How AI agents are priced, and what each pricing unit means for your bill

AI agent pricing models come down to one question: what event makes the meter tick. Vendors sell agents per seat, per task or action, per resolution, per minute, or as a share of an outcome such as money recovered. The pricing unit decides who carries the risk when the agent does a lot of work, little work or the wrong work. Before you compare two prices, check that they count the same thing. "$0.99 per outcome" and "$2 per conversation" are not two prices for one product; they are prices for two different events.

This guide classifies the units you will meet, with one published example of each, quoted from the vendor's own pricing page as of November 2026. The examples show how the units work. They are not recommendations, and they are not a market average. Prices change; check the live page before you budget.

The pricing units at a glance

Unit	You pay when	Example cited here	Who carries the volume risk
Per seat	A named person has a license, whatever the agent does	Salesforce Agentforce add-ons, $125 per user per month	You pay for idle seats; the vendor absorbs heavy use, within any quota
Per task or action	The agent performs a metered step	Salesforce Flex Credits, 20 credits per Agentforce action	You pay for every step, including steps in work that fails
Per conversation	A conversation happens	Salesforce Agentforce Conversations, $2 per conversation	You pay whether or not the conversation solves anything
Per resolution	The vendor's definition of "resolved" is met	Intercom Fin, $0.99 per outcome	Shared; depends on how "resolved" is defined
Per minute	Time runs on a call	Retell AI voice agents, $0.07–$0.31 per minute	You pay for long calls, whether or not they end well
Share of outcome	A defined business result happens	Chargeflow, 25% per recovered chargeback	Mostly the vendor; you pay only when the result arrives

Per conversation sits between task and resolution pricing, so it gets its own row. Hybrid plans combine two or more rows; they are covered below.

Per seat

A seat price charges for each person who can use the agent, usually per month. It is the familiar software license, and it is the easiest unit to budget.

Salesforce lists its Agentforce add-ons for Sales, Service, and Field Service at $125 per user per month, which it describes as "unmetered Agentforce usage for employees"; Industries add-ons are $150 per user per month and Max Edition starts at $550 per user per month. Salesforce notes on the same page that the add-ons are "not available" with its PayGo and Pre-Commit buying models. Cognition's Devin Teams plan is $80 per month for the team plan plus $40 per month per full developer seat, up to 100 users; each paid plan comes with a usage allowance that refreshes daily and weekly, and usage beyond it is bought "at API pricing."

What a seat price means for you:

• It counts people, not work. Seat pricing fits agents that assist your staff (an assistant inside the CRM, a coding agent a developer drives). It fits poorly when the agent does a job on its own, because the agent itself has no seat.
• Read the usage terms. "Unlimited" or "unmetered" seats can sit beside quotas, as in Devin's plans. Ask what happens when a heavy user hits the limit: throttling, a slower model, or overage at a stated rate.
• Idle seats cost the same as busy ones. Count who will actually use the agent each week, not who might.

Per task or action

Task pricing charges for each step the agent performs. The vendor defines the step, so the first job is to find that definition.

Salesforce sells Flex Credits at $500 per 100,000 credits. Its pricing page defines an action as "a specific function that an AI agent executes on the platform, such as updating a record, summarizing a complex case, answering a product inquiry, or executing a custom prompt or flow," and states that "Agentforce actions are 20 Flex Credits, while Agentforce Voice actions are 30 Flex Credits." At list price that is $500 ÷ 100,000 × 20 = $0.10 per Agentforce action, and $0.15 per Voice action. Salesforce's own "Where is My Order?" example uses two actions per request, so $0.20 per request.

Three terms on the same page matter for a buyer. Unused Flex Credits "do not roll over into subsequent subscription terms." Exceeding the balance carries "no overage penalty"; extra use is billed "monthly in arrears" at the contracted rate. And the worked examples "do not include other costs like Data 360 credits or other consumption services."

What task pricing means for you:

• One request can be several billable steps. Ask the vendor how many actions your typical unit of work takes, then confirm it in a trial.
• Failed work still consumes actions. A request that ends with a handoff to a person has already used its steps.
• Prepaid pools can expire. Size a credit purchase to measured usage, not to the vendor's estimate.

Per conversation

Salesforce also sells Agentforce Conversations at $2 per conversation, for customer-facing agents. It describes this as "flat-pricing," against Flex Credits that "align cost to value." The two models cannot be mixed: "Flex Credits and Conversations will not be supported in the same org." Switching from Conversations to Flex Credits means swapping "all of your existing Agentforce Conversation SKUs."

A conversation price is predictable per contact, and it charges the same for a conversation the agent solves as for one it hands to your team.

Per resolution

Resolution pricing charges only when the agent's work counts as resolved. It moves part of the risk to the vendor, so the definition of "resolved" is the whole contract.

Intercom prices its Fin AI Agent at $0.99 per outcome. Its pricing page says an outcome is counted when "a customer confirms their issue is resolved," when "they don't ask for more help after Fin responds," or when "Fin completes a workflow (Procedure), including handoffs," and that you are "only charged once per conversation, even if multiple questions are answered." For Fin used with another help desk, the page lists $0.99 per outcome, "no seat costs," "no setup, integration, or platform fees," and a "minimum monthly commitment" with "50 outcomes" given as an example. On Intercom's own plans, Fin outcomes are billed on top of seats, with "1 Full seat minimum": Essential lists $29 per seat per month (shown as $19 under a new-customer offer), Advanced $85 and Expert $132.

Zendesk uses a different definition for the same idea. Its help article says AI agent usage is measured through automated resolutions, meaning "you pay only for customer requests that were successfully resolved by an AI agent, without any escalation to a human agent," with an allowance included in each plan and more available to buy. The article does not give a price.

Compare those two definitions. Intercom's counts a completed workflow that includes a handoff, and a customer who simply stops replying. Zendesk's, as written in that article, excludes requests escalated to a human. Neither is wrong, but the same month of tickets can produce different billable counts under each. Ask every vendor for the written definition, and check it against a sample of your own conversations during the trial.

Per minute

Retell AI prices its voice agents by call time. Its pay-as-you-go plan lists "$0.07-$0.31 / min for AI Voice Agents," with $10 in free credits and "no commitments"; its Enterprise plan is "custom pricing." The per-minute total is built from components priced separately on the same page: Retell's voice infrastructure at $0.055 per minute, a text-to-speech voice (from $0.015 per minute for Retell's platform voices), the language model you pick (each model has its own per-minute rate) and telephony. The page's default calculator shows $0.11 per minute: $0.04 LLM, $0.055 voice infrastructure, $0.015 text-to-speech and $0.00 telephony.

What per-minute pricing means for you:

• Your configuration sets the rate. The model and voice you choose move the price within the range.
• Long calls cost more, whether or not they solve anything. A call that loops for eight minutes before a transfer is billed for eight minutes.
• Check what is outside the range. Add-ons and telephony are listed as separate components on Retell's page.

The phone support job page covers what an inbound phone agent does and where a person stays involved.

Share of outcome

Outcome pricing can tie the fee to a business result measured in money. Chargeflow charges 25% per recovered chargeback for its Automation product, with "no payment until you recover a chargeback." On the same page it prices Alerts at $29 per deflected chargeback (merchants with "over 50 alerts per month" are directed to request a better price) and Prevent at $0.10 per scanned transaction, with the first 1,000 scanned transactions free. One vendor uses three units for three products.

What outcome pricing means for you:

• The vendor only earns when the result happens, which is the closest thing to paying for work done.
• The definition and the measurement belong to the contract. Who decides that a chargeback was recovered, and from which system's data?
• A percentage scales with value, not effort. One large recovery can cost more than many small ones. Model it on your own case sizes.

The accounts receivable job page covers collections work, where recovered amounts are the natural measure.

Hybrid plans

Several of the examples above are hybrids. Intercom charges seats plus outcomes. Salesforce's Agentforce User License is $5 per user per month and "requires Flex Credits," and Max Edition bundles seats with "2.75M Flex Credits per org per year." Devin combines a team fee, seats and a usage allowance with extra usage at API pricing. When you see a hybrid, list every line separately: fixed fees, seats, included usage, the overage rate and any minimum commitment.

A worked example: one month of support chats, three units

Assume, for this example, that a support team receives 1,000 chat conversations a month. On Intercom it uses one full seat at the $29 Essential list price. On Salesforce it buys no seats, and every conversation is a single request using two Agentforce actions, as in Salesforce's own "Where is My Order?" example. Assume also that 500 of the 1,000 conversations count as Fin outcomes. These numbers are illustrations, not measured rates.

Unit	Arithmetic	Month, list price
Intercom Fin, per outcome, Essential plan	500 × $0.99 + 1 seat × $29	$524
Salesforce Agentforce, per conversation	1,000 × $2	$2,000
Salesforce Agentforce, Flex Credits	1,000 × 2 actions × 20 credits = 40,000 credits; 40,000 ÷ 100,000 × $500	$200

The table shows how much the bill depends on the unit and on your assumptions, not which product is cheaper. Change the assumed outcome count to 900 and the Intercom line becomes 900 × $0.99 + $29 = $920. Add a third action per request and the Flex Credits line becomes $300. None of the three lines includes your team's time on conversations the agent does not finish, any platform licenses the vendor requires, or negotiated discounts. Run your own numbers from a trial; the trial guide explains how to record cost per unit while you test.

How to compare quotes

Put every quote into one sheet with the same columns:

1. The billable event, in the vendor's words. Copy the definition from the pricing page or contract, not the sales deck.
2. The rate and the currency.
3. Fixed costs. Seats, platform fees, setup, add-ons.
4. Commitments. Monthly minimums, prepaid pools, expiry, annual terms.
5. Overage terms. Rate, and whether service stops or continues.
6. What is excluded. Telephony, data credits, other consumption services.
7. Your measured count of billable events from a trial on your own work.

Then compare the monthly total with what the same job costs you today. The AI agent vs employee cost guide sets out that comparison with sourced wages, and the hiring guide puts pricing in the full hiring sequence. Put the pricing unit and the billable-event definition into your brief as well; the job description guide shows where.

Questions buyers ask

Which pricing model is best? None of the sources cited here supports a general answer. Our recommendation: match the unit to how you can measure the job. If you can count resolved tickets reliably, resolution pricing lines up cost with results. If the agent assists named staff, seats are simpler. If volume swings, a usage unit avoids paying for idle seats.

Is per-resolution pricing always cheaper than per-conversation? No. It depends on the rate, the share of conversations that count as resolved under the vendor's definition, and any seats or minimums alongside it. The worked example above shows the arithmetic.

Can I negotiate? Your final quote may differ: Salesforce says its listed prices are "subject to change" and points to a sales representative for detailed pricing; Retell Enterprise is custom-priced; and Chargeflow invites merchants with over 50 alerts a month to request a better price.

What should I ask for in writing? The billable-event definition, the rate, every minimum and expiry, overage terms, and how you can see usage while it happens. Salesforce, for example, points to its Digital Wallet for usage tracking and threshold alerts.

Sources

1. 1
   Intercom Pricing

   IntercomVendorAccessed Nov 14, 2026

2. 2
   Salesforce Agentforce Pricing

   SalesforceVendorAccessed Nov 14, 2026

3. 3
   AI Phone Agent Pricing

   Retell AIVendorAccessed Nov 14, 2026

4. 4
   Chargeflow Pricing

   ChargeflowVendorAccessed Nov 14, 2026

5. 5
   Devin Plans and Pricing

   CognitionVendorAccessed Nov 14, 2026

6. 6
   About AI agents

   ZendeskVendorAccessed Nov 14, 2026