AI Agent Pricing: Per Seat vs Resolution vs Minute

AI agent vendors bill in at least five different units: per seat, per conversation, per resolved conversation, per action or credit, and per minute. Run the same rough workload, roughly 800 customer conversations a month with a 70 percent resolution rate, through each vendor's own published rate card, and the monthly total ranges from about $280 to $1,600. The difference is not the AI. It is which unit the vendor chose to bill.

That spread is why two business owners can compare three vendor quotes and walk away with no idea which one is actually cheaper. A quote in "$0.99 per outcome" and a quote in "$2 per conversation" and a quote in "$150 per user per month" are not comparable numbers until someone converts them into the same unit at the same volume. Vendors do not do that conversion for you, because the vendor whose model looks worse at your volume has no reason to show the math.

Why the same AI agent workload produces such different bills

Five pricing models cover most of what is sold as an "AI agent" in 2026, and each one rewards a different kind of buyer.

Per seat charges a flat fee per human license, regardless of how much the AI agent actually handles. GitHub Copilot Business runs $19 per user per month and Enterprise runs $39, both now bundling a monthly pool of AI credits consumed by usage rather than a flat seat fee alone, a change GitHub made in June 2026. Per-seat pricing is predictable and rewards a stable team, but it does not go down when the AI gets better at its job, and it can get expensive fast if you add headcount instead of letting the agent absorb volume.

Per conversation charges for every interaction the agent touches, resolved or not. Salesforce Agentforce bills $2 per conversation in this mode, counting the session from first contact to close regardless of whether it ends in a resolution, a human handoff, or an abandoned chat. This model is simple to budget but penalizes an agent that starts a lot of conversations it does not finish.

Per resolution (or "outcome") only charges when the agent actually closes the loop. Intercom's Fin charges $0.99 per outcome, where an outcome is a confirmed resolution, a customer who does not ask for more help, or a completed workflow including a handoff. HubSpot's Breeze Customer Agent moved to this model on April 14, 2026, dropping from $1.00 per conversation to $0.50 per resolved conversation, billed as 50 credits per resolution at $10 per 1,000 credits. Below roughly a 50 percent resolution rate, per-resolution pricing tends to beat per-conversation; above it, per-conversation can come out cheaper, because the vendor is no longer getting paid only for the wins.

Per action or credit breaks a conversation into its component steps and charges for each one. Salesforce's Flex Credits sell at $500 per 100,000 credits, with a standard agent action costing 20 credits and a voice action costing 30, so the per-action price works out to $0.10 and $0.15 respectively. The total bill under this model depends entirely on how many actions one conversation takes, which is the one number a vendor's pricing page never publishes up front.

Per minute is how voice agents bill, since a phone call does not have a "resolution" the way a chat does. Retell AI runs $0.07 to $0.31 per minute all-in with no separate platform fee, built from component costs of roughly $0.055 per minute for voice infrastructure, $0.015 for text-to-speech, $0.015 for telephony, and $0.0016 to $0.64 for the underlying language model depending on which one you pick. Vapi unbundles those same pieces: a $0.05 per minute platform fee plus pass-through rates for transcription ($0.0095 to $0.0099/min), the language model ($0.0077 to $0.0452/min), voice ($0.0146 to $0.0238/min), and telephony ($0.008 to $0.014/min), landing around $0.09 to $0.14 per minute all-in before you choose a pricier model.

What customer-facing text agents actually cost at the same volume

To make these comparable, we ran one illustrative scenario, a business fielding 800 customer conversations a month with 560 of them resolved by the agent, through each vendor's own published rate. This is a standardized scenario for comparison, not a claim about what any specific business will see.

Zendesk's per-resolution rate is not published by Zendesk itself. Independent trackers and customer-reported billing converge on roughly $1.50 per resolution on a committed annual plan and $2.00 pay-as-you-go, on top of the published Team ($55/agent/month) or Professional ($115/agent/month) seat tiers, each including a small allowance of 5 to 15 resolutions per agent before overage applies.

Notice that the cheapest and most expensive rows in that table are selling a comparable job. The spread is the pricing model, not the underlying technology.

What AI voice agents cost per minute, and what the rate hides

Voice is billed per minute because a phone call has duration instead of a resolution state, but "per minute" hides almost as much as "per conversation" does. A flat quoted rate like Retell AI's $0.07 to $0.31 per minute already bundles voice infrastructure, speech-to-text, text-to-speech, and the language model into one number. A BYOK (bring your own key) platform like Vapi unbundles the same four pieces and lists each one separately, which looks cheaper on the platform-fee line and is not actually cheaper once you add transcription, model, and telephony costs back in.

For a business running 500 calls a month averaging three minutes, that is 1,500 minutes. At Retell AI's bundled rate that lands between $105 and $465 a month with no separate platform fee. At Vapi's unbundled rate, roughly $0.09 to $0.14 per minute once every pass-through cost is added, the same 1,500 minutes runs $135 to $210 a month, plus whatever the $999 minimum monthly commitment on Vapi's Pro plan adds if you need it. Neither number is wrong. They are answering slightly different questions, and only one of them is the number you will actually pay.

What internal AI copilots cost per seat

Tools built for a business's own staff, not its customers, mostly stay in the per-seat lane, because the usage pattern (a developer typing in an editor all day) does not map cleanly to a conversation or a resolution. GitHub Copilot Business costs $19 per user per month and Enterprise costs $39, each seat now bundling that same dollar amount in pooled AI credits consumed by token usage rather than a fully flat fee, since GitHub's June 2026 pricing change. For a 10-person team, that is $190 or $390 a month, flat, until a handful of heavy users burn through the shared credit pool faster than the rest of the team and the whole org has to upgrade a tier to keep everyone working. Predictable pricing is still pricing with a ceiling.

Questions to ask before comparing two vendor quotes

A quote is only comparable to another quote once you know these five things:

• What exact unit is billed: seat, conversation, resolution, action, or minute.
• What counts as a billable event. Intercom's "outcome" includes a handoff to a human, not just a clean resolution, which inflates the billable count compared to a model that only counts true resolutions.
• What is bundled into the quoted number versus billed separately. A voice vendor's $0.07 per minute quote is meaningless without knowing whether telephony and the language model are already inside it.
• What the required minimum platform or seat fee is, since several of these models cannot be purchased standalone.
• What happens past the included allowance: an overage rate, a hard cap, or a forced tier upgrade.

What actually moves your bill: resolution rate, not sticker price

The single number that determines whether per-resolution or per-conversation pricing wins for a given business is the agent's own resolution rate, not either vendor's advertised rate. A business whose agent resolves 40 percent of conversations without a handoff comes out ahead on a per-resolution model, because three out of five conversations cost nothing. A business running the same agent at a 75 percent resolution rate can come out ahead on a flat per-conversation rate instead, because the per-resolution vendor is effectively charging a premium for its own success rate. Ask a vendor for your own expected resolution rate on your own ticket mix before signing, not the industry average on their homepage.

Setup and integration cost sit outside all of these per-unit rates entirely and vary by how much of your existing stack the agent has to connect to. Our AI setup and deployment work for a hosted assistant connected to email, calendar, and a CRM typically runs on a multi-week timeline before the first billable conversation ever happens, and that cost does not show up in any vendor's per-minute or per-resolution rate card.

Frequently asked questions

Is per-resolution pricing always cheaper than per-conversation pricing?

No. Below roughly a 50 percent resolution rate, per-resolution pricing tends to cost less because you are not paying for conversations the agent fails to close. Above that threshold, a flat per-conversation rate can cost less, since the per-resolution vendor is charging more per success as the agent gets better. Ask a vendor to model your own expected resolution rate, not an industry-average figure.

Why doesn't any vendor publish a side-by-side price comparison with competitors?

A side-by-side comparison only favors the vendor publishing it if their own model wins at a typical customer's volume. Every comparison we found during research, including one published directly by a vendor, compared platforms within one category, such as customer-service chatbots, and none crossed over into voice agents or internal copilots, which is the gap this piece fills.

Does per-seat pricing include the AI usage, or is that billed separately?

It depends on the vendor and increasingly it is both. GitHub Copilot's per-seat price used to be a flat fee; since June 2026 each seat also bundles a pool of AI credits consumed by actual token usage, so a seat fee that looks flat on the pricing page can still run into usage limits if your team's actual consumption outpaces the included pool.

What is a "credit" or "action" and why don't vendors publish how many a conversation uses?

A credit or action is the smallest billable unit inside a platform like Salesforce's Flex Credits, where a standard agent action costs 20 credits and a voice action costs 30, at $500 per 100,000 credits. Vendors rarely publish how many actions a typical conversation consumes because it varies by how many steps the agent takes to resolve a given request, which means the real per-conversation cost under this model is not knowable from the rate card alone.

How much does setup cost on top of the per-unit rate?

Setup and integration cost is separate from any per-seat, per-resolution, or per-minute rate and depends on how many systems the agent has to connect to, such as a CRM, a calendar, a help desk, or a phone system. A narrowly scoped single-purpose agent can go live in a few weeks; an agent wired into several existing systems takes longer and costs more before the first billable conversation happens.

Should a small business pick the cheapest pricing model?

Not in isolation. The cheapest rate per unit at low volume can become the most expensive rate per unit once volume grows, particularly with per-conversation or per-action billing that scales linearly with no cap. Model your actual expected monthly volume and resolution rate against each vendor's real rate card before signing, the same way this piece modeled an 800-conversation month across five vendors.

Picking the right model matters more than picking the cheapest number on a pricing page. If you want that modeling done against your own numbers instead of a standardized example, get in touch and we will run it before you sign anything. For a look at what we build once a pricing model is settled, see our custom AI agents and voice agent work, or the full AI Studio line.