Zendesk AI metrics to track automated resolutions: an audit scorecard

The scorecard

Zendesk’s current reporting separates Unassisted conversations, Assisted escalations, Contained resolutions, and Verified resolutions. Build a scorecard that preserves these names rather than inventing a single “AI success” bucket.

Zendesk’s reporting change explains that AR% is (Contained + Verified) / all categorized outcomes. It may rise after the definition change, so do not claim that a new percentage alone proves a better agent.

Inspect the records, not just the total

For accounts on resolution allowances, open Admin Center > Account > Usage > Resolutions. Zendesk’s usage guide lets admins filter the dashboard by range, brand, and tier. The contributing list shows ticket ID, tier, channel group, AI agent, brand, and timestamp. Open the ticket or AI-agent conversation from the row when you need to understand an outlier.

Tickets include Resolution type, Resolution tier, and Channel group fields. The event log records when an automated resolution occurred. These are the basis for an audit: a ticket sample tells you whether a sudden tier shift reflects a policy change, a new customer need, a channel change, or a reporting misunderstanding.

A practical review rhythm

1. Fix the date range, brands, and channels before exporting.
2. Record AR%, tier counts, usage, and assisted escalations.
3. Pull a small sample from each material movement, including tickets that surprise you.
4. Write one evidence-backed hypothesis, such as “new billing questions are becoming assisted escalations.”
5. Assign one bounded response: update a source, change routing, or leave the rule alone and inspect again next week.

Do not borrow Zendesk’s tiering, performance, or billing numbers as eesel results. They describe Zendesk’s reporting. The same discipline applies to an eesel deployment: measure the actual configured teammate and its evidence.

Use eesel CLI to prepare a scoped audit, not alter Zendesk data

The eesel CLI lets a person, script, Claude Code, Codex, or Cursor operate the same eesel teammate and workspace as the dashboard. It returns JSON and does not configure Zendesk analytics.

When the scorecard shows more assisted escalations, have a coding agent help investigate a small, approved sample of de-identified cases rather than simply recomputing the total. Ask the eesel teammate to distinguish missing knowledge from an intentional human-only policy. Review connected actions and restrict those outside the analysis before supplying any file or sending a test:

npx @eesel/cli --agent "Support" instructions
npx @eesel/cli --agent "Support" files upload ./approved-handoff-sample.csv
npx @eesel/cli --agent "Support" chat "Review the uploaded handoff sample. Separate missing information from required human review, and flag cases where the record is insufficient."

The coding agent can turn the JSON into a short review list with supporting case references. For example, installation questions may lack a current guide while access disputes correctly require a person. The support owner checks the classifications against Zendesk, approves any source or instruction change for the eesel teammate, and reruns representative cases before accepting it. Preview writes with --dry-run; review activity and the receiving tickets after tests. A sample explains possible causes, not the whole population or an automated-resolution rate. Node.js 18.17+ is required.

Turn a metric review into useful support work

Try eesel alongside Zendesk when you want support analysis to begin with visible inputs, permissions, and evidence. Try eesel.