AI customer service agents: what they really cost per resolved conversation

The short version

At a realistic mid-market load — 10,000 support conversations a month, 60% handled without a human — the same work costs $0.09 per resolved conversation on Crisp and $1.50 on Zendesk. Same tickets, same outcome, 16× the bill.

That gap is not a quality gap. It is a unit gap. Every vendor in this category has picked a different thing to charge for — a resolution, an outcome, a session, an action, a credit, a minute — and the headline price only becomes comparable once you convert all of them into the one number a support lead actually budgets against: cost per conversation the AI closed on its own.

No vendor does that conversion for you. So here it is.

Everything below is verified against vendor primary sources on 13 August 2026.

What each vendor actually charges for

"Billed when the AI fails?" is the column vendor pages bury.

Eight of the 21 tools we list in this category publish no price at all. That is itself a finding: if a vendor won't name a unit in public, you're negotiating without a reference point, and the first number you hear will be anchored to your ticket volume rather than to their cost.

The arithmetic vendors won't do for you

Assumptions held constant across every row: 10,000 conversations a month reach the AI first, 60% are closed without a human, human seat licences excluded. This is the AI line item only.

Three things fall out of this that no vendor page will tell you.

The success-based pricing is the expensive pricing. Intercom and Zendesk only bill when the AI wins — and they are the two priciest options per win by a wide margin. You are paying a premium for risk transfer. Whether that premium is worth it depends entirely on how good your knowledge base already is. If you expect a 70% resolution rate, you're buying insurance against an event that mostly won't happen.

Session pricing punishes a bad quarter twice. Freshdesk bills $0.49 whether Freddy resolves the ticket or not. At 60% that's $0.78 per win. At 30% — an ordinary first-quarter number for a thin knowledge base — the identical contract costs $1.55 per resolved conversation, more than Zendesk's success-only rate. The cheaper-looking unit gets most expensive exactly when the deployment is going badly, which is when you can least defend it.

Salesforce's per-action model is a volatility bet. At 3 actions per conversation it's $0.50 a win. At 8 actions — a returns workflow that checks an order, validates a policy, issues a credit, updates a record and sends a confirmation — it's $1.33. You discover which number you're on after you've built the agent, not before.

Ignore every deflection rate you've been quoted

Support leads shop on resolution rate. The published rates are not comparable, and at least one headline number isn't measuring what a skimming buyer will assume.

Fin's benchmark page advertises an 85% average resolution rate, drawn from 110 million conversations across 12,000+ customers and 15 industries. Read the label: that 85% is the average of the top 10 performers per industry, not the average customer. The industry toplines on the same page — 98.2% for software and technology, 91.0% for education — are top performers too. It's a real number measuring something other than what it looks like at a glance.

The deeper problem is that "resolution" is defined by whoever is billing you:

• Zendesk requires an automated reply followed by a 72-hour quiet period — the customer doesn't reopen. Conservative, and it generally favours the buyer.
• Intercom Fin counts a resolution when "no further help is requested after Fin's last answer" — but also bills a procedure handoff, where Fin gathers context, takes an action, then passes the conversation to a human. Intercom moved from "resolutions" to "outcomes" in March 2026 and was explicit about why: work that ends in a handoff now generates revenue where it previously didn't. A qualification outcome — Fin matching a prospect to your criteria — bills at $9.99, ten times the support rate. Fin does state that a conversation simply passed to your team with no outcome is free.
• Freshworks needs no definition at all. It bills sessions.

A 70% resolution rate on Zendesk and a 70% on Fin are therefore not the same 70%, and neither is auditable by you unless you instrument it yourself.

The one contractual answer in this set: Tidio's Premium tier pairs pay-per-resolution billing with a guaranteed 50% Lyro resolution rate. It's the lowest bar of the group, and it's the only one written into the price.

Don't buy a phone system for deflection

Several tools in this category are priced per minute, because they're voice-first: PolyAI (per-minute, enterprise contract, no published rate), Voicefleet (€99/€299/€599 for 500/1,000/2,000 minutes), Workforce Wave, Vida. Aircall and CallHippo are cloud phone systems with AI features attached — they are not deflection engines and won't be mistaken for one after a demo.

Minute-based pricing inverts the incentive of resolution-based pricing: an agent that talks a customer through a long call costs more than one that fails in thirty seconds. If you're buying voice, the number to negotiate is average handle time, not containment.

Five questions to ask before signing

1. What is the billable unit, in writing? Not "outcome" — the exact trigger condition.
2. Do partial resolutions and handoffs bill? Fin bills procedure handoffs and leaves plain handoffs free. Make your vendor draw that line explicitly.
3. Is there a quiet period before a resolution is confirmed? Zendesk's 72 hours is the benchmark. No quiet period means a reopened ticket can bill twice.
4. What does this cost at a 30% resolution rate? Model your worst quarter. Session- and action-based pricing degrades badly there.
5. Who audits containment? If the only record is the vendor's dashboard, you have no recourse on a disputed invoice.

What we'd actually do

Under roughly 1,300 AI conversations a month, Crisp Plus at $295 is hard to beat on published pricing, and it includes the AI rather than metering it as an add-on. The caveat: Crisp's credits are consumption-based, so long or tool-heavy conversations will burn faster than the ~450-per-$25 estimate on their pricing page. Verify your own burn rate in month one rather than trusting the estimate.

Above that volume, CoSupport's $0.19 per resolution is the strongest published number in this category by a factor of five, and the flat $99/month server plan is enough of an outlier to be worth a pilot on volume economics alone. The risk you're pricing there is vendor risk, not per-ticket risk — CoSupport is a far smaller company than Zendesk or Intercom, and that belongs in the decision.

If you're already deep in Zendesk, Freshdesk or Salesforce, the honest advice is that switching costs will exceed the per-resolution delta for most teams. What's negotiable is the unit. Zendesk's $1.50 is a committed-volume rate, which means it moves. Go into that conversation with the table above and a number of your own.