AI Customer Service Pricing in 2026: Per-Resolution vs Per-Seat, Modelled

Every comparison of AI support pricing you can find was written by somebody selling AI support. That is not a reason to dismiss them, but it does explain why they all compare the same three options and never the fourth.

Who writes the pages that rank for this

On the day we checked, page one of Google for per-resolution versus per-seat pricing was eight results. Every one belonged to a company selling AI support software, and three of the eight belonged to a single vendor: Fin, which charges per resolution.

That is not a scandal, it is an incentive. But it means the framing you are reading everywhere was chosen by someone with a preferred answer, and it explains a specific omission we will come back to.

The five models, and what each one actually meters

Vendors describe these as pricing options. They are better understood as five different answers to the question of what you are buying.

The third row is the one that catches people, and it deserves stating plainly: under per-resolution pricing, every improvement you make to your AI increases your invoice. Better documentation, wider scope, a higher confidence threshold. You are paying for success, so success costs more.

None of the other four models have that property. It is not a criticism of the vendors who use it — paying only for outcomes is genuinely fairer than paying for attempts — but it does mean your incentive to tune the system and your incentive to control cost point in opposite directions.

The published rates, in one place

Here is what the major vendors publish. All read from vendor documentation and dated; check the current page before committing to anything, because these move.

Two things stand out before any arithmetic. The published per-resolution rates run from $0.60 to $2.00 — a spread of more than three times on what is nominally the same unit. And the enterprise tier does not publish at all, which means the cheapest and most expensive options in this category cannot be compared on the same page.

The comparison everyone publishes

Take the standard worked example: a 50-agent team handling 20,000 AI resolutions a month.

Fin wins that table by a wide margin, and the table is published by Fin. The numbers are not wrong. The comparison set is.

The option the comparison leaves out

Every platform in that table meters AI usage. Nobody includes the alternative: a per-seat help desk where the AI is billed per session rather than per outcome.

Freshdesk Pro is $59 per agent per month. Freddy AI is $0.49 per session, with the first 500 included once per account. A session is not a resolution — Freshworks bills when a session starts regardless of how it ends, so to reach 20,000 resolutions at a 60% resolution rate you need roughly 33,000 sessions. Let us count it honestly, at the rate that flatters Freshdesk least.

Against Fin's $19,800 a month, that is 3.2 times cheaper. The annual gap is $162,200 — for the same 20,000 resolved conversations, on a comparison that does not appear on any page ranking for this query.

Does the gap close at smaller scale?

No. It is remarkably stable, and it only converges when volume is small enough that neither number matters.

Which produces a conclusion the category does not print: per-resolution pricing is not a cost optimisation. At every volume that matters it costs multiples of the per-seat alternative, and the gap does not narrow as you grow.

So why does anyone buy it?

Because cost is not the only axis, and this is where we should be fair to the vendors.

• Capability. Fin is widely reported as the most capable out-of-the-box agent available, including by people criticising its price. Freddy is competent. Those are not the same product.
• Bundling. Fin's $0.99 is stated to include the helpdesk and copilot. Zendesk's $1.50 sits on top of $55–169 seats plus $50 copilot plus $35 QA, which is why Zendesk looks worse than Fin in every table including its own competitors'.
• Risk transfer. Paying only for successful outcomes means a bad month costs you nothing. A per-seat licence bills whether the AI worked or not.
• Speed. Per-resolution products deploy in hours. Nobody is negotiating a procurement cycle for them.

That is the honest trade. You are buying capability and speed at roughly a 3x premium over the per-seat route. For many teams that is worth it. It is simply not the same claim as saving money, and the pages ranking for this query consistently present it as the latter.

The costs that never appear in the comparison

Four of them, and between them they routinely add more than the difference between two vendors' headline rates.

1. The help desk underneath

Several per-resolution agents assume you already pay for a platform. Zendesk's $1.50 per resolution is on top of $55–169 per seat. On a 20-agent team that is $1,100 to $3,380 a month before a single resolution is billed.

2. Add-ons priced per agent while the AI is priced per outcome

Zendesk Copilot at $50 per agent and Zendesk QA at $35 per agent are billed on headcount even though the thing they support is billed on usage. Grow the team and these rise; grow the AI and the resolutions rise. You are exposed on both axes at once.

3. Implementation

At the enterprise end this is a line item rather than an afterthought, and it can rival first-year licences. Sierra is reported at $150k+ annually and Decagon at a $50k+ platform fee before usage. Neither publishes, so neither figure is auditable — treat both as indicative.

4. Engineering

Anything that queries your own systems is development work. Answering from your help centre is configuration; looking up an order, checking a subscription or issuing a refund is an integration project. This never appears on a pricing page and it is frequently the largest number in year one.

Model your own, in five minutes

You do not need a spreadsheet from a vendor. We built a calculator that does this against published rates for 18 platforms — but the arithmetic is simple enough to do by hand, and doing it by hand is how you understand it.

• A - your monthly ticket volume.
• B - the share you believe AI can resolve. Use 15% for year one unless your documentation is genuinely current, in which case 30%.
• C - your agent headcount.
• D - the per-resolution rate you have been quoted.
• E - the per-seat rate of the platform you would otherwise buy.

Per-resolution cost is A x B x D. Per-seat cost is C x E plus any session charge. Compare, then re-run with B doubled, because that is what happens if the deployment goes well.

That second run is the one that matters and almost nobody does it. Under per-seat pricing, doubling B changes nothing. Under per-resolution it doubles your bill. A deployment that succeeds beyond expectations is a budget problem under one model and a straightforward win under the other.

A worked example on a mid-size team

Twelve agents, 9,000 monthly tickets, quoted $0.99 per resolution, against Freshdesk Pro at $59 a seat with Freddy sessions at $0.49.

The crossover sits near 55% on this team, not 30%. Twelve seats at $59 is $708; 9,000 tickets open 9,000 sessions, 500 are included once per account, and the remaining 8,500 bill at $0.49 for $4,165. That is $4,873 flat, whatever the agent achieves. Per-resolution stays cheaper until the agent is resolving more than half of everything, which is above what most deployments reach.

Notice the per-seat column does not move. Ticket volume drives the session charge, not AI performance, which is exactly why the comparison inverts as the deployment improves.

If a quoted per-resolution deal only works at a low resolution rate, you are being sold something that is cheapest when it fails. Worth saying out loud on the call to hear the answer.

What to ask before signing

Five questions. The first two decide your invoice and neither is answered on any pricing page we have read.

• What exactly triggers a billable event, in writing? Specifically: what happens when a customer abandons mid-conversation, and what happens when the same person returns two days later about the same issue.
• Is a conversation that the AI escalates to a human billed? Some vendors charge for the attempt.
• What does this cost at three times our current volume? Model it rather than accepting a rate.
• What is required underneath — seats, platform fees, minimums, annual commitments?
• What is the rate at renewal, and is there a cap on increases?

The billable-event definition is the single most important term in the contract and the one least likely to be published. If a vendor will not put it in writing before your trial ends, that is information.

How to choose, in one paragraph

If your volume is small or seasonal, per-resolution is fine and the absolute numbers are low enough not to matter. If you want the most capable agent available and can absorb a 3x premium for it, buy per-resolution knowingly rather than because a comparison table told you it was cheapest. If cost per resolved conversation is genuinely what you are optimising, a per-seat help desk with session-priced AI wins at every volume above trivial, and nobody ranking for this query will tell you that.

All rates read from vendor documentation and published roundups in July 2026 and cited below. We do not run billable volume through these platforms — the arithmetic is ours, the rates are theirs, and both are checkable.

WRITTEN BY AR · UPDATED 2026-07-31

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

Editorial policy