Three of the largest vendors in customer service changed what they bill you for this year, and they did not pick the same definition. Zendesk now sorts every AI conversation into three tiers and charges for exactly one of them. Salesforce started charging per autonomous resolution in July, with nothing owed if the customer walks away unhappy. Fin still counts twenty-four hours of customer silence as a resolution and invoices it.
Same word. Three prices. If you are replacing an IVR and a deflection bot this quarter, the definition in the contract will move your bill further than the rate will.
The verdict, before the analysis
Buy on the resolution definition, not the resolution rate. The cheapest headline number in this market belongs to the vendor with the loosest test, and that is not a coincidence.
Zendesk's "verified resolution" is the strictest definition any vendor publishes — an LLM has to confirm the customer's request was actually satisfied — and it is the benchmark to hold every other quote against, even if you never buy Zendesk. Salesforce Agentforce Help Agent is the right default if your service org already runs on Service Cloud, because it refuses payment on escalation, negative feedback and abandonment. Amazon Connect is an order of magnitude cheaper than everything else here and will cost you two platform engineers you do not currently have. Fin has the lowest per-resolution rate and the weakest test behind it.
The losers are Genesys Cloud CX and NICE CXone. Both charge per seat for the seats the project exists to remove, then meter AI consumption on top — and neither publishes the consumption rate on its pricing page. A per-seat platform with an unpublished AI meter is structurally misaligned with a deflection program.
Normalised to one workload — 1,000,000 contacts a year: 700,000 chat conversations at eight messages each, 300,000 voice calls at a five-minute average, 200 human agents still on the floor, and 500,000 issues genuinely resolved end to end — this is what the meters do. Telephony, implementation and professional services are excluded from every row.
| Platform | Billed unit | Billed when the AI fails? | Rate published? | Annual at this workload |
|---|---|---|---|---|
| Amazon Connect | Per voice minute / per message | Yes — every second meters | Yes, full rate card | ~$185K + telephony |
| Fin | Per resolution, incl. 24h silence | Yes — silence, configured handoffs and disqualifications all bill | Yes, $0.99 | ~$495K+ (chat only) |
| Genesys Cloud CX 3 | Per named seat + AI tokens | Yes — seats bill regardless | Seats yes, token rate no | $372K + tokens |
| NICE CXone Ultimate | Per agent + per session | Yes — seats and sessions bill regardless | $249 + $0.25 | ~$848K |
| Salesforce Agentforce Help Agent | Per autonomous resolution | No — escalation, negative feedback, abandonment are free | $2 per resolution | ~$1.0M + Service Cloud seats |
| Zendesk AI agents | Per LLM-verified resolution | No — assisted escalation and containment are free | Not on the pricing page | ~$1.15M |
| Sierra | Per agreed-in-advance outcome | "In most cases, there's no charge" | Nothing published | Not quotable |
All rates checked on vendor pricing pages on 2 September 2026.
Containment and resolution are now separate line items
A containment rate is the share of contacts that end inside the automated channel without transferring to a human. It says nothing about whether the customer's problem was solved. CallMiner's framing is the right one: when success is defined by what does not happen, the system optimises for the non-event, and a bot that refuses to transfer scores better than one that escalates cleanly. Adjusting for repeat contacts typically opens a 5% to 15% gap between raw containment and genuine resolution. A 2026 synthesis published by Aissist.io — a vendor that sells agentic customer-service AI, and which pooled more than 40 published sources including vendor disclosures rather than measuring programs directly — puts the inflation higher still, at 20 to 40 points versus counting only issues solved end to end, against a cross-program automation median near 41%. Weigh that one accordingly; it is an estimate assembled from the same vendor disclosures it is criticising.
What changed in 2026 is that this stopped being a metrics argument and became a billing schedule.
Zendesk split the outcome into three tiers in a change introduced on 18 May 2026. An assisted escalation — "the AI agent contributed to the interaction before a human agent completed the resolution" — does not count against your allowance. A contained resolution, where "the AI agent handled the interaction to completion without the customer requesting further assistance," also does not count. Only a verified resolution, which passes an LLM check that "the customer's request was satisfactorily resolved," is billable. Read that again: Zendesk built a product that detects containment and then declined to charge for it.
Salesforce arrived at the same place from the opposite direction. The original Agentforce model was $2 per conversation, and Salesforce Help still lists it alongside Flex Credits at "$500 USD per 100,000 Credits," with "one Agentforce action" consuming 20 credits, or $0.10. That model billed the attempt. Agentforce Help Agent, announced 25 June 2026 and generally available in July, charges $2 per autonomous resolution instead — and per the launch coverage, "if a customer requests human escalation or provides negative feedback, no charge is applied." Reporting on the announcement adds abandonment to the no-charge list.
Fin did not move. Its pricing page lists $0.99 per resolution, the lowest published rate in this comparison, and Fin's own help centre, last updated 19 June 2026, defines two kinds. A confirmed resolution is one where the customer says something affirmative. An assumed resolution is triggered when "a customer disengages from the conversation for 24 hours after Fin's last answer." Silence is billed as success. There is a clawback — if the customer reopens that conversation, "that resolution will be deducted and not charged" — but it keys on the same conversation. A customer who gives up and phones instead, or opens a fresh ticket, does not trigger it. Resolution is also not the only billable outcome: the same page charges $0.99 for a procedure handoff — "Fin completes a Procedure you've configured to hand off to a human" — and $0.99 for a disqualification, with qualifications at $9.99. An unconfigured transfer is free, but a routed-to-a-human contact that Zendesk files under assisted escalation and does not charge for is $0.99 on Fin.
Sierra invented this category and still publishes no rate — it has no pricing page at all. Its outcome-pricing post says that if a case is escalated or unresolved, "in most cases, there's no charge," and that criteria are "clear, agreed-upon" per customer. The hedge is doing real work in that sentence, and it is negotiable — which is the point of asking. Decagon publishes nothing either; its pricing URL 404s as well.
Price the abandonment, not the resolution
Run the gap through the meter. At 500,000 billed resolutions on Fin, a 15% overstatement means roughly 75,000 of them were customers who went quiet rather than customers who were helped. That is about $74,000 a year paid for failures — trivial against the $495,000 subtotal, and completely beside the point.
The real cost is the second contact. Those 75,000 people come back through a channel you also pay for, and it is usually the expensive one. On Amazon Connect's published rates a five-minute voice call costs $0.038 per minute for the voice service, $0.008 per minute if self-service AI is enabled and $0.015 per minute for conversational analytics — about $0.31 in platform cost — before you add telephony and a human agent's time. The AI vendor books the revenue on the failure; you book the cost of it twice.
This is the number to instrument before you sign anything: the 24-to-72-hour repeat-contact rate on conversations the vendor marked resolved, split by channel. Not CSAT, which the abandoned customer never fills in. Not containment. If a vendor will not expose that in a proof of concept, that is your answer about what its resolution definition is worth.
The handoff is where the money leaks
Escalation quality is the second-order failure, and the gap between what buyers believe and what customers experience is the widest measured number in this market. Five9's 2026 Business Leaders CX Report — vendor-published research from a CCaaS provider that sells the agent-assist and handoff tooling the gap argues for, surveying 600 decision-makers and 3,000 consumers across the US, UK and Germany — found that 96% of CX decision-makers say their organisation preserves context effectively during AI-to-human handoffs, while 83% of consumers say they have to repeat themselves after being transferred, 35% of them "often or always."
The same survey has those decision-makers admitting the components: 37% say agents lack visibility into the prior AI interaction, 34% cite context loss when switching channels, and 30% blame redaction for stripping context. That last one is not a bug. It is your compliance controls stripping the transcript, and it means the handoff quality you buy is partly a function of how you configured PCI masking — a detail no demo will show you.
It also decides whether the deflection program survives contact with customers. In the same report, 33% of consumers rate themselves very or extremely frustrated at the point of transfer, and consumer trust runs at 55% when there is a clear path to a human versus 26% without one. Hiding the escalation path is the fastest way to raise containment and lose the account.
Voice is a latency problem before it is an AI problem
Every vendor here will demo voice. The physics are worse than the demo suggests, and they are not vendor-specific.
Time to first audio byte — the silence a caller sits through after they stop speaking, measured from a recording of a real phone call rather than an API timestamp — is the metric that decides whether a voice agent feels usable. A public benchmark whose last call was placed on 1 August 2026 measured five voice-agent infrastructure platforms over 432 attempted turns each — it publishes its caller harness and every per-turn timing, though it does not disclose who operates it — and found no median below one second: Telnyx 1,296 ms, ElevenLabs 1,424 ms, Bland AI 1,520 ms, Vapi 1,558 ms and Retell AI 1,740 ms, with p95 figures running from 1,768 ms to 2,259 ms. Human conversational turn-taking averages roughly 200 ms.
Those are the infrastructure layers. A CX suite built on one starts from that floor and adds its own orchestration, retrieval and CRM lookups on top. Which is why the only honest voice evaluation is a recorded call to the vendor's own reference customer, with a stopwatch on the gaps — and why our own eleven-stack voice containment test found no clean winner.
Voice also prices differently. ElevenLabs Agents lists $0.080 per minute with burst at $0.160, noting that "the LLM model and any telephony are billed separately on top." Fin voice is not publicly priced at all — "available to select customers working directly with our sales team." Amazon Connect's $0.038 voice minute plus $0.008 for self-service AI is the only fully published voice rate card in this group. Do not let a vendor quote you a per-resolution rate for chat and a per-minute rate for voice without normalising both to cost per contact; the arbitrage between the two meters is where a deflection business case quietly dies.
What each of these actually costs at 1M contacts
The arithmetic, run on published rates as of 2 September 2026.
Amazon Connect — about $185,000. From the customer pricing appendix: 1.5M voice minutes at $0.038 ($57,000), self-service AI voice at $0.008 ($12,000), conversational analytics at $0.015 ($22,500), 5.6M chat messages at $0.010 ($56,000), self-service AI chat at $0.0015 ($8,400), and performance evaluations at $12 per agent per month for 200 agents ($28,800). Billing is per second with a ten-second minimum. Telephony is extra. There are no seat licences at all.
Fin — about $495,000, plus outcomes this figure does not count. 500,000 resolutions at $0.99. Procedure handoffs and disqualifications bill at the same $0.99 and qualifications at $9.99, none of which are modelled here — so treat $495,000 as a floor, not an estimate. No setup, integration or platform fees when running against a non-Intercom helpdesk, per its pricing page. Voice is quoted separately.
Genesys Cloud CX 3 — $372,000 plus tokens. Genesys publishes CX 1 at $75, CX 2 at $115, CX 3 at $155 and CX 4 at $240 per user per month. AI Experience tokens are the AI meter: "250 named / 350 concurrent tokens per org/month included," with CX 4 adding "+30 named / +39 concurrent tokens per agent/month." Virtual agents, native bots, predictive routing, knowledge and Copilot all consume them. The page does not publish a per-token price. Get it in writing before you model anything.
NICE CXone Ultimate — about $848,000. NICE lists five suites from Omnichannel at $110 to Ultimate at $249 per agent per month, with Ultimate carrying an additional $0.25 per session. The page does not define "session." If it means every contact, that is $250,000 on top of $597,600 in seats at this volume. If it means only AI-handled sessions, it is less. That ambiguity is worth more than the discount your rep is offering.
Salesforce Agentforce Help Agent — about $1.0 million, at $2 per autonomous resolution across 500,000 resolutions, plus Service Cloud licences for the 200 remaining agents. If your conversations average fewer than 20 actions, Flex Credits at $0.10 per action work out cheaper than the legacy $2-per-conversation model; above 20 actions they do not.
Zendesk AI agents — about $1.15 million. Zendesk's pricing page lists Suite Professional at $115 per agent per month billed annually and Copilot at $50 per agent per month paid yearly, but does not publish a per-resolution rate. A Zendesk partner quotes $1.50 per resolution; treat that as a contract-negotiated figure, not a list price. At $1.50 the resolutions are $750,000, seats $276,000 and Copilot $120,000.
Two costs sit outside every row above. Professional services are the big one: an enterprise CCaaS TCO guide published 30 April 2026 cites Forrester's estimate that "professional services fees add another 50 to 200 percent of first-year subscription costs," with migration running $20,000 to $100,000 mid-market and $700,000 and above for large enterprises. The second is telecom surcharges, which the same guide puts at 10 to 15 percent of base platform cost. A $185,000 Amazon Connect bill with $370,000 of integration work behind it is not the bargain the rate card implies — though it is still the cheapest option here, and it is the only one where the meter cannot be renegotiated upward at renewal.
Who should not buy each of these
This is the section vendors skip.
Do not buy Amazon Connect if you do not have at least two engineers who can own contact flows, Lex intents and Lambda integrations as a permanent job. The rate card is honest and the platform is genuinely modular, but nothing is assembled for you. A CX team without platform engineering will spend the savings on a systems integrator and lose the flexibility that justified the choice.
Do not buy Fin if a meaningful share of your contacts are the kind a frustrated customer abandons rather than escalates — cancellations, billing disputes, outage complaints. Assumed resolution bills silence, and those are exactly the queues where silence means the opposite of success. Fin is strong where questions are answerable and customers say thank you.
Do not buy Genesys or NICE if the business case is headcount reduction. You are buying a per-seat licence with a consumption meter bolted on, and the seat count is the thing you promised to shrink. Both are the right answer if you need mature workforce management, omnichannel routing, recording and compliance as one estate — that is a real product and these two build it well — but that is a modernisation project, not a deflection project. Be honest about which one you are funding.
Do not buy Salesforce Agentforce if you are not already on Service Cloud. The resolution definition is good and the no-charge conditions are the most buyer-favourable published, but the value depends on the CRM data being there already. And note the pricing history: the same product shipped a $2-per-conversation meter, took two years of customer complaints, and replaced it. Price the possibility of a third model.
Do not buy Zendesk if you expect the strict verification tier to be free. The definition is the best in the market and you will pay for it — verified resolutions plus seats plus Copilot is the most expensive row in the table. That is a defensible trade if you are buying auditable outcomes. It is a bad trade if you were buying cheap deflection.
Do not buy Sierra or Decagon without a written outcome definition in the MSA. Neither publishes a rate, both sell custom contracts, and "in most cases, there's no charge" is a sentence with an exception clause in it. Ask what the exceptions are, in writing, before the pilot.
Recording, redaction, and the deadline that already passed
Compliance is the constraint that quietly eliminates options, and it moved in 2026.
Under PCI DSS v4.0.1, as of 31 March 2025, disk-level encryption no longer satisfies PAN protection in call recordings — the primary account number must be rendered unreadable in all stored formats, including audio. That is a real-time redaction requirement, not a post-processing one, and it is the same redaction 30% of the decision-makers in the Five9 data blame for losing handoff context. You are buying both sides of that trade-off whether you price it or not.
The FCC's February 2024 declaratory ruling put AI voice agents inside TCPA's "artificial or prerecorded voice" restrictions with no carve-out for AI that simulates a live agent. Disclosure law has stacked up on top: Utah S.B. 226 narrowed that state's AI Policy Act rather than widening it: in an ordinary consumer transaction the disclosure is owed only when the customer clearly and unambiguously asks whether they are talking to AI, with proactive disclosure required for regulated occupations and high-risk interactions — financial, legal, medical and mental-health contexts — and available everywhere else as a safe harbour. California AB 2905 requires the live-voice announcement that precedes an automated outbound message to state that the recording uses an artificial voice, and Texas TRAIGA took effect 1 January 2026. EU AI Act Article 50 transparency obligations took effect on 2 August 2026 — if you take calls from EU customers, the agent has to disclose that it is an agent, and that requirement is already live.
Three procurement questions follow. Where does redaction happen — before the transcript is written, or after? What exactly survives redaction into the handoff payload? And can the disclosure be configured per jurisdiction, or is it one global toggle? Vendors that answer the first two precisely are usually the ones that have been through an audit.
What changes the answer
Four things move this recommendation, and none of them is a feature comparison.
Your existing CRM. If service already runs on Service Cloud, Agentforce's integration advantage outweighs a rate difference. If you run Zendesk, the verified-resolution tier is right there. If you run neither, that is the case for Amazon Connect or a helpdesk-agnostic layer.
The Salesforce–Fin transaction. Salesforce signed a definitive agreement on 15 June 2026 to acquire Fin for approximately $3.6 billion, bringing "an established global customer base of more than 30,000 companies" and an agent "resolving on average 76% of support volume end-to-end." The deal is expected to close in Salesforce's fiscal Q4 2027 and is not closed today. If you sign a multi-year Fin contract this quarter, you are signing with a company that will be owned by Salesforce inside your renewal cycle. Get change-of-control and pricing-continuity language in the MSA — the same clause that matters in every AI vendor acquisition.
Voice share. Above roughly 40% voice, the per-minute economics dominate and Amazon Connect's published rate card starts winning outright. Below 20%, chat-first outcome pricing is simpler and the latency problem mostly disappears.
Whether you can measure repeat contact at all. If your data warehouse cannot join a resolved AI conversation to the same customer's next contact across channels within 72 hours, you cannot audit any vendor's resolution claim, and every model above is theatre. Build that join first. It is a week of work and it is the only thing on this page that survives a vendor switch.
The Bottom Line
The last time this industry agreed on a metric, it was average handle time, and it took a decade to unwind the damage of paying people to end calls quickly. Containment is the same shape: a proxy that is easy to measure, easy to game, and pointed at the wrong party. The difference this time is that the proxy is printed on an invoice, which at least makes it negotiable.
The market has already conceded the argument. Zendesk built a tier specifically to not charge for containment. Salesforce rebuilt its pricing twice to stop billing attempts. Sierra says it does not charge for escalations. Those are not marketing positions, they are admissions — and they tell you exactly which number the vendors think will not survive a customer's audit.
So audit it. Instrument the 72-hour repeat-contact rate before the pilot, write the resolution definition into the contract, and make the vendor's clawback cover a new conversation and not just a reopened one.
Buy the definition. The rate is what you negotiate afterward.
Continue Reading
- Agentic AI Pricing: Don't Buy Consumption Without a Cap
- Eleven Voice Agents, One Bank Call, No Clean Winner
- OpenAI Voice AI: $0.30 Calls vs $17 Human Agents
- What Is Fin AI Agent? Salesforce's $3.6B CIO Breakdown
- Yellow.ai Is Going Public to Buy Your Outsourcer
- Salesforce Agentforce: We're Not Integrating, We're Replacing
- Agents Averaged 73. Only 30% Were Usable. Grade Pass/Fail.
- Why 50% of Voice AI Costs Vanish with Owned Infrastructure