What the sources actually say
In March 2025, OpenAI’s Zendesk case study said early pilots were increasing automation rates toward 80% and that broader metrics would follow later that year. It did not report a general, independently validated 80% result for every account.
Zendesk’s August 26, 2025 article says its platform is resolving up to 80–90% of conversations. It also gives attributed customer examples: Phonero automated 59% of resolutions across channels during a volume spike, while Hello Sugar said 66% of incoming questions were resolved without human involvement. Those are vendor-reported, account-specific outcomes—not a prediction for another account.
| Do not infer | Verify instead |
|---|---|
| 80% is guaranteed for our queue | The outcome mix for our own jobs |
| A closed conversation was good | The ticket, conversation, feedback, and follow-up |
| AI should own every routine request | Which jobs have approved knowledge and a safe handoff |
| A high rate removes human work | Whether exceptions reach the right people with context |
Measure with today’s definition
Zendesk changed reporting on May 18, 2026. Its reporting announcement defines AR% as (Contained + Verified) / (Unassisted + Assisted escalation + Contained + Verified). Both Contained and Verified count in AR%; only Verified consumes the allowance. The contributing-resolution CSV is not enough to calculate the rate because it excludes Unassisted conversations in the denominator.
Use Zendesk's resolution-usage page at Admin Center > Account > Usage > Resolutions to locate contributing records, then inspect the linked ticket and AI-agent conversation. Segment by customer job. A good verification set includes a normal request, ambiguous or missing information, a failed action, a human handoff, and a return contact. Record the expected policy and route before looking at the outcome.
Turn an approved baseline into an eesel proposal
eesel’s CLI operates the same teammate and workspace as its dashboard. Before any write or real chat, review the selected teammate's instructions and action permissions, and get the owner's approval for a small redacted baseline: date range, metric definition, fictionalised case patterns, and known escalation reasons. Do not upload raw ticket exports by default.
npx @eesel/cli status --agent "Zendesk support"
npx @eesel/cli integrations --agent "Zendesk support"
npx @eesel/cli new --name "delivery-pilot-baseline" --agent "Zendesk support"
npx @eesel/cli chat "Using this approved fictional baseline: 100 delivery-status requests, 15 missing order numbers, 8 failed lookups, and a human handoff for account changes, propose one bounded job to test. State supporting evidence, missing facts, and an escalation condition. Do not change tickets, send messages, or alter configuration." --agent "Zendesk support"
Node.js 18.17+ is required and output is JSON. A support lead can run it directly; scripts and coding agents such as Claude Code, Codex, and Cursor can use the same workspace teammate with its existing permissions. Claude Code can turn the JSON proposal into a review note naming the proposed job, excluded requests, missing policy, and cases to test. A human should compare it with the baseline and approved sources. The fictional counts illustrate a brief; they are not measured performance and may overlap. Chat is billed teammate work, not a sandbox. It does not prove Zendesk’s billing classification or establish an 80% result. If the owner later approves an instruction or source change, inspect command help and use --dry-run where supported before the real write; then test the same cases and a nearby exception before comparing real outcomes.
Use the number as a question, not a promise
Use eesel CLI to turn an approved baseline into a concrete pilot proposal. The useful question is which customer work can be safely automated now. Answer it from current tickets, sources, actions, and handoffs—not a copied launch claim.
Frequently asked questions
Does Zendesk guarantee that its AI agent will solve 80% of support issues?
No. The 80% figure is a dated vendor claim or target, not a guarantee for every account.
What did OpenAI say about Zendesk’s 80% automation figure?
It said early pilots were increasing automation rates toward 80%, not that every account achieves it.
What are Zendesk’s 80–90% and customer examples?
They are vendor-reported claims and account-specific outcomes.
How is Zendesk AI-agent AR% calculated?
AR% is (Contained + Verified) divided by (Unassisted + Assisted escalation + Contained + Verified).
How can a team validate a result?
Test current, defined jobs and exceptions, then review conversations, handoffs, feedback, and repeat contacts.
How can eesel CLI help with this evaluation?
A person, script, or coding agent such as Claude Code can inspect the same eesel teammate and workspace as the dashboard, prepare a topic-specific review, and propose configuration changes. Owners approve writes and test outcomes; a CLI proposal does not prove Zendesk-native routing or resolution rates.
