Agent 365, three SKUs and a meter that counts actions
Agent 365 is Microsoft's framework for governing AI agents inside the Microsoft 365 tenant, across Copilot agents, custom Copilot Studio agents, Agent Store agents, and platform agents. The licensing runs three SKUs, per user for human centric agents, per agent metered for autonomous ones, and a tenant level governance plan, and the meter counts actions, not seats, which is where every forecast broke.
Prepared by Redress Compliance · August 7, 2026 · Microsoft advisory. Based on 15 to 25 Agent 365 and AI agent engagements advised 2024 to 2025.
Executive summary
Three SKUs cover three populations, and mixing them up is the first overspend. Per user assignment covers agents running in a single user's context, stacking with Copilot per user pricing; per agent metered covers autonomous, asynchronous, scheduled, and multi user agents, billed per Copilot Message unit in prepaid packs; and the Agent 365 plan runs the governance layer at tenant level with per managed agent elements. Most estates over license the per user SKU and under size the message pool, the exact inverse of where the cost actually lands.
The message meter compounds by construction. A simple query is one message, multi step reasoning three to five, each tool call adds one, an agent to agent handoff costs two at each agent, and long context retrieval carries a surcharge. The worked example shows the scale: a nightly compliance agent across two thousand records, three reasoning steps and two tool calls each, consumes roughly ten thousand messages a night, 3.6 million a year, about $360,000 at typical rates before discount, from one agent.
The forecasts broke low and the entitlement assumptions broke expensive. Agent message and capacity consumption overran the prepaid pool by 20 to 40 percent across our engagements, because seat based intuition prices actions badly; and Agent 365 was assumed included in existing Copilot or E5 rights when it is a separate entitlement, adding 10 to 25 percent of unbudgeted cost when the assumption met the invoice.
Governance scoped after rollout is governance priced at true up. The Agent 365 plan covers identity, audit, lifecycle, sensitivity, DLP, and data loss boundaries across the whole agent estate, broader than most buyers assume, and on most estates it was scoped after the agents shipped rather than before. The renewal levers are the meter's standard set: the message pool unit price and the per agent metered rate capped at signing, inside the M365 line of the Enterprise Agreement where the SKUs land.
The message meter, decoded action by action
| Action type | Message cost | The note |
|---|---|---|
| Simple LLM query | 1 message | Single round trip, single agent |
| Multi step reasoning | 3 to 5 messages | Each reasoning step adds |
| Tool use | 1 message per tool call | Additional, on top of the reasoning |
| Agent to agent handoff | 2 messages per handoff | Counted at each agent involved |
| Long context retrieval | 2 messages plus a surcharge | Large context windows cost more |
| A scheduled run | The sum of all of the above | Multiplied by the schedule, forever |
The three SKUs, mapped to the agent estate
- Per user assignment: agents in a single user's context, the Copilot chat experience and personal productivity agents, priced per user per month and stacking with Copilot pricing at signing.
- Per agent metered: autonomous, asynchronous, scheduled, and multi user agents, billed per Copilot Message unit from prepaid packs at a defined rate, the SKU the forecast discipline exists for.
- The Agent 365 plan: the tenant level governance layer, identity, audit, lifecycle, sensitivity, and DLP across all agents, an annual tenant fee with per managed agent elements, and a separate entitlement from Copilot and E5.
The Microsoft EA renewal playbook
The agreement the Agent 365 SKUs land inside: the M365 line placement, the meter caps, and the persona arithmetic the agent estate rides on.
Get the white paper →The right sizing, inverted from where estates start
The recurring misallocation runs both directions at once: the per user SKU over licensed, assigned broadly on the same instinct that over assigns every seat product, while the message pool under sizes because nobody ran the action arithmetic on the agent roadmap. The correction inverts it, per user assignment against demonstrated single user agent use, the pool sized bottom up from scheduled agents at their measured weights, and the governance plan scoped before rollout rather than priced after it. The commercial protections are the consumption meter's standard set, the pool unit price and the metered rate capped for the term at signing, negotiated inside the EA where the SKUs sit. The seat side economics next door, Copilot's own idle rates and true cost per active user, are worked in the Copilot true cost analysis, and the same actions not seats forecasting failure repeats across every platform meter, ServiceNow's among them in the Now Assist consumption analysis.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A negotiation playbook, talking points, and a two page executive brief on day one
What we saw across Agent 365 engagements, 2024 to 2025
Across roughly 15 to 25 Microsoft AI agent and Agent 365 conversations Fredrik Filipsson advised between 2024 and 2025, the licensing model was new enough that buyers struggled to forecast cost:
Agent consumption against the prepaid pool, forecast on seat intuition rather than action arithmetic.
Unbudgeted cost where Agent 365 was assumed inside Copilot or E5 rights it sits outside.
The governance finding was the quietest and the most consequential: identity and lifecycle controls for agents scoped after rollout on most estates, which means agents shipped into a tenant whose audit, sensitivity, and data loss boundaries were retrofitted around them. The sequence that holds runs governance first, the plan scoped and priced before the first autonomous agent ships, the pool sized from the roadmap's action math, the per user SKU assigned against demonstrated use, and the caps written at the EA table where the leverage lives.
Your first five moves
- Run the message arithmetic on every autonomous agent before it ships: actions per run, runs per period, weights applied, because one nightly agent can be $360,000 a year.
- Confirm Agent 365 as a separate entitlement, never assumed inside Copilot or E5, the 10 to 25 percent that arrives unbudgeted otherwise.
- Size the pool bottom up from the agent roadmap and assign the per user SKU against demonstrated single user agent use.
- Scope the governance plan before rollout, identity, audit, lifecycle, and DLP, because retrofitted boundaries price worst.
- Cap the pool unit price and metered rate at signing, inside the M365 line of the EA. The Microsoft practice runs the model with you.
Frequently asked questions
What is Microsoft Agent 365?
Microsoft's framework for governing AI agents inside the Microsoft 365 tenant, spanning Copilot agents, custom Copilot Studio agents, third party Agent Store agents, and platform agents. It licenses through three SKUs: per user assignment for human centric agents, per agent metered for autonomous ones, and a tenant level Agent 365 plan carrying the governance layer.
How does Agent 365 message metering work?
Autonomous and asynchronous agents bill per Copilot Message unit from a prepaid tenant pool: a simple query is one message, multi step reasoning three to five, each tool call adds one, handoffs cost two per agent involved, and long context carries a surcharge. A scheduled run sums all of it, multiplied by the schedule, which is why action arithmetic, not seat intuition, is the forecast.
Is Agent 365 included with Copilot or E5?
No, and the assumption was expensive: Agent 365 is a separate entitlement, and estates that assumed it sat inside existing Copilot or E5 rights added 10 to 25 percent of unbudgeted cost when the invoice corrected them. The governance plan, the metered packs, and the per user SKU each price separately, landing inside the M365 line of the EA.
What does an autonomous agent actually cost?
The worked example frames it: a nightly compliance agent across two thousand records, three reasoning steps and two tool calls each, consumes roughly ten thousand messages a night, 3.6 million a year, about $360,000 at typical per message rates before discount. One agent, one schedule, which is why the arithmetic runs before anything ships.
Why do Agent 365 message pools overrun?
Because they were sized on seat intuition: consumption overran the prepaid pool by 20 to 40 percent across our engagements, with scheduled agents drawing complexity weighted actions on timers regardless of who watched. The pool sizes bottom up from the agent roadmap's action math, and the unit price and metered rate cap at signing before the overrun prices at defaults.
What does the Agent 365 governance plan cover?
The tenant level layer across all agents: identity, audit, lifecycle management, sensitivity labels, DLP, and data loss boundaries, with some elements scaling per managed agent. It is broader than most buyers assume and was scoped after rollout on most estates we advised, which retrofits the controls around shipped agents, the worst order and the standard one.
