AI Agent Pricing Models, Compared

Finance has the quote, and the unit rate is the only number anyone can defend in the meeting. One vendor charges per seat. Another charges per task. A third wants a cut of meetings booked. The same workload can cost several times as much under one of those structures as under another, at the same volume, because the contract is defining a different event as "one." Gartner predicts that more than 40% of agentic AI projects started in 2025 will be canceled by 2027, with escalating costs and unclear value as the two leading causes (June 2025).

What are the main AI agent pricing models?

Vendors sell agents under four structures. Knowing which structure you are buying matters more than comparing the unit rates, because the same workload can cost three times as much under one model as under another at identical volume.

Before any demo, mark which of the four applies to each line on the quote. Most enterprise contracts mix two or three, which is why a single "per month" number on the pricing page is not the number finance should model.

Per-seat pricing.

A seat is a license for one active agent. You pay a fixed monthly or annual fee per agent, regardless of how many times it runs or how many tokens it consumes.

Usage-based pricing.

Usage pricing charges per unit of work: per API call, per token processed, per task completed, or per action that changes something, such as sending an email or updating a contact. Cost scales with activity.

Outcome-based pricing and hybrid models.

Outcome pricing charges on results: per meeting booked, per lead qualified, per deal the vendor says the agent touched. Very few vendors offer a pure outcome contract, because attribution is difficult and reversals happen. For a wider look at what a pre-built agent covers compared with building on your own team, see the build versus buy comparison.

How does per-seat pricing work for AI agents?

Per-seat pricing is predictable, and it is coarse. Finance can budget it the way they budget other software: a fixed line.

When a seat is the right shape.

A seat works best when the agent runs a defined task at roughly steady volume. Think of an inbound qualification agent processing 50 to 100 form submissions a month, a content-sorting agent handling a fixed intake, or a scheduling agent handling a known number of calendar events.

Where a seat stops matching the work.

A seat breaks down when the workload swings. A campaign that sends ten times the usual traffic to a form means your per-seat agent processes ten times the volume at the same fee. Budgeting the quiet months at peak-volume pricing is waste that accumulates across a 12-month contract.

The concurrency cap in the service description.

Some per-seat contracts include a concurrency limit or a rate cap in the service description, not on the pricing page. A seat set to run 100 concurrent tasks will process your baseline and then hold the overflow during a peak.

When does usage-based pricing make sense for AI agents?

Usage-based pricing is the plainest model for a production agent: your cost tracks what you consume. It is also the hardest to budget, because the same agent in a heavy month can cost five to eight times its baseline.

What counts as a billable event.

The definition of a billing trigger decides your cost, and vendors do not share one definition. Common triggers include each API call to the language model (charged per token or per request), each tool call the agent makes (reading a contact, writing an activity, querying enrichment data), each task completion, or each action with a side effect such as sending an email or booking a meeting.

How to forecast before you sign.

The useful forecast is a metered pilot. Run the agent for 30 days with usage logging on, count the billing events in each category, multiply by the contract rate, and multiply by 2x for volume growth over the contract term. If the vendor cannot turn usage logging on during that pilot, you have a contract problem before you have a cost problem, because you will be signing a meter you cannot read.

The spike that arrives with a paid campaign.

Usage-based pricing creates a structural risk when the agent's workload is tied to paid media. A campaign that generates ten times the form volume generates ten times the agent cost in the same week you are paying ten times the media budget. The operations budget needs either a burst allowance or a usage cap that falls back to a person when the cap is hit. A hard stop in the middle of a campaign, or an automatic overage with no ceiling, are the two outcomes to rule out before the media goes live.

What is outcome-based pricing and can you actually get it?

Outcome-based pricing sounds like the alignment everyone wants: you pay when the agent delivers a result, and the vendor carries the cost of poor performance. In practice it is rare outside of a performance add-on on a larger enterprise deal, and the attribution mechanics are harder than the pitch.

The two shapes vendors actually offer.

Vendors who offer outcome pricing generally use one of two shapes. The first is a performance floor: a base platform fee plus a per-outcome charge, so the vendor has floor revenue and you pay extra only when the agent performs. The second is a pure success fee, with no base and a percentage of each result. Pure success fees are rare, because the vendor cannot control your conversion rate, your offer, or how fast your team follows up. A sales team that lets agent-qualified leads sit for 48 hours before anyone calls will underperform on any outcome model, whatever the agent did, and the agent will take the blame for a follow-up the team never made.

For how chatbot contracts and agent contracts differ on this point, the comparison of agents and chatbots covers the deployment patterns that push a vendor toward one pricing structure or another.

Clauses worth reading twice.

Watch for attribution windows the vendor defines alone, especially anything past 30 days. Watch for outcome definitions that include events the agent did not cause, for clawback language you cannot apply to a specific record in your CRM, and for a success metric you cannot pull yourself. If the vendor both runs the agent and decides what the agent caused, the bill and the measurement sit in the same hands.

How do you calculate total cost of ownership for an AI agent?

The unit rate is not the total. Every deployment carries costs that never appear on the pricing page: integration work, monitoring, prompt maintenance, and a person reviewing the cases the agent escalates.

What shows up on the quote, and what shows up later.

Direct costs are the vendor license, whether seat or usage, the language-model API cost on a bring-your-own-key model, and enrichment or data-provider fees. The costs that show up later are the engineering time to connect the agent to your CRM, calendar, and webhooks; ongoing prompt edits as your offer and your buyer change; the time spent when the agent escalates an edge case; and the compute for a monitoring layer. Those later costs compound over a 12-month contract, and they rarely appear in a vendor calculator.

An illustrative year, so you can see the shape.

The following is an illustrative example, not a client result. A three-agent outbound stack on usage-based pricing at $0.02 per task, running 5,000 tasks per month, costs $100 per month at the listed rate. Add $2,000 in one-time integration engineering, $300 per month in enrichment fees, and four hours per month of prompt maintenance at an internal cost of $150 per hour. Year-one total: ($100 x 12) + $2,000 + ($300 x 12) + ($600 x 12) = $12,200. A comparable per-seat contract at $800 per month for three seats puts year one at $9,600. The seat model is cheaper in this illustration only if usage stays below the seat's volume ceiling.

Three costs the pricing page leaves off.

Escalation handling is the time a person spends on edge cases, typically two to five hours per month for a production agent. Prompt upkeep is the revision prompts need as your product, your buyer, and your competitors change, usually one to two hours per quarter per agent.

Build vs. buy: what does the cost structure actually look like?

When you buy a pre-built agent, you pay the vendor's pricing model and you accept their architecture. When you build with your own team, you pay engineering up front and infrastructure on an ongoing basis.

Buying a pre-built agent.

The main advantage is time. A vendor-supplied agent is configured rather than built from a blank page, which puts a deployment in weeks rather than months. A task that costs $0.10 per call from a vendor may cost $0.008 per call to run yourself at the same volume. The guide to what agents cost to build shows engineering-hour estimates and infrastructure costs for common agent types.

Building with your own team.

Building in house requires a language-model API key, integration engineering, and prompt development. The ongoing cost is the token bill. AgentBench (Liu et al., arXiv:2308.03688, ICLR 2024) evaluated commercial and open-source language models across eight interactive agent environments and found that top commercial models significantly outperform open-source models under 70 billion parameters on agent tasks.

The review time a weaker model creates.

An agent that fails 20% of tasks instead of 5% needs roughly four times the human review time. That review cost often exceeds the API savings from a cheaper model. Before you pick an open-source model to shrink the token bill, test it on the booking, enrichment, or outreach task you actually want it to run. A score on an academic benchmark does not tell you how it will behave on that task, and the review hours will show up on your week either way.

What should you ask before signing an AI agent contract?

6sense's 2025 Buyer Experience Report (more than 4,000 B2B buyers) found that 94% of buying groups rank a preferred vendor before they ever contact the sales team. A related 6sense study, Science of B2B 2025 (n=634 B2B marketers plus 100 synthetic VC-backed responses), found that only 13% of B2B organizations report closed-won revenue to their boards. If the number you are paying the vendor against is not a number the business already reports, you will not be able to tell whether the agent earned its place.

Questions about the structure.

Ask how the vendor defines a billable event, specifically, and read the service description rather than the pricing page. Ask the rate for overages, and whether that rate is capped.

Questions about what happens when volume grows.

Ask what the vendor charges at ten times your current volume. Ask whether the per-unit rate decreases as volume grows, and at which thresholds. For a view of what a first deployment includes, from the initial map through the build and the monitoring, the systems page shows, in one place, how that work is organized in practice.

Six questions for the vendor call.

1. How does the vendor define a billable task or event at the level of each smaller action inside it?
2. What concurrency limit applies to each agent seat, and is that limit in the contract or only a soft setting?
3. What is the rate for volume above the contracted tier?
4. What is included in the base platform fee, and what is metered separately?
5. Who controls usage reporting, and can you pull the raw data from your own systems?
6. What is the performance guarantee, and what credit or refund applies if the agent misses it?

Common questions.

What is the most common pricing model for AI agents in 2025?

Usage-based pricing is the most common structure for production AI agent deployments. Vendors charge per API call, per token processed, per task completed, or per action with a side effect, such as sending an email or updating a contact in the CRM.

Is usage-based pricing cheaper than per-seat for AI agents?

It depends on your volume and how much that volume swings. Usage-based pricing is cheaper at low and variable volumes, because a slow month costs less than a busy one. Per-seat pricing is cheaper when volume is high and steady, because you pay the same fee no matter how many tasks the agent runs. Run a 30-day metered pilot before you commit to a 12-month usage contract. Multiply the observed task rate by the contract rate, then by 2x for growth, to estimate year-one cost.

What does outcome-based pricing mean for an AI sales agent?

Outcome-based pricing charges per delivered result: per qualified meeting booked, per lead advanced to an opportunity stage, or per deal influenced inside an attribution window. Pure outcome contracts are rare because the vendor cannot control your conversion rate, your offer, or your follow-up speed. Watch for attribution windows longer than 30 days, for outcomes the agent did not directly cause, and for usage reporting controlled only by the vendor.

How do you calculate the total cost of ownership for an AI agent?

Add the direct costs (vendor license, language-model API fees on bring-your-own-key models, enrichment provider fees) to the costs that show up later (integration engineering, prompt maintenance, escalation handling, and monitoring). Direct costs appear on the pricing page. A production agent typically needs one to two hours of prompt maintenance per quarter and two to five hours per month of escalation triage. Those hours compound across a 12-month contract. Run a 30-day pilot with usage logging to establish the direct-cost baseline, then add the later-cost estimates to reach a total you can defend.

When does building an AI agent in-house cost less than buying?

Building with your own team costs less at high volume when the vendor's per-unit rate exceeds your language-model API cost plus the engineering you amortize. AgentBench (ICLR 2024) found that top commercial language models significantly outperform open-source models under 70 billion parameters on agent tasks. An agent that fails 20% of tasks instead of 5% requires roughly four times the human review time, which often wipes out the API savings from a cheaper open-source model.

Where these figures come from.

Two Gartner press releases are cited here. The August 2025 release is the source of the prediction that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from under 5% in 2025. The June 2025 release is the source of the prediction that more than 40% of agentic AI projects will be canceled by the end of 2027, with escalating costs and unclear value among the causes.

6sense's 2025 Buyer Experience Report (more than 4,000 B2B buyers) is the source for 94% of buying groups ranking a preferred vendor before they contact sellers. 6sense's Science of B2B 2025 (n=634 B2B marketers plus 100 synthetic VC-backed responses) is the source for only 13% of B2B organizations reporting closed-won revenue to boards. AgentBench (Liu et al., arXiv:2308.03688, ICLR 2024), which tested commercial and open-source models across eight interactive agent environments, is the source for commercial models outperforming open-source models under 70 billion parameters on agent tasks.

The year-one illustration uses round figures to show the shape of the cost, not a deployment result. The per-task rate of $0.02, the 5,000 tasks, the $100 month, the $2,000 integration, the $300 enrichment fee, the $150 hourly internal cost, the $600 monthly maintenance, the $12,200 year-one total, the $800 seat fee, the $9,600 seat year, and the $0.10 versus $0.008 per-call comparison are an illustrative example, not a client result. The build versus buy comparison lays out the questions to settle before a 12-month vendor contract.