For Microsoft customers, however, one distinction changes the discussion: Agent 365 licenses people for agent oversight and use—not every temporary helper as a separate monthly employee. Microsoft’s published commercial structure is more nuanced than the idea of putting every agent on the subscription payroll.
The useful question is therefore not simply whether seats or meters are fairer. It is what each charge buys, how the bill responds to workload growth, and what a successfully completed business task actually costs.
Microsoft’s “seat opportunities” remark was not a rate card
At the UBS Global Technology Conference on December 2, 2025, Rajesh Jha said Microsoft was not seeing AI reduce Microsoft 365 seats. He described a possible future with more agents than people, including agents with identities, mailboxes, and participation in workplace communications. His conclusion: “To me, all of those embodied agents are seat opportunities.”
That was a strategic outlook, not a published licensing rule requiring every spawned process to buy Microsoft 365. The distinction matters because an enduring workplace agent and a short-lived helper are not necessarily the same commercial object.
Microsoft Learn confirms that Agent 365 became generally available for commercial customers on May 1, 2026, with per-user licensing. Microsoft’s launch announcement lists the standalone price as $15 per user per month, covering an individual who manages or sponsors agents, or uses agents to work on their behalf.
Consequently, the four-agent illustration does not establish a four-seat Agent 365 bill. Buyers should identify the people requiring coverage and separately establish the execution costs of the platforms running their workflows.
Agent 365’s governance price should not be mistaken for an all-inclusive execution price—or used to assume that every governed agent necessarily incurs both Copilot Studio and Azure charges.
Salesforce shows why the comparison needs more columns
Salesforce’s published Agentforce pricing illustrates several distinct purchasing models rather than one universal combination of a seat and a meter.
| Published option | Listed price | Important distinction |
|---|---|---|
| Flex Credits | $500 per 100,000 credits | Standard Agentforce actions consume 20 credits, equivalent to $0.10 |
| Conversations | $2 per conversation | A conversation-based consumption option |
| Agentforce User License | $5 per user per month | Requires Flex Credits; employee usage is metered |
| Agentforce add-ons | $125 per user per month | Includes unmetered employee Agentforce usage within the applicable offering |
These are Salesforce’s listed options, not interchangeable entitlements. Its pricing page identifies product-specific availability and warns that detailed pricing is subject to change.
Two corrections follow. First, $125 is not the entry price for every Agentforce user license. Second, not every option necessarily adds a usage meter to the same employee activity. The bill depends on which offering is purchased and what work falls within it.
Salesforce also states that its illustrative consumption calculations exclude other costs, such as Data 360 credits or other consumption services. A tidy action-price example is therefore not automatically a complete business-case calculation.
Pricing can influence architecture—but not mechanically
Bustamante argues that per-agent seats could encourage customers to consolidate work into larger, less specialized agents. He also identifies messy enterprise documents and missing data as obstacles to production deployment. His perspective deserves its commercial context: Bem sells function-call-based processing, so he has an interest in alternatives to seat licensing.
The architectural conclusion should remain conditional. A contract that charges for each persistent agent could discourage decomposition. A contract that charges for actions could instead encourage fewer billable steps. Neither proves that specialist agents or consolidated designs will always be better.
Our practical assessment is to evaluate cost and correctness together. Removing a validation step may make a workflow cheaper while making its results worse. Combining several tasks may reduce one licensing expense while complicating testing and maintenance. Those trade-offs belong in an architecture review, not just a procurement spreadsheet.
The price list can influence the design. It should not become the architect.
Measure an outcome, not merely an invocation
For a pilot, we recommend defining one acceptable result before estimating savings: a correctly reconciled invoice, a resolved support case, or a document extraction that passes specified validation checks.
Then use a working calculation:
Cost per accepted outcome = allocated fixed costs + execution costs + exception-handling costs, divided by accepted outcomes.
This is a proposed evaluation method, not a vendor billing formula. Apply it consistently:
- Define acceptance. Specify the checks required before a task counts as complete. An attempted answer should not automatically become a successful outcome.
- Record the workflow. Assign a business-task identifier and associate agent activity, billed operations, retries, and human interventions with it.
- Separate cost layers. Track access and governance licenses separately from execution charges, supporting services, and human review.
- Test difficult inputs. Include incomplete documents, conflicting records, and cases requiring escalation—not just demonstration-friendly examples.
- Model several volumes. Compare ordinary demand, peak demand, and quiet periods. State assumptions about commitments and unused capacity.
- Reconcile estimates with invoices. Establish whether actual billed units match the workflow’s recorded activity before expanding deployment.
This approach gives finance and engineering a shared unit of comparison. It also prevents a lower price per call from masquerading as a lower price per completed task.
Governance is a separate deliverable
Microsoft describes Agent 365 as a centralized way to observe, govern, and secure agents. Its documented capabilities include an agent registry, lifecycle and access management, and integration with Microsoft Entra, Purview, and Defender. Those functions address oversight and risk; they are not simply another name for model execution.
For consequential workflows, we recommend a separate evidence test. Ask the vendor to demonstrate which records are available for a completed decision: input references, actions taken, relevant configuration versions, timestamps, and human approvals. Specify retention, access, and export requirements.
Do not equate an activity log with a complete explanation of a business decision—or demand an AI model’s private internal reasoning as an audit artifact. The goal is reviewable operational evidence.
Before signing, put four questions into the commercial evaluation:
- What exactly is billable, including helpers, testing, retries, and failed operations?
- How does cost change when workload doubles but staffing does not?
- What evidence can we retrieve when an outcome is challenged?
- What is the measured end-to-end cost of an accepted result?
The 90-second agent is a useful stress test for licensing assumptions. It is not proof that Microsoft sells four seats for that workflow. The stronger procurement lesson is to distinguish the person, the agent, the execution, and the outcome—before the invoice bundles them into one expensive surprise.