self-hosted Anthropic Claude or OpenAI Assistants stack is a rebuild, not a port.

This is not an argument against managed agents. The trade for most enterprises — speed of deployment, integration with Microsoft Graph, no ML platform team to staff — is favorable. It is an argument for knowing what you are trading, and pricing the lock-in into your renewal posture.

The three procurement questions before you sign

Before the next renewal:

1. Per-cohort value math. For each cohort that will get agents, what is the measured outcome — drafting velocity, ticket deflection, time saved on routine reasoning? At what license rate per seat does the math fail? That number is your renegotiation floor.

2. Governance readiness gate. Is your Purview posture ready to absorb agentic data flow? If sensitivity labels are not published and DLP is not enforced, no licensing tier saves you from the post-launch governance scramble. (We have written about the Purview posture problem separately.)

3. Exit posture. If Microsoft adjusts the Copilot Studio message-pack pricing 30% upward at renewal — which has historically been within their pricing flexibility — what is your migration cost? If the answer is "we'd be stuck," your negotiation leverage is zero.

How we sequence this work at Protime

Our intake on Microsoft licensing engagements starts with the cohort map, not the SKU. Within the first two weeks we have a written matrix of who-uses-what-for-what, what data those workflows touch, and what Microsoft constructs the agents will use. From that, the licensing tier becomes obvious — not chosen first.

The work after that is unglamorous and decisive: oversharing audit on the SharePoint content the agents will read, sensitivity-label gap analysis, DLP enforcement modeling, Copilot Credit consumption forecasting per cohort, and the Power Platform integration cost line that does not show up in the Microsoft Copilot SKU itself.

When the renewal conversation happens, your account team is talking to someone who knows what their math actually says. That changes the conversation.

---

Managed agents are here. The question is not whether you have them — Microsoft has shipped them and the licensing is on your renewal sheet whether you have asked for it or not. The question is whether you walk into procurement with your cohort and governance work done, or with the vendor's slide deck open in front of you. We help with the first.