Your numbers
Default true resolution rate = 56% (claimed 70% × kept 80%), applied to every vendor unless you override it per vendor below. Independent year-one medians sit nearer 35% — this default is on the optimistic side.
Pricing models to compare
Rates are US list prices, editable. Adjust to your own quote. “charges for unresolved” means you pay even for conversations the AI fails to resolve.
What it actually costs
Cheapest of 9 models selected$40,200
Lowest all-in cost/mo — Tidio / Lyro
$0.89
Lowest cost per true resolution — Tidio / Lyro
$477,600
Annual saving vs all-human — Tidio / Lyro
| Vendor | True res. | AI cost/mo | + Residual human/mo | Total/mo | Per true resolution |
|---|
| Tidio / Lyrocheapest | 56% | $5,000 | $35,200 | $40,200 | $0.89 |
| Gorgias AI Agent | 56% | $5,040 | $35,200 | $40,240 | $0.90 |
| Intercom Fin | 56% | $5,544 | $35,200 | $40,744 | $0.99 |
| Kustomer AI | 56% | $6,000 | $35,200 | $41,200 | $1.07 |
| Zendesk AI agents | 56% | $8,400 | $35,200 | $43,600 | $1.50 |
| Crescendo | 56% | $9,900 | $35,200 | $45,100 | $1.77 |
| Generic per-conversation | 56% | $10,000 | $35,200 | $45,200 | $1.79 |
| Ada / Forethought | 56% | $11,200 | $35,200 | $46,400 | $2.00 |
| Decagon (estimated) | 56% | $18,000 | $35,200 | $53,200 | $3.21 |
All-human baseline: $80,000/mo. “Residual human” is the cost of the tickets the AI does not truly resolve — they still reach a person — so a vendor that resolves a lower share carries a higher residual, which is the apples-to-apples correction: a cheaper unit price rarely buys the same resolution. Per-conversation models also bill the failures, which is why their AI cost is higher even when the unit rate looks lower. “Resolution” is vendor-defined; set each rate to what you measured on your data.
How this is calculated
A worked example: at 10,000 conversations a month, a 70% claimed resolution rate, a 20% reopen rate, and an $8 loaded cost per human ticket, the true resolution rate is 56% (7,000 claimed minus 1,400 that come back). Intercom Fin at $0.99 per resolution costs $0.99 per genuinely-solved ticket; a generic $1.00 per-conversation model costs $1.79, because it bills the 44% the AI never solved. Set each vendor’s resolution rate separately: a cheaper tool that resolves less leaves a bigger residual-human bill, which is what makes the comparison fair.
Every vendor calculator reduces to one formula: volume × resolution rate × cost per ticket, minus the AI bill. The lever that decides the answer is the resolution rate — and it is the number vendors inflate. This tool makes you split it. A containment number can look like 70–80% while the share of problems genuinely solved is far lower, because a closed conversation is not a solved problem.
The true resolution rate is your claimed rate minus the tickets that recontact within a week. Tickets that are not truly resolved still reach a human, so their cost is added back as “residual human cost.” That is why the all-in total can be close across vendors even when the headline unit prices look very different — and it is where the real ROI from AI customer service either appears or evaporates.
Per-resolution pricing only bills the successes; per-conversation pricing bills the failures too. For a fuller picture of unit economics, see the cost per resolved contact for AI vs a human agent.
What it leaves out
Implementation, integration, and ongoing prompt/knowledge maintenance are not modelled here — they typically add one to two times the first year’s licence cost. Neither is the CSAT cost of pushing deflection too hard. Both are deliberate next steps, not oversights. The defaults are sourced starting points; replace every one with your own numbers.
Common questions
What is the difference between deflection and resolution?
Deflection (or containment) counts conversations that ended without a human — including ones where the customer gave up or switched channel. Resolution counts the problems genuinely solved. The calculator separates them: it takes your claimed resolution rate, subtracts the share that recontacts within a week, and uses the remainder — the true resolution rate — for every cost figure.
Why does per-conversation pricing cost more than its unit rate suggests?
Per-conversation models bill every conversation the AI touches, including the ones it fails to resolve. So a $1.00 per-conversation rate can cost more per genuinely-solved ticket than a $1.50 per-resolution rate, because the per-resolution vendor only bills the successes.
How is cost per true resolution calculated?
It is the vendor's monthly AI cost divided by the number of genuinely-resolved tickets (volume × claimed resolution rate × (1 − reopen rate)). Tickets the AI does not truly resolve are added back as residual human cost, because they still reach a person.
Are the default prices accurate?
They are US list prices checked in 2026, with a source link on each model. Several vendors (Decagon, Sierra) price by negotiation, so those are marked estimated. Every rate is editable — replace it with your own quote.