AI Agent Pricing Models Explained: Seats, Credits, Tasks and Outcomes

AI agent pricing models are the ways vendors charge for software that carries out work on its own: per user seat, per credit consumed, per task or conversation, or per successful outcome such as a resolved support ticket. Unlike a chatbot subscription, an agent's cost depends on how much work it does, how many tools it calls and which model it uses, so two products with similar headline prices can produce very different bills. Many vendors now mix several models in one offer, for example a seat licence that includes a monthly credit allowance with metered overage. This page explains each pricing model, how the units are defined in published vendor pricing, how to estimate and compare costs, and how to control spend. It is for buyers, finance and FinOps teams, and engineers who have to forecast agent costs.

Last checked: 2026-11-15. Vendors and products are listed alphabetically; order is not a ranking. Prices are quoted in US dollars as published on each vendor's own page on the date checked, and change often.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Four base models, many hybrids. Seats, credits (consumption), per task or conversation, and per outcome are the building blocks; most published offers combine two or more.
• A credit is a vendor-defined unit. One credit can mean a fixed dollar value, a share of tokens or a weighted count of actions, so credits from different products are not comparable without converting to cost per task.
• Outcome pricing needs a definition you can audit. Per-resolution prices only make sense once you know what the vendor counts as a resolution and how you can dispute one.
• Allowances usually expire. Several vendors state that unused credits do not roll over, so oversized commitments are paid for whether or not they are used.
• Cost control is an identity problem too. Agents spend through their own service identities and API keys, so budgets, quotas and alerts need to be set per agent identity, not just per account.

What are the main AI agent pricing models?

Each model allocates risk differently between buyer and vendor. Seats put usage risk on the vendor; consumption and per-task pricing put it on the buyer; outcome pricing ties cost to a result the buyer values but makes the definition of that result the main negotiating point.

How do published vendor prices use these models?

The table shows how six vendors describe their agent pricing on their own pricing or documentation pages. It records units and published list prices only; it is not a comparison of value, and enterprise contracts often differ from list prices. As of 15 November 2026.

For model API token prices see the LLM Pricing Tracker; for consumer and business chatbot plans see AI Chatbot Subscription Plans Compared; for coding tools see the AI Coding Tools Pricing Tracker.

Seat-based pricing

A seat licence charges per named user, usually per month on an annual term. It suits assistants that employees use throughout the day, because cost scales with headcount rather than with activity. Seat pricing for agents increasingly includes limits: a monthly credit allowance, a usage quota or fair-use terms, with metered usage beyond it. When reading a seat offer, check:

• What a seat allows: some vendors distinguish full and lighter seat types with different rights.
• Whether agents need seats: Intercom lists Fin as requiring no seats; Salesforce offers both per-user add-ons and consumption options.
• What happens at the limit: GitHub states that once the credit allowance is used, users can upgrade, set a budget for additional usage or wait for the monthly reset.

Seats fit poorly where agents work without a human user, such as overnight batch agents or customer-facing agents. A seat model applied to autonomous agents can also hide how much work they actually do.

Credit and consumption pricing

Credits are a prepaid or metered currency that different activities draw down at different rates. They let a vendor price many activities (a classic answer, a generative answer, a tool action, a voice minute) under one balance. Three things define a credit system:

1. The value of a credit: GitHub publishes 1 AI credit = $0.01; Salesforce publishes $500 per 100,000 Flex Credits.
2. The burn rate per activity: Microsoft publishes a rates table (for example 2 Copilot Credits per generative answer and 5 per agent action); GitHub ties consumption to tokens and the model used.
3. The enforcement rule: what happens when credits run out. Microsoft states that usage beyond purchased capacity is subject to technical enforcement that can deny service, and that agent flow enforcement blocks new flow runs when no credits are available.

Credit models also differ on model choice: on token-based systems, a more capable model or a longer agent session uses more credits for the same request. GitHub notes a 10% discount on model costs for paid plans that use auto model selection.

Per-task, per-action and per-conversation pricing

Here the unit is an interaction the buyer can count. Salesforce lists $2 per conversation, designed mainly for external, customer-facing agents, and states that Flex Credits and Conversations cannot be used in the same org. Microsoft's agent action rate and Salesforce's credit-per-action examples are also task-level prices expressed in credits. The questions to settle are when a task or conversation starts and ends, whether retries or tool calls inside a task count separately, and how abandoned sessions are billed.

Outcome-based pricing

Outcome pricing charges only when the agent achieves a defined result. Zendesk describes AI agent pricing as based on the successful outcomes the agents deliver and bills per automated resolution, charged only for requests resolved without escalation to a human agent. Intercom prices Fin from $0.99 per Fin outcome. The appeal is alignment: failed attempts are not billed. The risk is definitional:

• Definition: what counts as resolved (customer confirmation, no reply within a time window, no escalation)?
• Reopens: is a ticket that comes back within days refunded or billed twice?
• Audit: can you export the list of billed outcomes with conversation IDs and dispute individual entries?
• Incentive: a pure "no escalation" definition can reward an agent that makes escalation hard, so pair billing with customer satisfaction and reopen-rate checks.

How to compare and estimate agent costs

Because units differ, convert every offer to the same measure: cost per successfully completed task at your expected volume. A simple method:

1. Define the task. Pick a unit of work you can count in your own systems (a resolved ticket, a merged pull request, a processed invoice).
2. Measure volume. Use last year's monthly volumes, including peaks.
3. Estimate automation rate. The share of tasks the agent will complete without a human, from a pilot rather than vendor claims.
4. Price each model. For seats: seats times price. For credits: tasks times average credits per task times credit value, plus failed attempts, which still consume credits. For outcomes: automated tasks times price per outcome.
5. Add the fixed parts. Platform fees, required base licences, minimum commitments and implementation.
6. Add the human cost of the remainder. Tasks the agent does not complete still need staff time.
7. Stress test. Re-run at double the volume and at half the automation rate to see which model breaks your budget first.

The figures above are invented to show the arithmetic and are not any vendor's prices. Note how the cheapest model changes with the automation rate and the number of attempts per task.

Choosing a pricing model

• Small teams with steady use: seats or plans with included allowances give predictable bills; check what happens at the limit.
• High-volume customer service: outcome or per-conversation pricing aligns cost with value, provided the outcome definition is auditable.
• Builders and platform teams: credit or token-based pricing gives flexibility across many agents; it needs metering, budgets and alerts from day one.
• Uncertain demand: pay-as-you-go first, then commit once a few months of usage data exist, because several vendors state that unused prepaid credits expire.
• Large enterprises: negotiate rate cards, burn rates per activity, rollover terms and a right to audit billed units; list prices are a starting point.

Controlling spend

Agents can loop, retry or be driven by attackers into running up costs. The Denial-of-Wallet and AI Agent Cost Runaway Guide covers the threat side in depth; the commercial controls are:

1. Set hard budgets where the vendor offers them, such as GitHub's USD budget for additional usage or a capped Azure billing policy for Copilot Studio pay-as-you-go.
2. Meter per agent and per environment so a single runaway agent is visible.
3. Alert on burn rate, not only on totals, so a spike is caught within hours.
4. Choose models deliberately, since credits per request rise with model capability on token-based systems.
5. Review commitments quarterly against actual usage to avoid paying for expiring credits.

Where identity fits

Every agent spends money through an identity: a service account, an API key, an OAuth client or a licensed user it acts for. That makes identity the practical place to attach cost controls. Give each agent its own identity rather than a shared key, so usage and spend can be attributed; scope quotas and budgets to that identity; and revoke or suspend it as the fastest way to stop runaway spend. Billing data is also a detection signal: an unexpected rise in credits consumed by one agent identity can indicate a leaked key or a hijacked agent. See the Agent Identity and Access Guide and FinOps Security.

For product-specific security settings see the Devin Security, GitHub Copilot Security Guide, Intercom Security Guide, Securing Microsoft Copilot Studio Agents Guide, Securing Salesforce Agentforce Guide and Zendesk Security Guide.

What should you measure?

• Cost per completed task: total agent spend divided by tasks completed without a human.
• Units per task: average credits, tokens or actions per completed task, and its trend.
• Allowance utilisation: share of included or prepaid credits used before expiry.
• Billed outcomes later reopened: for outcome pricing, the share of billed resolutions that come back.
• Spend by agent identity: top agents by cost each week, with owners.
• Forecast accuracy: actual against forecast spend per month.

Common pricing mistakes

• Comparing credit prices across vendors without converting to cost per task.
• Forgetting that failed and retried attempts still consume credits or tokens.
• Buying large prepaid commitments before usage data exists, then losing unused credits at term end.
• Accepting an outcome definition without a way to export and dispute billed outcomes.
• Sharing one API key across many agents, so spend cannot be attributed or capped.
• Leaving pay-as-you-go billing with no budget or alert attached.

Practitioner checklist

• Write down the billing unit, its definition and its dollar value for every agent product you use.
• Convert each offer to cost per completed task at current and doubled volume.
• Confirm rollover, overage and enforcement rules in the vendor's own terms.
• For outcome pricing, obtain the outcome definition and an export of billed outcomes.
• Give each agent its own identity and attach quotas, budgets and alerts to it.
• Set hard budgets on every pay-as-you-go or overage arrangement.
• Review spend by agent identity weekly and commitments quarterly.
• Re-check list prices before each renewal; vendor pricing pages change often.

Standards and references

• Devin: Pricing
• GitHub: Usage-based billing for individuals
• GitHub: Copilot individual plans
• Intercom: Pricing
• Microsoft: Copilot Studio billing and licensing
• Microsoft: Copilot Credits and billing rates
• Salesforce: Agentforce pricing
• Zendesk: Pricing

Related NHI Mgmt Group resources: Model Pricing Explained · LLM Pricing Tracker · Denial-of-Wallet and AI Agent Cost Runaway Guide · AI Agent Platforms Compared