Zendesk AI agent review: what to evaluate in 2026

What has changed

Zendesk's migration guidance says the new experience replaces Essential and legacy bot functionality. Zendesk's deprecation notice places the maintenance transition on August 31, 2026 and removal on December 10, 2026. New AI agents are single-channel, so one agent cannot serve both messaging and email.

The current product can use generative procedures for flexible, policy-guided work. That is valuable only if the procedure, knowledge sources, actions, and human route are all deliberate.

What it is good at, and where it needs a boundary

Zendesk's new experience combines knowledge-backed replies with generative procedures, dialogues, and authorized actions. That can make a repeatable job such as checking an order status or explaining a return policy much less scripted than an old bot flow.

It does not remove the need to define the job. A procedure can follow your policy, but it cannot make an outdated policy safe. An action can save time, but it should not have authority to make an irreversible change merely because a customer phrased a request convincingly. Review the source owner, action permission, identity check, and human escalation for each job separately.

The channel change matters too. Zendesk says each new AI agent works on one channel. A team that previously used one Essential agent for messaging and email needs to plan, build, and test separate agents. Do not assume a successful messaging result proves the email experience.

Ask buying questions before a rollout

Zendesk is changing both packaging and product. Its migration article says the Essential/Advanced distinction was removed during the 2026 rollout and that the newer capabilities are being expanded across Suite and Support plans. Entitlement and usage still depend on the account, so confirm the plan shown in Admin Center and the commercial terms Zendesk gives your account rather than budgeting from an old comparison table.

Ask the account team or procurement owner for four things in writing: what AI functionality and usage are included, what counts as an automated resolution or other billable use, which channels and actions are available, and what migration or implementation services cost. Then estimate the work your team owns: cleaning source content, setting permissions, reviewing failures, and maintaining separate channel agents. A price per successful conversation is not the whole cost if the queue still needs heavy human review.

Run a controlled evidence cycle

Start with one job, one channel, and a small set of current sources. Make a test set from recent tickets after removing personal data. Include normal questions, policy edge cases, missing information, a request for a person, and an action the agent must decline. For every case, record the expected answer or route before looking at the agent's reply.

Zendesk QA is useful here because it supports scorecards, reviews, dashboards, and BotQA metrics for AI-agent conversations. Use a scorecard that checks correctness, policy compliance, tone, escalation, and whether the customer got a usable next step. Sample successful resolutions as well as escalations; otherwise a high containment number can hide confident but wrong answers.

A review framework that produces evidence

Zendesk's AI-agent analysis tools include reporting, conversation logs, labels, and error logs. Zendesk QA can evaluate messaging AI-agent conversations with configured scorecards, manual or automatic reviews, a Reviews dashboard, and BotQA escalation metrics. Those are stronger review inputs than a vendor-wide accuracy number.

Use the same eesel teammate from the terminal

eesel's CLI works with the same workspace teammate as the dashboard, so a support lead can ask Claude Code, Codex, Cursor, or a script for a reviewable evidence pack. Get the workspace owner's approval and inspect connected permissions before testing.

npx @eesel/cli integrations --agent zendesk-review
npx @eesel/cli approvals --agent zendesk-review
npx @eesel/cli new --name "refund-policy-review" --agent zendesk-review
npx @eesel/cli chat "Can I return an opened item after 45 days?" --agent zendesk-review
npx @eesel/cli activity --agent zendesk-review

With Node.js 18.17 or later, the CLI returns JSON. Ask the coding agent to compare the refund reply with the approved policy: does it identify the purchase date and opened-item exception, ask for missing details, and avoid promising eligibility? Save the reply, cited source, and activity evidence with a human verdict. If a rule is missing, have the coding agent propose a narrower instruction, get the policy owner's approval, and repeat the case in a fresh session. That produces a reviewable change, not just a transcript.

These are checks of the eesel teammate, not a native Zendesk AI-agent score. Apply the same human scorecard to both products if you are evaluating them side by side, and check their actual ticket outcomes separately. The CLI is not an isolated simulation; approvals list held actions, not every possible action.

Try a Zendesk AI teammate with a defined job

eesel is an AI helpdesk teammate for the Zendesk queue you already use. Give it one job, approved sources, and a clear escalation route. Then inspect its configuration and activity before expanding the role.

The CLI is another way for a person, script, Claude Code, Codex, or Cursor to operate that same teammate. It is useful when the review needs a repeatable record, while the dashboard remains useful for the same workspace configuration and activity.

Try eesel