Which AI agent, and what it really costs
IC Guides · AI Agents
An AI agent’s price page tells you the unit, not the bill. One vendor charges per resolved conversation, another per credit, another per minute. The same thousand support conversations can cost $500 with one and $2,000 with another before anyone adds seats or platform fees. This guide helps you choose an agent by job, then work out what it would cost to run.
Pick by job
Start with the work, not the vendor. The job decides the billing unit you are likely to meet.
| Your job | Typical unit | Start with |
|---|---|---|
| Answer customers, with published prices | Per resolution or outcome | Fin, Zendesk, HubSpot compared |
| Answer customers, price on request | Contract specific | Decagon, Ada, Sierra, Forethought |
| Run a CRM or service platform agent | Credits | Salesforce Agentforce, Zendesk |
| Take or make phone calls | Per minute | Vapi vs Retell |
| Automate repeatable workflows | Executions or credits | n8n, Make, Gumloop, UiPath |
| Delegate inbox and calendar work | Seats and pooled credits | Lindy, Manus |
| Write and ship code | Plan limits and usage | Claude Code vs Devin |
| Build an agent agent | Tokens and tool calls | OpenAI Agents API |
The billing units
A seat is a fixed price per user. It is predictable, but you pay for people who barely use it. Outcome or resolution pricing means you pay when the agent completes a defined result, and since the vendor defines a resolution, read that definition. Conversation or session pricing charges per exchange, whether or not it solved anything.
With credits or tokens you buy capacity, and different actions use different amounts, so a pack does not equal a fixed number of tasks. Minutes are common in voice, and the listed rate often excludes the model, voice and telephony. Hybrid deals combine a platform fee with usage, and most enterprise deals end up there.
What the same work costs
These are estimates from public rates, before seats, base subscriptions, model costs and contract terms.
Take 1,000 support resolutions first.
| Provider | Public meter | About |
|---|---|---|
| HubSpot | $0.50 per resolved conversation | $500 |
| Fin (Intercom) | $0.99 per outcome | $990 |
| Yellow.ai | $0.99 per resolution after 500 free sessions | $990 |
| Salesforce Help Agent | $2 per qualifying resolution | $2,000 |
The spread is fourfold, but the numbers are not like for like. A resolution means something different at each company, and seats, add-ons and minimums sit on top. Zendesk is left out because its automated resolution pricing depends on a commitment and a resolution rate you must confirm.
Now take 2,000 voice minutes.
| Provider | Public meter | About |
|---|---|---|
| Vapi | $0.05 per minute hosting | $100 hosting alone. Vapi’s calculator puts a full stack at $82 to $129 per 1,000 minutes, so about $164 to $258 for 2,000, before the phone carrier. |
| Retell AI | $0.07 to $0.31 per minute | $140 to $620 by configuration. Retell’s example stack at $0.11 a minute is $220. |
Once you add model, voice and transcription, the two platforms land in the same band. The choice of model moves the bill more than the choice of platform. See voice agent pricing for the component tables.
Compare against the live rows in the AI Agent Pricing Tracker, where each price links to its source and shows when we checked it.
Where bills surprise you
Credits are not tasks. Salesforce sells 100,000 Flex Credits for $500 and a standard action uses 20, so a pack covers about 5,000 standard actions. Microsoft sells 25,000 Copilot Credits for $200, with different actions using different amounts. Free allowances can also use a different unit, since Yellow.ai’s 500 free sessions are not deducted from billable resolutions.
Floors and minimums also change the bill, as Vapi’s Pro plan adds 10% of hosting fees with a $999 a month minimum. ServiceNow publishes the Assist unit but no standard dollar price, so treat a missing price as an unknown and not a bargain. Different meters can share the same word, and ten thousand credits do not buy ten thousand workflow runs.
Decision rules by situation
These rules turn the tables above into a first screen. They use public rates checked on 12 October 2026, and each one says where it stops working.
Your support volume is steady and countable
Per resolution pricing fits, because cost follows completed work. At 1,000 resolutions a month the published rates run from about $500 at HubSpot ($0.50) to $2,000 at Salesforce Help Agent ($2). Before ranking them, ask each vendor whether a conversation handed to a human is billed, since the vendor defines what a resolution is. The rule weakens when seats and platform fees are large next to usage, because the per resolution rate is then the smaller part of the bill.
Your volume swings from month to month
Minimums and commitments cost most in the quiet months. Vapi’s published Pro plan adds 10% of hosting fees with a $999 a month minimum, which is the kind of floor to look for. Ask whether unused credits expire and what happens at the limit, whether that is throttling, overage rates or a hard stop.
The price is on request
Decagon, Ada, Sierra and Forethought do not publish rates, so they cannot be ranked against the figures above until you hold a written quote. Give each vendor the same workload and ask for the billing unit, the minimum and the exclusions in writing. Our guide to quote only support agents lists what to settle in the contract.
You are buying voice minutes
Price the full stack and not the hosting line. Vapi’s published calculator puts a full stack at $82 to $129 per 1,000 minutes, and Retell’s example stack at $0.11 a minute comes to $220 for 2,000 minutes. On public rates the two platforms land in the same band, so the model and voice you choose move the bill more than the platform does. Telephony sits outside both figures.
The platform sells credits
A credit is capacity and not a task. Salesforce sells 100,000 Flex Credits for $500 and a standard action uses 20, so a pack covers about 5,000 standard actions, or $0.10 each. A task that takes six standard actions would use 120 credits, about $0.60 at that rate, which is an illustration and not a quote. Microsoft sells 25,000 Copilot Credits for $200 with different actions using different amounts, so no per task figure holds until you count the actions in your workflow.
You need a budget ceiling
A seat plan fixes the bill, and you pay for light users as well as heavy ones. Capped credit plans set a ceiling too, but check what happens when the cap is reached before you rely on it.
None of these rules replaces testing. Run each shortlisted product against the same sample of real work before you sign, and compare the cost of what was completed.
Questions to ask before you buy
- What exactly counts as a billable outcome, and what is explicitly not billed?
- Does a conversation handed to a human still count?
- Which costs sit outside the listed rate, such as seats, model usage, telephony and platform fees?
- Is there a minimum commitment, and do unused credits expire?
- What happens at the limit, whether throttling, overage rates or a hard stop?
- Can I audit what the agent did for each billed item?
Sources and method
Most figures come from the public rates recorded in the IC AI Agent Pricing Tracker, which links each price to the vendor’s page. We re-read the vendor pages for Intercom Fin, HubSpot, Zendesk, Salesforce Agentforce, Microsoft Copilot Studio, Vapi, Retell, n8n, Gumloop, Lindy, UiPath and OpenAI on 12 October 2026, and each guide lists its dated sources. Prices that no vendor publishes appear as third party estimates in the guide that covers them, labelled as unconfirmed. We did not test these products, and vendor claims stay attributed to the vendor. Prices change, so check the tracker row’s date before relying on a number. Spotted something out of date? Send a correction.
Read next
IC Lab · AI Agent Pricing Tracker
Vendor prices change. The tracker records when.
Every entry links to the vendor’s pricing page and shows the date we checked it. Open the tracker, or get major pricing changes by email every Friday in TalkHead.
Related guides What is an AI agent? · AI agent glossary · Build or buy · Free AI agents · Permissions and safety · AI Agent Pricing Tracker · All AI agent guides
Frequently asked questions
Which AI agent should I choose first?
Choose by the work you need completed, customer support, coding, voice calls, or repeatable workflows. Compare the products in that category before comparing their prices. A low price per credit is not necessarily a low price per completed task.
How do I compare agents with different billing units?
Define one monthly workload and estimate the billable events each product would generate. Add seats, platform fees, model usage and likely retries. Divide total monthly cost by work completed successfully, not by credits purchased.
Is the cheapest published rate the cheapest agent?
Not necessarily. Rates may exclude subscriptions, minimum commitments, telephony and other usage. The definition of a billable outcome also varies. Treat the figures in the tables as scenarios, not like-for-like quotes.
Where can I check the latest price?
Use the AI Agent Pricing Tracker for linked vendor sources and last-checked dates. Confirm contract-specific terms with the provider before purchase.
Read beyond this guide
Why AI workers challenge the per-seat SaaS model Read the business analysis behind this implementation decision. See the wider AI agent pricing comparisons Compare billing units and vendor documentation before choosing.