Why AI Support-Agent 'Resolution Rate' Claims Aren't Comparable

Every AI support-agent vendor’s homepage leads with a number: a resolution rate, a containment rate, a deflection percentage. Buyers reasonably read these as one comparable stat (“this agent handles X% of tickets”) and shop accordingly. It isn’t one stat. We went through the sourced facts already published on our own support-agents listings for Intercom Fin, Decagon, Cresta, Crescendo, Zendesk AI Agents, Sierra and Ada, and laid the headline claims side by side. The result: five different metric names, at least two vendors that quote a different number for every case study they publish, one vendor that won’t disclose a rate at all, and at least two documented disputes over what counts as a “resolution” in the first place.

Five names for what sounds like one claim

Term	What it typically means	Who uses it
Resolution rate	Share of conversations the agent closed without human help	Intercom Fin, Zendesk
Containment rate	Share of interactions that never escalated to a human	Cresta
Deflection rate	Share of volume diverted away from a human channel entirely	Decagon (in one case study)
Automation rate	Share of volume handled without human intervention	Zendesk
Resolution accuracy	Share of resolved conversations judged correctly resolved	Crescendo

These sound interchangeable, and vendors often use them that way in the same paragraph. They aren’t the same measurement: “contained” doesn’t mean the customer’s problem was actually solved, only that no human touched it; “resolution accuracy” implies a quality judgment resolution rate alone doesn’t make; “deflection” can describe volume that was diverted before an agent ever tried to help. None of the seven listings above define these terms identically, and none publish a shared methodology a buyer could use to convert one vendor’s number into another’s.

The same vendor, three different numbers

Decagon’s own published case studies don’t converge on one figure. They report a different metric for each customer: Duolingo’s is an 80% “deflection rate,” ClassPass’s is a 95% “cost reduction,” and Chime’s is a 70% “chat-and-voice resolution rate.” Three named enterprise customers, three different measurements, none of them the same claim. That isn’t necessarily dishonest, since different customers likely tracked different things, but it means Decagon has no single number a prospective buyer can hold it to; whichever case study gets cited is whichever metric happened to look best for that account.

Zendesk AI Agents does something similar at smaller scale: its own published case studies cite Hello Sugar at 66% automation, TeamSystem at 80%, Babbel at roughly 45%, and Zendesk’s own internal desk above 60%, a spread the case studies themselves put at roughly 45–81% overall. Cresta’s named outcomes follow the same pattern in the other direction: Xanterra’s containment rate is 74%, Propel Holdings’ is 58%, and Snap Finance is reported not as a rate at all but as a “5.5x increase in containment,” a multiplier with no visible starting point. A case study is a best example, not an average; a spread this wide from one vendor’s own selected stories is itself informative.

The headline number and the independently checked number can disagree

Intercom Fin is the clearest example of a vendor-reported number diverging from outside estimates. Intercom’s own materials report a 76% average resolution rate across 12,000+ customers, with many above 85%. Independent analysis of Intercom’s own published case studies, however, found real-world rates commonly landing in the 40–50% range, roughly half the headline figure. Neither number is fabricated: 76% may be a genuine average across Intercom’s full customer base, and 40–50% may be a genuine read of the specific case studies Intercom chose to publish. The gap itself is the finding. A headline average and a case-study-level check of the same vendor’s own evidence don’t agree, which is exactly the kind of thing a vendor’s own pricing page will never flag for you.

When “resolution” is disputed at the billing level

Three of the seven listings surface a sharper problem: what counts as a billable “resolution” is contested, not just imprecise.

• Intercom Fin bills for an “assumed” resolution when a customer simply stops replying to a conversation. A customer going silent is not the same as being helped, but it is billed the same as a real fix.
• Decagon’s resolution-based billing has drawn G2 reviewer complaints specifically about how a “resolution” is counted, on top of a third-party estimate (not an official Decagon figure) that the platform costs roughly $74K–$600K+/year depending on volume.
• Zendesk AI Agents restructured its own billing on 11 May 2026 into three explicit tiers for exactly this reason: “Assisted Escalation” (the agent helped but a human closed it) and “Contained Resolution” (the agent replied and the customer never followed up, but nothing confirmed the issue was actually solved) are both free, and only a “Verified Resolution” (one where automated checking actually confirms the issue was solved) draws from a paid allowance. That three-way split is Zendesk’s own admission that “the agent replied and the customer went quiet” isn’t the same claim as “the agent fixed the problem,” and that the two had been getting billed the same before the change.

Crescendo takes a different approach to the same underlying problem: its per-resolution pricing carries a “Total Outcome Guarantee” that doesn’t charge for conversations that go unresolved or score poorly on customer satisfaction. That shifts the dispute from “was this counted correctly” to “trust our own quality scoring,” which is a different trade, not a solved one. (Crescendo’s own site separately claims a 99.8% resolution accuracy rate; we found no independent figure to check it against.)

Some vendors don’t publish an operational number at all

Sierra prices purely on outcomes: you pay when an agent resolves a customer’s issue, and escalations are free. But Sierra’s own materials don’t disclose what rate of conversations actually get resolved versus escalated. The pricing model itself implies Sierra is confident enough in its resolution volume to bet its revenue on it, but a buyer evaluating Sierra against a competitor that publishes 70–80% has no comparable figure from Sierra to weigh it against.

Ada is the starkest case: across everything in its own published materials, we found no resolution, containment or deflection rate at all, only a case study crediting one customer, IPSY, with a 943% return on investment. ROI and resolution rate answer different questions; a buyer who wants to know “what share of my tickets will this actually close” gets no number from Ada either way.

What to actually ask a vendor

Given all of this, a headline percentage on a vendor’s homepage is close to useless on its own. Four questions get you a comparable answer instead:

1. What exactly counts as “resolved”? Ask specifically whether a customer going silent, or an issue that later reopens, counts toward the number.
2. Is this a fleet-wide average or a named case study? A single customer’s best result and a rate across the full customer base are not interchangeable, and vendors rarely label which one they’re showing you.
3. Who verified it? Self-reported, vendor-audited and independently checked are three different confidence levels, and most published numbers don’t say which one they are.
4. Does it match how you’re billed? If the vendor charges per resolution, the billing definition of “resolution” is the number that actually matters to your invoice, not whatever headline figure is on the marketing page.

For a full breakdown of pricing models, deployment options and which of these seven (plus Botsify, the eighth agent in this category) fits a given support setup, see our AI support-agent comparison guide.