Two years ago almost every business AI tool was sold the same way: a monthly fee per user. In 2026 an AI agent can be billed per seat, per token, per action, per conversation or per resolved ticket, and the unit prices on the official price lists range from $0.10 to $2.00. This guide compares how Salesforce, HubSpot, Intercom, Zendesk, Anthropic, OpenAI, Mistral and Manus charge for agents, what buyers actually prefer, and how to work out the number that matters: the cost per completed task.
Why the per-seat model is breaking for AI agents
Seat pricing works when value scales with the number of people who log in. An AI agent breaks that link: a support agent that answers 10,000 customer questions a month may not need a single human seat, and if it works well it reduces the number of seats the customer buys. Vendors therefore look for a unit that grows with the work the agent does. The candidates are the input (tokens, credits, runtime), the activity (actions, conversations) or the result (a resolved ticket, a qualified lead).
Each unit moves risk between vendor and buyer. With token or credit pricing the buyer pays for every attempt, including loops and failures. With per-conversation pricing the buyer pays for every customer who talks to the agent, solved or not. With per-resolution pricing the vendor only earns when the agent succeeds, and the fight moves to the definition of "resolved".
What one unit of agent work costs: the price lists
The chart compares the official list prices of the large customer-service and CRM platforms in October 2026. They are not directly comparable, because an action is a much smaller unit than a resolved conversation, but they show how far apart the models are.
Sources: official pricing pages of Salesforce (Flex Credits: $500 per 100,000 credits, 20 credits per action and 30 per voice action), HubSpot (announcement of 13 April 2026), Intercom and Zendesk. Seat licences and platform fees are not included. Enterprise contracts are usually discounted.
- Salesforce Agentforce runs three models side by side: Flex Credits ($500 per 100,000 credits, so about $0.10 per standard action and $0.15 per voice action), $2 per conversation for customer-facing agents, and per-user add-ons at $125 per user per month for unmetered employee use. A $5 Agentforce user licence with metered credits sits in between.
- HubSpot switched on 14 April 2026 from $1.00 per conversation to $0.50 per resolved conversation for its Customer Agent, and from a monthly per-contact fee to $1.00 per recommended lead for its Prospecting Agent. HubSpot reports that the Customer Agent resolves 65% of conversations across more than 8,000 activated customers.
- Intercom Fin costs $0.99 per outcome, on top of a helpdesk seat ($29, $85 or $132 per seat per month).
- Zendesk charges $1.50 per automated resolution on a committed plan and $2.00 pay as you go. Each agent seat includes 5 automated resolutions a month (10 on Suite Professional).
Model providers: tokens, runtime and credits
The model labs price lower in the stack. You pay for the compute the agent consumes and you build the definition of success yourself.
- Anthropic. Claude Managed Agents is billed on two dimensions: tokens at the standard model rates and $0.08 per session-hour of runtime, measured to the millisecond and only while the session is running. Waiting for the user or for a tool confirmation is not billed. Web searches cost $10 per 1,000. Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens; Opus 5.5 costs $4 and $20.
- OpenAI. Codex moved from per-message to token-based pricing on 2 April 2026, aligned with API rates. OpenAI's help centre estimates about $100–200 per developer per month, with wide variation. In ChatGPT Business, Codex, Workspace Agents and ChatGPT for Excel draw on a shared pool of credits once the included limits are used up.
- Mistral. The API is billed per token, and its tool APIs (such as web search or code execution in the Agents API) are priced per call. The Vibe assistant (formerly Le Chat) costs $14.99 a month for Pro and $24.99 per user for Team; Enterprise is quoted individually.
- Manus. The autonomous agent is sold on credits. A task consumes credits for as long as it runs, depending on its complexity and duration, so the same plan can cover very different numbers of tasks. Third-party trackers list paid plans from about $20 a month.
Assumptions: 400,000 input tokens, of which 300,000 are cache reads; 40,000 output tokens; 20 web searches; 30 minutes in running state. Total about $0.87, of which runtime is $0.04. Prices: Anthropic pricing documentation, October 2026.
What buyers want: the seat comes back for AI
Vendors are moving to usage and outcomes, but buyers are more cautious. Futurum's 2H 2026 Enterprise Software Decision-Makers Survey asked which commercial model buyers prefer. For core software, per-seat pricing came last at 12.6%. For AI functionality billed separately it came first, at 42.3%, ahead of consumption-based (36.6%) and outcome-based pricing (21.1%).
Source: Futurum Research, 2H 2026 Enterprise Software Decision-Makers Survey, published 17 August 2026. Buyers ranked five commercial models; the chart shows the three reported in the text. In Futurum's 1H 2026 survey, 43% preferred consumption-based and 27% outcome-based models, so preferences are moving.
The reason is budget control. A seat is a known number in next year's budget; a token or resolution bill is not. When AI costs are unpredictable, finance teams ask for the model they can plan. The lesson for buyers is not that one model is right, but that predictability has a price, and it is worth negotiating caps, commitments and alerts whatever the unit.
What vendors do: hybrid is the new default
On the vendor side the shift is clear. Growth Unhinged's 2025 State of B2B Monetization survey of more than 240 software and AI companies found that pure seat-based pricing fell from 21% to 15% within twelve months and flat-fee subscriptions from 29% to 22%, while hybrid pricing (a base fee plus usage) rose from 27% to 41%. Pure outcome-based pricing is still rare: 5% use it as their primary model, but 25% expect to by 2028.
Source: Growth Unhinged, 2025 State of B2B Monetization (survey April–May 2025, more than 240 companies, typically $1–20M ARR). "A year earlier" is the respondents' model twelve months before the survey. Outcome-based: current share and share expected by 2028.
Salesforce is the clearest example: seats, credits and per-conversation pricing on the same price list. Intercom and Zendesk combine a seat with a per-resolution fee. Even the model labs, which sell tokens, now package agents into subscriptions with included usage and credit top-ups.
Per conversation vs per resolution: the effective price
A list price per conversation and a list price per resolution cannot be compared directly. What matters is the cost per resolved conversation, and that depends on how good the agent is. If an agent charged per conversation resolves only 40% of them, every resolution costs two and a half times the list price. A per-resolution price stays flat.
Calculation from list prices: $2 per conversation divided by the resolution rate, compared with flat per-resolution prices of $1.50 (Zendesk, committed), $0.99 (Intercom Fin) and $0.50 (HubSpot). Seats, platform fees and the cost of the human handling unresolved conversations are not included.
The gap shrinks as the agent improves: at a 30% resolution rate, $2 per conversation means $6.67 per resolution; at 90% it means $2.22. In customer service automation the human cost of the unresolved remainder is usually bigger than the agent fee, so the resolution rate is the first number to test.
The five pricing models compared
| Model | Examples | Good for the buyer when | Watch out for |
|---|---|---|---|
| Per seat | Agentforce add-ons ($125/user), ChatGPT and Claude business plans | Usage per employee is heavy and predictable | Paying for unused seats; fair-use limits |
| Tokens / credits / runtime | Claude API and Managed Agents, OpenAI Codex, Manus credits | You build your own agents and can optimise prompts | Loops and failed runs are billed; credits may expire |
| Per action | Salesforce Flex Credits ($0.10 per action) | Small, well-defined steps inside workflows | Agents that take many steps per task |
| Per conversation | Agentforce customer agents ($2) | The agent resolves a high share of conversations | Unresolved conversations are billed in full |
| Per outcome / resolution | HubSpot ($0.50), Intercom Fin ($0.99), Zendesk ($1.50–2.00) | Outcomes are easy to define and audit | Vendor-defined "resolution"; reopened tickets billed twice |
How to compare offers: a checklist
- Calculate the cost per completed task. Unit price times the units a real task consumes, plus seats and platform fees. Use the same 100 to 200 real tickets, documents or leads for every vendor.
- Read the definition of the outcome. When is a conversation "resolved", for how long, and what happens if the customer comes back? Ask for the audit log of billed outcomes.
- Ask what is not billed. Escalations to a human, failed runs, idle time and retries are treated differently by each vendor.
- Put a ceiling on usage. Monthly caps, budget alerts and a fallback to a human queue prevent a bad prompt or a traffic spike from becoming an invoice surprise.
- Check expiry and commitments. Pre-purchased credits and committed resolutions are cheaper per unit but may expire at the end of the term.
- Compare build vs buy. A platform agent at $1 per resolution and your own agent on the API at a few cents in tokens are different products: the second needs development, evaluation and operation, as we explain in our guide to AI inference economics.
What this means for companies in Croatia and the DACH region
Most list prices are in US dollars, and euro prices often mirror them one to one (Salesforce lists Flex Credits at €500 per 100,000). For a mid-sized company the decision is less about the cheapest unit and more about where the risk sits. If your volumes are small or seasonal, per-resolution pricing keeps the downside low. If you have a stable, high volume and an in-house team, tokens on the API plus your own evaluation are usually far cheaper per task. In both cases the business case should start from the ROI of the process, not from the unit price, and our comparison of AI subscription plans covers the per-seat side.
Frequently asked questions
What does an AI agent cost per resolved customer conversation in 2026?
List prices range from $0.50 (HubSpot Customer Agent) and $0.99 (Intercom Fin) to $1.50–2.00 (Zendesk automated resolutions). Salesforce charges $2 per conversation for customer-facing Agentforce agents, whether or not the conversation is resolved.
Is outcome-based pricing better for buyers?
Not automatically. It shifts the risk of a weak agent to the vendor, but the definition of a resolution decides what you pay. In Futurum's 2H 2026 survey only 21.1% of buyers preferred outcome-based pricing for AI; 42.3% preferred per-seat.
How is Claude Managed Agents priced?
On two dimensions: tokens at the standard model rates, plus session runtime of $0.08 per session-hour while the session is running. Idle time is not billed. Web searches cost $10 per 1,000.
How should a company budget for AI agents?
Model the cost per completed task, not the unit price: multiply the unit price by the units a task really consumes, add seats and platform fees, and test with your own tickets or documents before signing a commitment.
Which pricing model fits your AI agent?
We help companies in Croatia and the DACH region estimate the real cost per task of an AI agent, compare platform offers with their own Claude-based solution and set up the budget controls before roll-out. As a Claude Certified Architect from Zagreb, we start with your process data, not with a price list.
Talk to an AI consultant