How to Price AI Agents: Seat, Tool, or Teammate?

Every AI company is wrestling with the same pricing question right now.

Per-seat pricing worked when software was a tool humans used. But AI agents don't just assist. They resolve tickets, write code, book meetings, and draft legal memos autonomously. When one agent replaces one or more full-time employees, pricing per human seat makes no sense for the buyer or the seller.

The problem is that "usage-based" is not one thing. Charging per-resolution, per-compute-unit, per-conversation, and per-outcome are all usage-based, but they produce wildly different revenue profiles and buyer psychology. Founders need a framework, not just a billing model.

What Is AI Agent Pricing?

AI agent pricing is the billing model a software company uses to charge customers for autonomous AI agents that perform work independently of human users. Unlike traditional SaaS per-seat pricing (which charges for access), agent pricing charges for work performed.

This is a meaningful distinction. Traditional pricing assumes a human operating software. Agent pricing assumes software operating on behalf of a human. The unit of value shifts from "who has access" to "what gets done."

The evolution happened in three waves. For over a decade, seats were the default. Then between 2023 and 2025, credits and usage-based pricing emerged as AI features introduced variable costs that flat subscriptions couldn't absorb. Now in 2026, the market is splitting further based on agent autonomy level, and the per-seat vs usage-based debate has evolved into a more nuanced classification problem.

Why Per-Seat Pricing Breaks for AI Agents

The broad case against pure per-seat pricing in the AI era is well documented. But AI agents introduce three specific dynamics that go beyond the general "seats are dying" narrative.

The "1 Agent Does 5 People's Work" Problem

When an AI SDR handles outbound prospecting for a team of five, charging one seat at $50/month captures a fraction of the value delivered. The buyer pays less than what a single human would cost, while extracting the output of an entire team. The vendor leaves 90%+ of the value on the table.

The "15 Agents" Ceiling

Per-agent seats work when a customer deploys one or two agents. At 15 agents, the math collapses. If each agent seat costs $5,000/month, the customer is paying $75,000/month. That's more than the actual human team would have cost. The model punishes scale, which is the opposite of what software should do.

The Value Asymmetry Problem

Not all agents deliver equal value. A coding agent that ships production code creates more value per hour than a scheduling assistant. A legal research agent handling complex litigation generates more revenue impact than a meeting summarizer. Flat per-agent seats ignore this entirely, treating a $500K-value agent the same as a $5K-value agent.

The challenge with AI agents is that "what feels right" depends entirely on whether the buyer sees your agent as a utility or a colleague. And that perception determines whether $0.10 per action or $5,000 per month is the number that signals quality.

The Seat, Tool, or Teammate Framework

Before picking a billing model, answer one question: what is your AI agent to the customer?

The answer falls into three categories. Each has a distinct pricing architecture, price anchor, and scale profile.

Category 1: Tool

A tool agent performs discrete, repeatable tasks on demand. The user triggers it, it executes, it returns a result. The interaction is transactional.

Pricing model: Per-usage, per-action, per-resolution, or per-credit.

How to recognize a tool agent:

⦿ Users invoke it manually. The agent responds to a specific request each time.

⦿ Output is a single deliverable. A resolved ticket, a translated document, a summarized report, a generated image.

⦿ Value scales with volume. More tasks completed = more value delivered. The agent's ongoing presence between tasks has no value.

Real examples:

• Intercom Fin: $0.99 per resolution. One charge per conversation, regardless of how many questions the customer asks. No charge if the conversation goes unresolved.
• ChatGPT API: Per-token pricing. Input and output tokens charged separately.
• Harvey.ai: Usage-based enterprise pricing for AI legal research queries.

Pricing math: If Intercom Fin resolves 1,000 tickets per month at $0.99 each, the customer pays $990/month. Predictable per-unit economics that scale with actual usage.

Category 2: Teammate (Digital Employee)

A teammate agent fills a specific role on a team, operating continuously or semi-autonomously. It has a "job title" equivalent. Customers talk about it the way they talk about a hire.

Pricing model: Per-agent seat, anchored in FTE cost savings. Price the seat at a fraction (typically 10-30%) of the equivalent human salary.

How to recognize a teammate agent:

⦿ Customers describe it with a job title. "Our AI SDR," "our AI recruiter," "our AI support agent."

⦿ It runs autonomously. No human triggers each task. The agent operates on standing instructions, workflows, or inbound signals.

⦿ It replaces or augments specific headcount. The buyer evaluates the agent against the cost of hiring a person for that role.

Real examples:

• 11x.ai ("Alice"): Roughly $5,000/month per AI SDR agent. Positioned against a human SDR salary of $80,000 to $120,000/year. The buyer saves 25-50% while getting an agent that works 24/7 with no ramp time.
• Devin (Cognition): $20/month Core plan or $500/month Team plan, plus $2.00-$2.25 per ACU (Agent Compute Unit, roughly 15 minutes of active work). A hybrid between teammate and tool pricing.
• Bland.ai: AI phone agents priced per agent deployment for sales and support calls.

Pricing math: A human SDR costs $100,000/year, which is $8,333/month including benefits and overhead. An AI SDR priced at $5,000/month saves the buyer 40%. The seller captures 60x more revenue than a $0.10 per-action model would at typical volumes. Both sides win.

Category 3: System of Agents

A system deploys multiple AI agents working together as an orchestrated platform. The customer doesn't think about individual agents. They think about the system's total output.

Pricing model: Hybrid. Platform fee (for access and orchestration) plus metered or outcome-based billing for actual agent work.

How to recognize a system:

⦿ Customers deploy many agents. Five, ten, or fifty agents running across different workflows.

⦿ Individual agent identity matters less than system output. The value is in orchestration, not individual task completion.

⦿ Per-agent seats at scale feel like paying for internal architecture. If you price every micro-agent as a separate seat, the buyer feels like they are paying for your engineering, not their value.

Real examples:

• Salesforce Agentforce: $2 per conversation or Flex Credits at $0.10 per action (20 credits per action at $0.005/credit). Enterprise customers get 100,000 free Flex Credits. Organizations choose one pricing model.
• Sierra.ai: Outcome-based pricing tied to business results. You pay when the agent delivers a successful resolution, a saved cancellation, or an upsell. Unresolved conversations are typically free. Sierra does not publish pricing; contracts are custom and enterprise-only.

Pricing math: A platform fee of $2,000/month plus $0.10 per action across all agents. At 500,000 actions/month, usage = $50,000, total = $52,000. This scales without the per-agent seat ceiling and aligns cost with actual throughput.

The Framework at a Glance

How Companies Are Pricing AI Agents Today

The framework is not theoretical. Here is how it plays out in production.

Intercom Fin is the clearest tool example. At $0.99 per resolution (with a minimum of 50/month), the pricing directly maps to value delivered. You only pay when the agent successfully resolves a customer's question. No charge for unresolved conversations. We covered Intercom's full pricing structure in our Intercom pricing teardown, including how they run two billing models simultaneously: subscriptions for humans, usage billing for AI.

Devin by Cognition straddles the line between teammate and tool. The $20/month Core plan with pay-as-you-go ACUs ($2.25 each, roughly 15 minutes of work) treats it like a tool. The $500/month Team plan with 250 included ACUs starts to feel like a teammate budget. This hybrid reflects Devin's nature: sometimes a quick coding assistant (tool), sometimes an autonomous engineer shipping PRs (teammate).

When Does an Agent Cross from Tool to Teammate?

The classification is not always obvious. Some agents sit on the boundary. Three diagnostic signals help you determine where your agent falls.

1. Autonomy level. Does the agent need a human to trigger every action, or does it operate on standing instructions? A summarization tool waits for input. An AI SDR proactively reaches out to leads without being asked. Higher autonomy pushes toward teammate pricing.

2. Identity persistence. Does the agent have a name, a role, or persistent context across sessions? If customers call it "Alice" or "our AI recruiter," they are framing it as a colleague. If they say "I used the summarizer," it is a tool. Identity persistence signals teammate.

3. Replacement framing. Do customers describe the agent as "replacing" a person, or as "handling" a task? "We replaced two SDRs with 11x" is teammate framing. "We use Fin to handle support tickets" is tool framing.

If all three signals point to teammate, price as a digital employee. If two or more point to tool, price per-usage. If the answer is "it depends on the customer," you may need to offer both (like Devin does with its Core and Team plans).

A Decision Flowchart for AI Agent Pricing

Use these three questions to find your pricing model:

1. Does your customer deploy more than 3 agents?

If yes: System pricing. Use a platform fee plus metered or outcome-based billing. Per-agent seats will hit the 15-agent ceiling.

2. Does your agent replace a specific job role?

If yes: Teammate pricing. Use a per-agent seat anchored at 10-30% of the equivalent FTE salary.

3. Neither of the above?

Then: Tool pricing. Use per-usage, per-resolution, per-credit, or per-token billing.

Within each category, choose your billing metric:

⦿ Tool: Per-resolution (Intercom), per-token (OpenAI), per-credit (HubSpot), or per-action

⦿ Teammate: Monthly agent seat at 10-30% of equivalent human salary, optionally with usage overages (like Devin's ACU model)

⦿ System: Platform fee + per-action (Salesforce Flex Credits), per-outcome (Sierra.ai), or per-conversation with volume tiers

Common Mistakes in AI Agent Pricing

Mistake 1: Pricing a System Like a Tool

Salesforce's $2/conversation launch priced a multi-agent system with a single-tool metric. The result: backlash and a pivot to Flex Credits within months. If your product orchestrates multiple agents, do not flatten the pricing into a per-conversation or per-task model. The buyer will feel like they are paying for architecture, not value.

Mistake 2: Ignoring the 15-Agent Ceiling

Per-agent seats work at 1 to 3 agents. At 15, the customer is paying more than a human team would cost. If your product is designed for multi-agent deployments, build volume discounts into the seat model or switch to platform-plus-metered pricing at scale.

Mistake 3: Anchoring on Compute Cost Instead of Value

If your AI agent saves a company $200,000/year in headcount, pricing at $20/month because your API costs are low is not "competitive." It is leaving 99% of the value on the table. Anchor your price on value delivered (FTE savings, revenue generated, costs avoided), not on your infrastructure costs.

A solid competitive pricing analysis can reveal how competitors in your space anchor their prices. If everyone else charges $5,000/month and you charge $50, buyers will question your quality, not celebrate your affordability.

Mistake 4: Hiding Pricing Behind "Contact Sales"

Many AI agent companies show no pricing at all. This works for pure enterprise plays like Sierra.ai (where all contracts are custom and sales-led). But if you want self-serve adoption, developer traction, or competitive visibility, transparent pricing is a requirement. Your pricing page is a positioning statement. Hiding it signals either that you don't know your own value or that you plan to charge whatever you can get away with.

For guidance on how to evaluate and benchmark your approach, see our guide on writing competitive intelligence reports.

What Comes Next: Outcome-Based Agent Pricing

The logical endpoint of agent pricing is not usage, not seats, but outcomes.

Sierra.ai is the most prominent example. Their pricing is tied to business results: a resolved support conversation, a saved cancellation, an upsell, or a cross-sell. If the conversation is unresolved, there is typically no charge. This creates perfect alignment between what the buyer pays and what they receive.

But outcome-based pricing introduces measurement complexity. Who defines "resolved"? What if the customer disputes the outcome? What if a resolution requires multiple conversations across days? Sierra's contracts are custom and enterprise-only, partly because building the measurement infrastructure is as complex as building the AI itself.

The prediction: by 2027, outcome-based pricing will be the dominant model for teammate-class agents in customer-facing roles (support, sales, success). Tool-class agents will remain per-usage. System-class agents will use platform-plus-outcome hybrids. The companies building outcome measurement infrastructure today are positioning themselves for this shift.

AI Agent Pricing FAQ

The Framework in One Sentence

Classify your agent as a tool, a teammate, or a system. Then match the pricing:

✦ Tool: Per-usage, per-resolution, or per-credit. Charge for work done.

✦ Teammate: Per-agent seat at 10-30% of the equivalent human salary. Charge for value replaced.

✦ System: Platform fee plus metered or outcome-based billing. Charge for throughput and results.

The real debate is not per-seat vs per-token. It is: when does an AI agent cross the line from tool to teammate, and how do you price that moment fairly for both sides?

The companies that answer this question well will write the next pricing playbook.