AI Knowledge HubAI Agents and the Shift to Real Work: From Monthly Seats to Usage-Based Pricing
The bottleneck for putting AI agents into real work is no longer model capability. It is billing and control. A monthly seat cannot explain the cost or value of an agent that works in the background around the clock. Major vendors are moving to a hybrid model: a platform fee plus usage, action, or outcome-based charges.
Seat Pricing Cannot Explain Background Work
Seat pricing measures how long a person uses an app. An agent, by contrast, keeps searching, summarizing, calling tools, and retrying after failures while the person is away. A single seat can produce widely different amounts of work depending on the task. Seat pricing struggles to recover cost, and customers struggle to predict what an agent will spend in a month.
That is why hybrid pricing is emerging. Vendors charge a base fee for platform access and governance, then meter the actual work in credits, actions, or conversations. Recent price lists from Microsoft and Salesforce show this direction.
Microsoft: Separating Credits from the Control Plane
Microsoft Copilot Studio changed its common billing currency from messages to Copilot Credits on September 1, 2025. Copilot Credits measure the time and effort an agent needs to retrieve information, respond to prompts, and execute actions. More complex work consumes more credits. Linking an Azure subscription turns this into pay-as-you-go billing at the end of each month. Prepaid packs and commitments are also available, but unused credits do not carry over, and exceeding committed capacity can lead to service enforcement.
Agent 365 is different. It is not priced by the work an agent performs. It is a control plane for managing agents, priced at $15 per user per month with an annual commitment. It provides a registry, access control, visibility, interoperability, and security so organizations can observe, govern, and protect their agent fleet. It is also offered with Microsoft 365 E7. In short, Microsoft is charging for work through credits and for governance through seats.
Salesforce: Selling Actions and Conversations
Salesforce Agentforce offers Flex Credits and Conversations. Flex Credits cost $500 per 100,000 credits, and Conversations cost $2 per conversation. Organizations deploying agents for both customer-facing and employee-facing work can choose the unit that fits the task, or combine it with per-user licensing.
The common pattern is clear. Customers pay for the work the agent does, but not through pure metering. A platform and governance fee sits underneath. Pure usage billing creates bill shock; pure seat billing fails to recover vendor costs. Hybrid pricing is the compromise.
The Billing Unit Becomes the Management Unit
Once a billing unit is chosen, the metrics you must operate are chosen too.
| Billing unit | Example | What the customer sees | Best fit |
|---|---|---|---|
| Seat | Agent 365 | Monthly price per user | Enterprise rollout, governance |
| Credit | Copilot Studio, Agentforce Flex Credits | Task complexity | Search, summary, actions |
| Conversation | Agentforce Conversations | Price per conversation | Customer support |
| Token | LLM APIs | Input and output tokens | Development, bulk processing |
| Outcome | Price per resolution | Price per result | Repeatable work with measurable outcomes |
Credit billing requires action logs and retries. Conversation billing requires resolution and escalation rates. Outcome billing requires acceptance criteria and a dispute process. The billing unit becomes the quality-control unit.
The cost per agent task is roughly composed of the following:
- Model tokens: input, output, context reuse
- Retrieval: vector search, web search, internal document lookup
- Tool calls: APIs, databases, MCP servers
- Orchestration: multi-agent branching, retries
- Human review: approval, correction, rework
- Failure cost: wrong sends, duplicate processing, rollback
Looking only at tokens can miss a large share of the real cost. Retrieval, tool calls, retries, and human review must be included to see cost per task.
A Practical Adoption Checklist
| Step | Question | Output |
|---|---|---|
| 1. Define the task | What counts as one unit of work? | Task spec, input and output examples |
| 2. Measure unit cost | Are tokens, retrieval, tools, retries, and review included? | Cost per task |
| 3. Set caps | Are there daily and monthly caps with 80% alerts? | Budget guardrails |
| 4. Add approval gates | Is there human approval for external sends, payments, and deletions? | Approval rules |
| 5. Keep audit logs | Are inputs, tool calls, outputs, and costs recorded? | Audit trail |
| 6. Review contract terms | What about overage, carryover, and data return at exit? | Negotiated clauses |
| 7. Run a pilot | Has one task been tested for two to four weeks? | Success and stop criteria |
Steps 3 and 4 matter most. Usage billing without caps is not a budget; it is an incident. Any agent that can send external email or execute payments needs a human approval gate.
What This Means for Real Deployments
Tasks with clear inputs and outputs, such as public API, statute, and case-law lookups, are easier to price per unit. Breaking work into units like "summarize one notice," "retrieve one case," or "classify one inquiry" fits credit billing well. Broad mandates such as "handle the work autonomously" blur cost, quality, and accountability, and are poor candidates for a pilot.
Buyers should also change their questions. Instead of "how much per seat," ask: "How much per task? Are failed tasks billed? How are retries counted? Is human review time included? What is the overage rate?"
Conclusion: Adoption Is Proven by Invoices and Audit Logs
Putting AI agents into real work does not end with a feature demo. Monthly subscriptions will not disappear, but usage-based charges are being layered on top. The competitive advantage in 2026 and 2027 will not come from the cheapest model. It will come from the operational ability to measure cost per task, enforce caps, and reduce failures.
If you are adopting agents, change the first question from "what can it do" to "what does one task cost." An adoption that cannot answer that question is still a demo.
Prices are based on publicly listed pricing as of October 2026 and may vary by region and contract terms.
Visitor Comments