A Microsoft 365 Copilot agent costs $15 per agent per month under Agent 365 preview pricing, plus message consumption against Copilot Studio at $0.01 per standard message or $0.02 per generative message. A 1,000-agent enterprise fleet, used at 10,000 monthly interactions per agent and three messages per interaction, costs roughly $315,000 per month at list before premium message types are added. The number procurement teams underestimate is not the per-agent fee. It is the long tail of premium message types invoked by agents that touch Dataverse, third-party connectors, and complex Power Fx logic.
This page is the 2026 reference for how Microsoft 365 Copilot agents are licensed, what each line item costs, and where the cost surfaces appear during deployment. The Agent 365 preview terms and Copilot Studio licensing documentation (May 2026) are the source for the unit prices. The deployment math is built from advisor-led enterprise agent rollouts during 2025 and 2026.
What Agent 365 actually licences
Agent 365 is Microsoft's productisation of agent identity, lifecycle management, and governance, launched in preview at Ignite 2024 and moving to general availability in mid-2026. Each agent created on the platform receives an Entra Agent ID, behaves as a first-class identity in the Microsoft 365 tenant, is governed by Conditional Access, audited in Microsoft Purview, and consumes Copilot Studio messages when it runs.
The Agent 365 licence is conceptually similar to a service principal licence in other platforms, but with three differences. First, the agent has its own identity and audit trail, not the identity of the calling user. Second, the agent can be deployed into Teams, SharePoint, and Outlook as a participant. Third, the per-agent cost is metered, not per-seat against the human users invoking it. The model is closer to a robot process automation seat than a Microsoft 365 user seat.
Per-agent base fee
Preview pricing for Agent 365 is $15 per agent per month, with caps for large fleets. Microsoft has signalled but not confirmed that GA pricing will introduce tiers: a higher rate for line-of-business agents that touch external systems, a standard rate for internal productivity agents, and a lower rate for departmental or scoped agents. The current $15 number applies uniformly in preview.
| Agent count | Per-agent monthly base | Annual base (1,000 agents) |
|---|---|---|
| 1 to 100 | $15.00 | n/a |
| 101 to 1,000 | $15.00 (preview, expected to tier in GA) | $180,000 |
| 1,001 to 5,000 | $15.00 (preview) | $180,000 + per-agent above 1K |
| 5,001+ | Negotiated | Bespoke |
The base fee covers the agent identity, Conditional Access participation, Purview audit logging, and the runtime hosting in Copilot Studio. It does not cover message consumption. It does not cover the storage of agent knowledge sources in Dataverse or SharePoint, which is billed against the underlying service.
Copilot Studio message consumption
Agent message consumption is metered against the Copilot Studio message pack. The starter pack is $200 per month for 25,000 messages. Overage is $0.01 per standard message and $0.02 per generative message. A capacity reservation negotiated as part of an EA or MCA-E can lower the effective rate by 15 to 30 percent against committed monthly volume.
| Message type | List rate | Triggered by |
|---|---|---|
| Standard message | $0.01 | Deterministic agent logic, Power Fx evaluation, simple lookups |
| Generative AI message | $0.02 | LLM call to summarise, draft, or extract |
| Dataverse connector message | $0.05 to $0.10 (premium) | Read or write against Dataverse table |
| Third-party connector (premium) | $0.10 to $0.30 | Salesforce, ServiceNow, SAP, custom REST |
| Document grounding (long context) | $0.05 to $0.20 | Large document retrieval, multi-page reasoning |
The bill for a single agent interaction is the sum of every message type invoked in the sequence. A bot that takes a Teams message, queries Dynamics 365 (premium), summarises with an LLM call (generative), and writes a note back to Salesforce (premium) consumes one standard, one premium, one generative, and one premium message. The effective per-interaction cost is $0.10 to $0.40 at list, not $0.02. Across a high-volume use case, the difference is six to twenty times what the back-of-envelope $0.01 number suggests.
1,000-agent enterprise rollout cost
The cost projection for a representative 1,000-agent enterprise fleet, sized at 10,000 monthly interactions per agent (200 working days, 50 interactions per day) and three messages per interaction, illustrates the order of magnitude:
| Cost line | Calculation | Monthly cost |
|---|---|---|
| Per-agent base | 1,000 × $15 | $15,000 |
| Message consumption (standard mix) | 1,000 × 10,000 × 3 × $0.01 (assumed all standard) | $300,000 |
| Premium message premium (30% premium mix) | 1,000 × 10,000 × 0.9 × $0.05 incremental | $450,000 incremental |
| Human Copilot seats (1,000 invoking users) | 1,000 × $30 | $30,000 |
| Total monthly run rate | n/a | $795,000 |
The annualised cost is $9.54M at list. With a 25 percent capacity reservation discount applied to message consumption and a 15 percent EA-bundled discount applied to the Copilot seats, the realised cost lands closer to $7.5M. The headline that most procurement teams quote, "$15 per agent per month," accounts for less than 2 percent of the total. Message consumption is the dominant cost. Premium message types are the single largest line.
The premium message blind spot: Microsoft's pricing communications emphasise the $0.01 per-message standard rate. Premium message types are documented in the Copilot Studio licensing guide but are not surfaced in the pricing tier sheets given to procurement teams. Deployments that connect agents to Dataverse, Dynamics 365, third-party CRMs, or external REST endpoints will run 40 to 70 percent of message volume at premium rates, not the assumed 5 to 10 percent. Build the cost projection at a 40 percent premium mix as the planning baseline.
Consumption forecasting method
The defensible way to forecast agent message consumption has four steps. First, instrument the candidate use cases in a 60-day pilot with a small agent fleet (5 to 20 agents) and capture the message log directly from Copilot Studio. Second, classify the message volume by type (standard, generative, Dataverse, third-party, document grounding). Third, derive the messages-per-interaction multiplier for each use case archetype. Fourth, project the full-fleet volume using the use case mix and headcount.
The two errors that destroy the forecast are: assuming a uniform messages-per-interaction across use cases when reality varies by a factor of 10, and assuming a flat premium mix when the actual mix is dominated by the use cases that touch external systems. The second error is the larger of the two and accounts for most enterprise over-runs against budget.
Governance, identity, and audit
Each Agent 365 agent receives an Entra Agent ID. Conditional Access policies can be applied to the agent, including session controls, location restrictions, and risk-based access denials. Purview audit logs every action the agent takes, including data access, message generation, and system calls. The audit retention is governed by the Purview retention policy on the tenant, not the Agent 365 product.
The licensing implication is that Agent 365 governance assumes a Microsoft 365 E5 tenant or an equivalent Entra Suite, Purview Audit (Premium), and Conditional Access entitlement. Tenants on E3 without the add-ons get a partial governance experience. The realised cost of running agents at enterprise grade therefore includes the E5 base or the add-on stack, in addition to the per-agent fee.
Distribution via Teams, Outlook, and SharePoint
Agent 365 agents are distributed inside Microsoft 365 via Teams (as participants in channels and direct messages), Outlook (as add-ins surfaced in the reading pane), SharePoint (as embedded experiences on team sites), and the Copilot panel in any Microsoft 365 app. The distribution is governed by Microsoft 365 admin policies, including Teams app permission policies and the Copilot deployment scope.
The licensing implication is that agents reach more users than the procurement team typically anticipates. An agent deployed to a single department site can be invoked by anyone with site access. Charging is against message consumption, not against the user invoking the agent, so a low-licence-count agent that becomes popular across a broad user base can produce a runaway message bill before quarter-end. The governance lever is the Copilot Studio capacity reservation, which caps monthly message consumption against a committed budget.
Build versus buy versus first-party Copilot
The decision between building an agent in Copilot Studio versus buying a first-party Microsoft Copilot SKU versus using a third-party agent platform depends on three factors: the data sources the agent must touch, the integration depth required with Microsoft 365, and the volume of user interactions.
| Use case | Best fit | Reason |
|---|---|---|
| Sales call summarisation in Outlook | Copilot for Sales | Pre-built, $50 SKU, no message consumption to manage |
| HR policy Q&A bot | Copilot Studio agent | Custom knowledge sources, low message volume per interaction |
| Cross-system order status (CRM, ERP, fulfillment) | Build outside, integrate via API | Premium message multipliers make Copilot Studio expensive at scale |
| Departmental document Q&A | Copilot Studio agent | SharePoint grounding is native, premium messages avoided |
| External-facing customer service agent | Dynamics 365 Customer Service + Copilot | Licensed for external users, governance built-in |
The pattern is that Copilot Studio agents are most cost-effective for use cases where the data lives inside Microsoft 365 and the interaction volume is moderate. They are least cost-effective for high-volume use cases that touch line-of-business systems, where the premium message multipliers compound. For those use cases, a custom agent built outside Copilot Studio against Azure OpenAI directly is typically 40 to 70 percent cheaper at scale.
Negotiation levers for agent licensing
Three levers move money in agent licensing negotiations. First, the Copilot Studio capacity reservation, which exchanges committed monthly message volume for 15 to 30 percent off the message rate. The risk is identical to any committed-consumption deal: under-utilisation is wasted spend. Second, the Agent 365 enterprise commit, which bundles the per-agent base into a fixed monthly fee for a committed agent count, removing the per-agent metering risk. Third, the bundled EA discount on the human Copilot seats invoking the agents, which is the largest commercial lever when agent fleets are tied to broad Copilot adoption.
The negotiation should focus on the message consumption commitment. The per-agent base is small. The Copilot seat discount is constrained by Microsoft's seat-level pricing discipline. The message commitment is where Microsoft has the most flexibility because the unit economics are most favourable to Microsoft at the standard message rate.
For the broader Copilot framework, see Microsoft 365 Copilot pricing 2026, Copilot Studio pricing, and Copilot licensing strategy. For the EA-level commercial framework, see Microsoft EA complete guide and the Microsoft vendor hub. For agent procurement specifically, see our AI procurement advisory.