AI Customer Service Agent Pricing in 2026: What It Actually Costs

Every SaaS category eventually settles on a billing unit everyone understands: a seat, a contact, a gigabyte. AI customer service agents have not done that yet, and the vendors are not pretending otherwise. Four different companies selling functionally similar products — an AI that resolves support conversations without a human — currently use four different units to bill for it, and one of them uses three units at once, depending which pricing page you land on.

This is not a temporary rough patch. Roughly three-quarters of software vendors now charge separately for AI features rather than folding them into the base subscription, and the reason is structural: AI inference has a real, variable cost per use, so a flat per-seat fee either overcharges light users or loses money on heavy ones. Usage-based pricing is how a vendor stays solvent when the thing it sells has a marginal cost. That is a legitimate reason. It also means the number on the pricing page is close to useless until you convert it into your own volume.

The five models running at once

Every AI agent pricing page you will read in 2026 is some combination of these. Knowing which one you are looking at is the first step to comparing two vendors honestly.

• Per-seat. A flat fee per human user — the legacy SaaS model. Rare for the AI layer itself now, though it still prices the human agents who supervise or escalate to it.
• Per-resolution / outcome-based. You pay when the AI successfully closes a conversation without a human. No resolution, no charge — in theory. What counts as "resolved" is defined entirely by the vendor, which matters more than the headline rate.
• Usage / consumption credits. You buy a pool of credits or tokens and each AI action draws down the pool. Closest to how the underlying AI model itself is billed, which is why infrastructure-heavy vendors prefer it.
• Hybrid. A base platform fee plus a usage or outcome component on top. The most common structure in enterprise deals by 2026, because it gives the vendor predictable revenue and the buyer a cost that at least loosely tracks value received.
• Custom enterprise contract. No public price at all. Quote-gated, sales-cycle-only, and — as the examples below show — sometimes a six-figure floor before you process a single conversation.

What vendors actually charge per resolution

These four publish a real rate, which is rarer than it should be in this category. All figures are the vendor's own published pricing, checked in September 2026.

• Quickchat AI: $0.50 per resolution (Enterprise tier, 2,000-resolution monthly minimum). Only conversations the AI closes without human handoff count — a self-serve plan starts separately at $9/month for lower volume.
• Intercom Fin: $0.99 per resolution, billed as an "outcome" — a resolution, a handoff, or a disqualification all count the same, a lead-qualification outcome is $9.99. 50-outcome monthly minimum. Fin for Platforms (standalone, non-Intercom helpdesk) adds a $49/month base fee.
• Zendesk AI: $1.50 per resolution on a committed pack, $2.00 pay-as-you-go. Each seat tier ships with a small bundled allowance (roughly 5 to 15 resolutions per agent per month depending on plan) before the per-resolution rate applies.
• Salesforce Agentforce: $2.00 per conversation under the original 2024 pricing, still offered alongside the newer consumption model covered below.

The fine print that changes the number

"Per resolution" sounds like a single, comparable unit. It is not — each vendor defines "resolved" differently, and the definition moves the real bill more than the headline rate does.

• Zendesk restructured its own definition in May 2026. The resolution model now has three tiers — Assisted Escalation and Contained Resolution are free, and only a "Verified Resolution" draws from your allowance or bills as overage. That sounds like a discount. It also means Zendesk, not you, decides which conversations count toward your bill, and the definition has already changed once.
• Overages auto-bill without warning. Since January 2026, Zendesk resolution overages charge automatically once you exceed the bundled allowance — there is no approval step. Budget against your allowance, not your plan's sticker price.
• Minimums exist to protect the vendor's average, not your bill. Intercom's 50-outcome minimum and Quickchat's 2,000-resolution Enterprise floor both mean a slow month still bills at the assumed volume.
• "AI-resolved" usually excludes anything that touched a human. Quickchat is explicit that only conversations closed without handoff count. If your AI agent is mostly triaging and handing off to humans — which is normal for a new deployment — you may be paying seats for the humans and a low resolution count that looks like the AI barely helped, when really the accounting is just strict.

One vendor, six prices: Salesforce Agentforce

Salesforce is the clearest evidence that nobody has settled on a unit yet. As of 2026 it runs Agentforce across six distinct pricing structures depending on how you deploy it: the original $2-per-conversation rate, a per-user license at $5/month (which itself requires buying consumption credits separately), add-on packages at $125–150, a free "Foundations" allotment bundled into Enterprise-tier orgs, and — the newest — Flex Credits: $500 per 100,000 credits, with each standard agent action metered at 20 credits.

Do the arithmetic and Flex Credits work out to roughly $0.10 per action — but a "resolution" is not one action. If closing a typical case takes an agent three to five actions (a data lookup, a reasoning step, a response, maybe a tool call), that is a rough $0.30–$0.50 per resolution-equivalent under Flex Credits — cheaper than the legacy $2 flat rate, and cheaper than every vendor above except Quickchat. That comparison is ours, not Salesforce's; the company does not publish a per-resolution figure for Flex Credits, because the true cost depends entirely on how many actions your specific workflow needs, and that number is invisible until you have already built and run it.

The vendors that will not tell you

Above a certain deal size, published pricing disappears entirely. Decagon runs on custom enterprise contracts with an estimated $50,000/year platform-fee floor plus a per-conversation charge in the same range as Intercom's — but that per-conversation figure comes from third-party estimates, not Decagon's own site. Reported contract values run from roughly $105,000 to $923,000 a year, with a median near $433,000. Ada goes further: no public rate at all, and no deal reported under 300,000 conversations a year.

Neither of these is doing anything improper — enterprise software has always priced this way above a certain contract size. The practical takeaway is simpler: if a vendor's pricing page has a "Book a demo" button where a number should be, treat the quote you eventually get as the start of a negotiation, not a rate card, and get at least one transparently-priced competitor's numbers in hand before that call.

What 1,000 resolved conversations a month actually costs

Translated into a monthly bill at a volume a genuinely busy small-business support inbox might see, the per-unit gap above stops being abstract:

The $1,500 gap between the cheapest and most expensive option on this list, at one plausible volume, is bigger than most small businesses' entire software budget for every other tool combined. This is not a category where "just pick the market leader" is a safe default — the pricing conversation has to happen before the product demo, not after.

What to ask before you sign

1. Get the vendor's exact definition of "resolved" in writing. Ask specifically whether a handoff, an escalation, or a "the customer stopped replying" conversation counts toward or against your number.
2. Model your real volume, not your target volume. Take your last 90 days of actual support conversations and estimate what fraction a new deployment would realistically close without a human in month one — usually well below the vendor's demo numbers.
3. Ask what happens above the bundled allowance — does it overage automatically, throttle, or require a plan upgrade? Zendesk's answer changed in the last year without much fanfare.
4. Price the human seats you still need alongside it. An AI agent rarely replaces 100% of a support team on day one; the seat cost of the humans who handle escalations is part of the real total, not a separate line item to forget.
5. For any custom-quoted vendor, get a second transparent quote first. Decagon and Ada's sales teams have a number in mind before you do; a same-conversation comparison from a vendor with public pricing changes that conversation.

Questions about AI agent pricing