What you're buying
The four buckets of contact center AI
"AI" on a CCaaS quote is not one product. It is four quite different things with different risk profiles, different economics, and — critically — different billing meters. Vendors rarely separate them on the proposal, which is exactly why quotes are hard to compare. Separate them yourself before you do anything else.
Bucket 1 · Faces the customer
Virtual agents and self-service
Conversational bots on voice and digital that try to resolve the request without a human: order status, password resets, appointment changes, balance enquiries. This is where the money is, because a contained contact costs a fraction of a handled one. It is also where the reputational risk is, because failures happen in front of the customer. Almost always metered on consumption — per minute, per message, per session, or per resolution.
Bucket 2 · Faces the agent
Agent assist
Live transcription, suggested responses, knowledge surfacing, next-best-action prompts, and automatic wrap-up summaries for a human who is already on the contact. Lower ceiling than self-service, but far lower risk and far more predictable: it shortens handle time, cuts after-call work, and measurably shortens the ramp for new hires. Usually a per-seat add-on, which makes it the easiest AI line to budget.
Bucket 3 · Faces the supervisor
Analytics, auto-QA, and forecasting
Automated quality scoring across 100% of interactions instead of the three calls a month a human QA analyst can review, plus sentiment analysis, topic and driver discovery, compliance checking, and AI-assisted forecasting and scheduling. The quiet winner of the category: it replaces a sampling process that was never statistically meaningful, and the ROI shows up in coaching quality and compliance exposure rather than in headcount.
Bucket 4 · Takes action
Agentic AI
The 2026 addition: AI that does not just answer but executes — issuing the refund, rebooking the flight, updating the record in your CRM, opening the ticket in your ITSM. This is the bucket with the biggest promise and the least mature governance. It needs write access to your systems of record, which turns an AI purchase into a security and audit conversation. Price and permission it accordingly.
The practical sequencing for most organizations is buckets 2 and 3 first, then 1, then 4. Agent assist and auto-QA pay back quickly, carry little customer-facing risk, and generate the interaction data that makes the customer-facing bots work later. Reversing that order is the most common way an AI programme stalls.
How it's priced
Four pricing models, and which one bites
Nearly every AI line on a CCaaS quote is one of four models. The difference between them is not cosmetic: it determines whether your AI bill scales with headcount, with contact volume, or with success — and therefore whether it is forecastable at all.
| Model | Typical shape | Budget risk |
|---|---|---|
| Bundled in the tier | Core AI included in the seat price, or in a top tier running roughly $160–$200+ per agent all-in. | Low. Predictable. Watch for a low ceiling on what "core" covers. |
| Per-seat add-on | $20–$60 per agent per month for agent assist or auto-QA, often as several separate SKUs. | Low. Scales with headcount, which you already forecast. |
| Consumption / metered | Per minute of analysed audio, per message, per session, or per "token" drawn from a monthly pool. | High. Scales with contact volume — exactly what spikes when things go wrong. |
| Per outcome / resolution | Roughly $0.99–$2.00 per resolved conversation, sometimes with a monthly minimum. | Medium. Aligned in principle — entirely dependent on the definition of "resolved." |
The one that bites is the third. A rate of a cent and a half per analysed minute reads as a rounding error in a proposal and turns into the largest variable line on the bill during a peak season, a product recall, or an outage — precisely the months when your contact volume triples and your budget is least flexible. Consumption pricing is not wrong; unmodelled consumption pricing is.
The one number to demand: ask every vendor to price the AI line at your current monthly volume, at double it, and at your single worst month in the last two years. If they can't or won't produce all three, you do not yet have a quote — you have a rate card.
By platform
How each platform meters AI
Published and third-party-tracked pricing as of August 2026. Rates move, and negotiated deals sit below list — the value here is the shape of each model, which changes far less often than the numbers.
| Platform | AI pricing model | What to watch |
|---|---|---|
| Dialpad | Core AI included in the seat price rather than sold as a premium add-on. | Simplest to budget in the market. Compare depth, not just inclusion. |
| RingCentral RingCX | Native AI (RingSense) and digital channels bundled into the seat. | Low entry price plus bundled AI makes the fully loaded comparison flattering. |
| Zoom Contact Center | AI Companion included at no additional cost across paid plans. | Genuinely no AI line item. Verify which features that covers on your tier. |
| Five9 | An AI minute allowance per seat (commonly cited at 3,000/seat) for transcription, summaries, and insights; virtual agents and AI agents billed on usage. | Hybrid model — the allowance covers agent-facing AI; self-service is metered. |
| Webex Contact Center | Cisco AI Assistant as a per-agent add-on (commonly $20–$35/agent/mo); Webex AI Agent metered on consumption. | Two separate meters. Confirm both are in the Flex Plan quote. |
| Genesys Cloud | AI Experience tokens: a monthly pool at the organisation level, topped up per licensed agent on higher tiers, with additional tokens sold on usage. | Different features burn tokens at different rates. Ask for the burn table. |
| NICE CXone | Enlighten AI sold as separate SKUs, tracked at roughly $30–$60 per agent per month each; the full suite can add $60–$120 on top of the base package. | The deepest suite, and the largest premium — potentially 50–100% over base. |
| Talkdesk | Some AI in higher tiers, some as priced add-ons on top of the seat. | Get the split in writing — tier inclusions have moved repeatedly. |
| Amazon Connect | Everything metered: Contact Lens analytics around $0.015 per analysed minute with volume tiering, chat per message, Amazon Q in Connect billed per message. | Purest consumption model. Cheapest at low volume, needs real modelling at scale. |
Read that table as four strategies, not nine products. The bundlers (Dialpad, RingCX, Zoom) trade AI depth for budget certainty. The metered platforms (Genesys, Amazon Connect) trade certainty for the ability to pay only for what you use. NICE sells the deepest capability at the clearest premium. Five9 and Cisco sit in between with a hybrid. None of these is wrong — but a bundler and a metered platform cannot be compared on seat price alone, which is the mistake that makes half of all CCaaS shortlists misleading. Base rates for every platform are in the 2026 CCaaS pricing guide.
The math
Does the AI line actually pay for itself?
Two business cases, both worth running before you sign. Use your own numbers — these are illustrative for a 60-agent operation handling roughly 30,000 contacts a month.
Case A — agent assist, per seat
Sixty agents at a $35/agent/month add-on costs $2,100 a month. The saving comes from handle time and after-call work: automatic summaries typically remove most of the wrap-up typing, and live knowledge surfacing shortens the search mid-call.
If a fully loaded agent costs you $28 an hour and each of them handles 45 contacts a day, shaving 20 seconds off average handle time frees about 15 minutes per agent per day — roughly $7 of capacity per agent per day, or about $9,200 a month across 60 agents at 22 working days.
That clears the $2,100 comfortably — if you actually convert freed capacity into either absorbed volume growth or reduced overtime. Capacity that just becomes idle time is not a saving; it is a slide. Decide in advance which of the two you are buying.
Case B — self-service, per resolution
A human-handled contact at $28/hour and 7 minutes average handle time costs roughly $3.27 in agent labour alone. A bot resolution at $0.99 to $2.00 is cheaper per unit — but only on genuinely contained contacts.
The trap is the escalated contact. If the bot attempts 10,000 contacts, resolves 4,000, and hands 6,000 to an agent, you have paid the resolution fee on 4,000 and the full agent cost on 6,000 — plus the handle time is often longer, because the customer arrives annoyed and the agent re-gathers context. At $1.50 per resolution that is $6,000 of AI cost against roughly $13,000 of avoided labour: still positive, but at half the containment rate it flips.
This is why the containment rate — not the per-unit price — is the number that decides the business case, and why you should model it at a conservative figure. Industry benchmarks put typical enterprise deflection near 40% of queries touched, with a much smaller share fully resolved end to end. Build the case at a rate you would be unembarrassed to defend, then treat anything above it as upside.
One more line most business cases omit: the work does not disappear, it changes shape. Bots absorb the simple contacts first, which means the contacts reaching your agents get harder, longer, and more emotionally demanding on average. Handle time on human-handled contacts goes up after a successful deflection programme. That is not a failure — it is arithmetic — but if your business case assumed a flat AHT, it will read as one.
Honest assessment
Where AI pays off, and where it's theater
Earns its money
- ✓ Automatic call summaries and wrap-up — immediate, measurable AHT saving
- ✓ Auto-QA across 100% of interactions instead of a 1% human sample
- ✓ Knowledge surfacing during the call, especially for new hires
- ✓ Self-service on genuinely high-volume, low-variance contact types
- ✓ Topic and driver discovery that tells you what to fix upstream
- ✓ AI-assisted forecasting and scheduling in shift-heavy operations
Usually theater
- — Real-time sentiment dashboards nobody acts on during the shift
- — A bot on a low-volume, high-variance queue — it will never see enough patterns
- — AI layered on an unmaintained knowledge base — it confidently repeats stale answers
- — Next-best-action prompts agents are not empowered to act on
- — Deflection targets set without a clean escalation path — CSAT pays the bill
- — Agentic AI with write access before anyone has defined the audit trail
The pattern is consistent: AI pays when it is attached to a process someone owns and a metric someone is accountable for. It becomes theater when it produces information nobody is positioned to act on. Before buying any AI SKU, name the person whose job gets easier and the number on their dashboard that should move. If you cannot do both in one sentence, you are buying a demo.
The prerequisite nobody sells you: AI quality is mostly a function of your knowledge base, your CRM data hygiene, and your routing. Vendors do not sell those, so nobody mentions them in the demo. A tidy, current knowledge base will do more for bot accuracy than any model upgrade — and it is the one part of the programme you can start this month, at no licence cost.
The 2026 shift
Per-resolution pricing, and the word doing all the work
The biggest pricing change of the last two years is the move from charging for capability to charging for outcomes. Published per-resolution rates now cluster around $0.99 to $2.00 per resolved conversation, with some vendors lower and specialist players declining to publish at all. On its face this is a better deal for buyers: you pay when it works.
The whole model, though, rests on one contested word. What counts as a resolution? The definitions in market range from generous-to-the-vendor to genuinely rigorous:
The gap between the first and last definition can be a factor of two on the same traffic. Three clauses to get into the contract: the precise definition with worked examples, the repeat-contact window that voids a billed resolution, and a dispute mechanism with access to the underlying transcripts. Also check for monthly minimums — at least one major vendor bills a floor regardless of usage — and whether volume discounts exist at all, since several outcome-priced products deliberately have none.
A final piece of context worth keeping in view: Gartner's widely quoted projection is that agentic AI will autonomously resolve around 80% of common customer service issues by 2029, while current enterprise benchmarks sit near 40% deflection. Vendors quote the first number and price against the second. Buy for where the technology is, with contract terms that let you benefit if it gets where it is going.
Before you sign
Twelve questions that change the quote
Take these to every vendor on the shortlist. The answers vary far more than the demos do.
Question 9 deserves special attention. Several vendors' compliance addenda cover the core platform while excluding some AI processing, which means a HIPAA-regulated or PCI-scoped operation can switch on a feature that quietly moves data outside the agreement. Ask for it in writing, feature by feature. Our healthcare contact center guide and FedRAMP guide cover the adjacent ground.
FAQ
Contact center AI questions, answered
How much does contact center AI cost in 2026?
It depends entirely on which of the four pricing models your vendor uses. Agent-facing AI is usually a per-seat add-on of roughly $20 to $60 per agent per month, or bundled into a top tier that runs $160 to $200+ all-in. Customer-facing self-service is metered — per minute, per message, per token, or per resolution — and a per-resolution rate of $0.99 to $2.00 is now common. The per-seat lines are easy to budget; the metered lines are where quotes go wrong, because cost scales with contact volume rather than headcount.
What is the difference between agent assist and a virtual agent?
Agent assist faces inward: it helps a human already on the contact with live transcription, suggested answers, knowledge surfacing, and automatic wrap-up summaries. A virtual agent faces outward: it talks to the customer directly and tries to resolve the request without a human. Agent assist reduces handle time and onboarding time and is low-risk. Virtual agents remove whole contacts and are worth far more when they work — and are far more visible when they fail.
Is per-resolution AI pricing better than per-seat?
It aligns incentives better, but only if the definition of a resolution is tight. Published rates cluster from about $0.99 to $2.00 per resolution, and vendors differ enormously on what counts: some bill any conversation the bot handled, some bill only when the customer did not subsequently reach a human, and at least one now bills only verified resolutions. Get the definition, the escalation exclusions, and the dispute process in the contract, not the demo.
Which CCaaS platform has the best AI?
There is no single winner, and the honest answer is that the packaging differs more than the underlying capability. Dialpad, RingCentral RingCX, and Zoom bundle core AI into the seat price. NICE and Genesys sell the deepest AI suites but meter them heavily. Five9 includes an AI minute allowance per seat and charges usage for virtual agents. Amazon Connect prices everything per unit consumed. Pick on how the meter matches your volume pattern, not on the feature list.
How much of our contact volume can AI realistically deflect?
Plan for a fraction of what the demo suggests. Industry benchmarks put typical enterprise deflection around 40% of queries touched by AI, with a much smaller share fully resolved without human involvement. Gartner's projection that agentic AI will autonomously resolve 80% of common issues is aimed at 2029, not today. Model your business case at a conservative deflection rate, and make sure it still works if the rate lands at half of what the vendor promises.
What is the most common mistake buyers make with AI pricing?
Signing a consumption-based AI line without a volume model. A per-minute or per-message rate looks trivially small in a quote and becomes the largest variable on the bill during a peak season or an outage-driven contact spike. Ask for the cost at your current volume, at double, and at your worst month last year — and get overage rates, caps, and rollover terms in writing.
Do we need AI before we replace our contact center platform?
Usually the reverse. AI features depend on clean routing, structured interaction data, and an accessible knowledge base, and legacy on-premises platforms rarely provide any of the three. Most organizations get more value from fixing the platform and the knowledge layer first, then attaching AI to a system that can feed it. Buying AI to compensate for bad routing produces expensive theater.
Keep reading