What does “resolved” mean when an AI support vendor says its agent resolved 80% of tickets? It depends on whom you ask. Intercom reports a 76% average Fin resolution rate across its customers, Sierra reports 80% at Airtable, and Decagon reports more than 90% at Substack. Each vendor counts differently, so the figures measure different things, and we found no outside confirmation of any of them.
This is independent analysis from public sources. None of the vendors took part in this article or approved it, and every result is as published by the vendor and unverified by SoftwareVital. We read the vendors’ own definitions where they publish one, then the customer stories that carry the numbers.
Intercom: silence can count as a yes
Intercom sells a helpdesk and an AI agent called Fin. Its help center says a resolution is counted when the customer either confirms the answer was satisfactory (a “confirmed resolution”) or exits the conversation without requesting further assistance (an “assumed resolution”). The same page says a customer who leaves after Fin’s answer without asking for more help “is counted as an Assumed Resolution and will be charged,” although an escalation after Fin detects frustration is not billed.
That is a fair, readable rule. It also means a resolved conversation and a quiet one look the same in the data. In a blog post dated 12 March 2026, Intercom says more than 12,000 teams use Fin and that the average resolution rate across customers “now stands at 76%.” The post also announces a move from pricing resolutions to pricing outcomes, which include a handoff to a human after Fin has done the groundwork. Intercom’s help center lists a resolution and a procedure handoff at the same $0.99.
The Anthropic story reports a 50.8% Fin resolution rate in just over a month, a 96% involvement rate, and 1,700 hours saved in the first month. The page doesn’t say whether 50.8% is a share of all conversations or only of those Fin joined, and it doesn’t say how the hours were counted. Synthesia reports a 55% resolution rate, set against monthly contacts that rose from 40,000 to 316,000, which the page calls a 690% increase. Synthesia’s support lead says the goal is 80%. That is a rare thing to see on a vendor page: a customer’s current number sitting well short of its target.
Zendesk bills by tier, and a language model checks the transcript
Since 18 May 2026, according to Zendesk’s help center, Zendesk AI agents have been billed on automated resolutions grouped into three tiers: assisted escalation, contained resolution, and verified resolution. At the end of a conversation, the page says, a large language model reads the text to confirm the request was satisfactorily resolved. Passing that check makes it a verified resolution. Failing it makes it a contained one. Only verified resolutions draw down the allowance.
The clock matters here. Zendesk says an email conversation ends 72 hours after the first email, a messaging conversation ends 2 hours after the first message by default, and a voice call ends at hangup. So the same customer behavior can land in a different tier depending on channel and settings.
The two stories we read use looser labels. Docebo reports that automated resolutions for incoming chats reached 35 percent within three months and that AI agents now resolve 43 percent of inbound chats and 30 percent of inbound email. The same page says Docebo replaced ten tools with one, so the credit is shared. BritBox reports a 47 percent automation rate, which the page describes as AI agents handling 47 percent of ticket volume, plus a 27 percent decrease in full resolution time. Handling a ticket and resolving it are different verbs. We can’t tell whether either story follows the 2026 tier rules, because neither page shows a date.
Sierra: 80% in one story, 70% in two others
Airtable’s page opens the argument for us. Sierra says Airtable’s AI resolution rate “hit 80%, a significant jump from its previous platform,” and that the agent carries the workload of roughly 95 full-time support agents. The earlier rate isn’t given. Sierra sells AI agents for customer service, and one of its stories sits higher: Ramp reports a 90% case resolution rate through automation.
SoftBank’s Customer Growth Division gives the most useful detail. Its self-resolution rate started around 50% at launch and has risen to 70%, and a footnote explains that Sierra’s monitoring feature uses AI to assess whether a customer’s issue was resolved within the conversation. That is a stated method, not a customer’s click. Chime reports resolution rates rising from 50% to over 70%, in an article first published on 23 October 2025.
On pricing, Sierra’s blog says an unresolved conversation usually carries no charge and that it provides clear, agreed-upon criteria for each outcome upfront. We found no public, formal definition of a resolved conversation on its site. If you’re a buyer, that criteria document is the thing to ask for.
What the five vendors each count
This table lines up the definitions, using only what each vendor’s pages say.
| Vendor | How its pages treat “resolved” | A headline figure and its label |
|---|---|---|
| Intercom | Confirmed by the customer, or assumed when the customer leaves without asking for more | Anthropic: 50.8% resolution rate in just over a month |
| Zendesk | Tiered; a language model reviews the transcript at the end of the conversation | Docebo: 43% automated resolution rate for chat |
| Sierra | No public formal definition found; SoftBank’s page says AI assesses whether the issue was resolved | Airtable: 80% resolution rate |
| Decagon | Share of conversations resolved without escalating to a human | Substack: more than 90% of user inquiries |
| Forethought | Stories headline deflection and containment more often than resolution | MeUndies: 80% containment rate on the chat widget |
Decagon says the quiet part about vendors
The glossary at Decagon, which sells AI agents for support, has a line we’d frame: “Vendors quote whichever of the two flatters the deployment.” It also says most platforms infer resolution “from the absence of a transfer,” and it describes a stricter verified resolution rate that counts anything unconfirmed as unresolved. Decagon’s pricing post then defines its own resolution rate as the share of conversations resolved without needing to escalate to a human, and admits an upset user who leaves can end up counted as a resolution. On that definition, Decagon’s number looks closer to the routing kind its own glossary cautions about, though that is our reading.
Substack is the large figure: Decagon says its agent is resolving more than 90% of user inquiries. Chime shows up again, this time on Decagon’s site, which reports “Consistently seeing 70%+ resolution” in chat and “Consistently nearly 70% resolution” in voice, plus a 60% decrease in support costs. Sierra’s Chime article says resolution went from 50% to over 70%. Both vendors list Chime, both land near 70%, and neither page says when each deployment ran or what share of Chime’s traffic it handled. We would not read that as a head-to-head result.
Curology reports a 65% cost reduction and says chat went from handling 5% of inquiries to 80% of ticket volume. That second figure is a share of tickets entering chat, not a resolution rate.
Forethought counts deflection, and its own pages disagree
At Forethought, the numbers are mostly deflection and containment. YNAB‘s deflection rate was stuck at 25% on its old chatbot, and Forethought says the new agent has “consistently deflected over 70% of tickets” since an October 2024 rollout. The page calls this a 45% improvement. By our estimate, 25% to 70% is 45 percentage points, or about 180% in relative terms, so the label is loose.
MeUndies reports 80% containment on its chat widget and says 11 agents handled Q4 2025 volume that previously required 41. The page names the product “Forethought AI Agents by Zendesk,” and Zendesk is also one of our five vendors. We didn’t research how the two companies are related.
The Achievers story headlines 93% first contact resolution for its tier 1 team. Its body text says Achievers reached a 69% deflection rate, while the results list further down says 44% deflection rate. The page doesn’t explain the gap. It could be different periods or different channels, but a reader can’t tell.
Test tickets to run before you trust a number
Ask each vendor to run these through a trial on your own help content, then check how its dashboard classifies each one:
- A customer asks a plain question, gets an answer, and never replies. Is that resolved, assumed, or unknown?
- A customer types “forget it” mid-conversation and closes the window. Does it count as a resolution, a deflection, or an abandonment?
- A refund that needs a system action plus a human approval. Does the handoff count as a resolution, a partial outcome, or a failure, and what do you pay?
- The same customer returns the next day about the same issue. Is the first conversation reopened, and is the earlier charge reversed?
- A question your help center answers wrongly because the policy changed last month. Does the agent say it doesn’t know, or answer confidently?
- A message at 3 a.m. in a second language. Is it handled, and is it counted the same way?
- Your own audit: pull 100 conversations tagged resolved and read them yourself.
Two quick answers
Which AI customer service agent has the highest resolution rate?
These pages can’t tell you. Vendors pick their own stories, define the term differently, and give different denominators, so Substack’s reported 90% and Synthesia’s reported 55% aren’t comparable. We haven’t named a winner, and we’d distrust any ranking built from this material.
Are these resolution rates verified?
No. Every figure comes from the vendor’s own website. We looked for outside sources that confirmed the numbers and found none.
Where this came from
SoftwareVital is a directory of software and AI tools, and vendors can pay us for advertising. For each vendor we read its own definition of resolution where one exists, plus one to three customer stories that reported a support outcome, favoring recent ones and varying how many we used. Pages were last checked on 7 November 2026. If you work at one of these vendors and can share fuller figures, we’d be glad to update this article.