Zendesk Agent Copilot Adoption Report

Most Zendesk AI reporting measures the bot: how many conversations it handled without a human. Agent Copilot is the other half of the story — AI that sits beside a human agent, suggesting replies and actions that the agent accepts, edits, or throws away.

That distinction matters for how you measure it. Bot performance is judged on containment. Copilot is judged on whether agents choose to use it and whether the tickets they use it on end up faster and no worse. A rollout where suggestions appear on every ticket and get dismissed 80% of the time is not adoption; it is a distraction with a licence fee.

This guide builds the Copilot view in Explore, separates accepted from edited suggestions (the split that tells you what to fix), and lays out how to run a before-and-after comparison that survives scrutiny.

What this report should answer

• What share of auto assist suggestions do agents actually apply?
• How much do they edit before sending, and which procedures need the most editing?
• Which agents and groups have adopted it, and which have not?
• Do Copilot-assisted tickets resolve faster without hurting CSAT?
• Which procedures earn their keep, and which should be retired?

For the metric see auto assist acceptance rate. For the bot side of AI reporting, see Zendesk AI Agent Performance Report and Zendesk Verified vs Contained Resolution Report. Keep this beside the support metrics dashboard.

Start with the prebuilt dashboard, then build what it misses

Before creating anything, open the Zendesk Copilot: Agent productivity dashboard in the Explore dashboards library. It covers three areas and will answer your first round of questions without any work:

• Auto assist — usage and acceptance of auto assist suggestions, including procedure adoption and per-agent usage
• AI suggestions — merge suggestions, quick answers, similar tickets, suggested first replies, suggested macros
• AI agent tools — generative writing tools such as expand, make formal, make friendly, and summarise

Clone it rather than editing it, and check the data refresh interval for your Explore plan before drawing conclusions from a partial day.

Two retention limits shape what analysis is even possible, and they are the single most important planning constraint here:

Ninety days is not enough for a year-over-year story. If you need long-run auto assist trends, export a monthly snapshot from day one — the data will not be there when you go looking for it later. Teams that skip this discover the gap exactly when leadership asks for an annual ROI figure.

Understand the three-way status before you build

Auto assist interactions carry one of three statuses, and getting the arithmetic right depends on understanding them:

• Accepted — the agent used the suggestion as written
• Edited — the agent used it after changing it
• Dismissed — the agent rejected it

Zendesk’s % Acceptance rate counts both accepted and edited as applied, dividing by all interactions shown. That is the right headline definition: an edited suggestion still saved the agent from starting from scratch.

But the accepted-versus-edited split is where the actionable detail lives, and the headline number hides it:

• High accept, low edit — suggestions are production-ready. This is the target state.
• High edit, low accept — suggestions are directionally right but not usable as written. A procedure and knowledge problem, not an adoption problem.
• High dismiss — suggestions are wrong, irrelevant, or badly timed. A targeting problem.

Those three point at three different owners. Reporting only the blended acceptance rate means never knowing which one you have.

How to build the report in Zendesk Explore

1. Acceptance, broken out by status

1. In Explore, open Reports > New report, choose the Copilot auto assist dataset, and click Start report.
2. In Metrics, add Auto assists, Accepted auto assists, Edited auto assists, and Dismissed auto assists.
3. Add % Acceptance rate and % Dismissal rate.
4. In Columns, add Time - Auto assist provided by week.

Report the three statuses as shares of total interactions shown, stacked. That chart tells you the trajectory of trust in a single glance, and it is the one to review weekly during a rollout.

There is a second grain available. Auto assists counts interactions; Suggestions counts the individual suggestions inside them, with % Suggestion acceptance rate as the parallel measure. Interaction level answers “did the agent engage with the help offered”; suggestion level answers “how much of what we offered was useful.” Pick one per chart and label it. Mixing them in one view produces numbers that look wrong and cannot be reconciled.

2. Coverage versus adoption

Two different questions that get conflated constantly.

Coverage — how often is Copilot even offered? Use Auto assist tickets (tickets where auto assist was provided at least once) against total Tickets.

Adoption — when offered, is it used? Use Applied auto assist tickets against Auto assist tickets, plus Agents used auto assist against total Agents.

Also useful: Tickets with all auto assists accepted, the count of solved tickets where every interaction was accepted. Those are your clean AI-assisted resolutions, and the best population for a like-for-like comparison against unassisted work.

Low coverage with high adoption is a good problem — expand the procedures. High coverage with low adoption means suggestions are firing where they do not belong.

3. Procedure performance

This is the most operationally useful view in the report. Auto assist follows procedures you author, and Auto assist procedure lets you rank them.

Put Auto assist procedure in Rows with Auto assists, % Acceptance rate, and % Dismissal rate as metrics.

Read it as a maintenance backlog:

• High volume, high acceptance — working. Look for adjacent workflows to cover next.
• High volume, high dismissal — actively wasting agent attention. Fix or retire it first.
• High edit rate — the procedure is close. Usually the wording or a missing condition, and the cheapest fix in the report.
• Low volume — either the workflow is rare or the trigger conditions are too narrow to fire.

Zendesk recommends adding procedure-specific tags during auto assist setup, and better still a procedure custom field because field values are human-readable where tags are not. Do this before you scale beyond a handful of procedures; retrofitting it across a mature setup is tedious.

4. Adoption by agent and group

Add Agent name with Auto assists, % Acceptance rate, and Agents used auto assist, then split by Ticket group.

Interpret carefully. Low acceptance from an experienced agent on complex tickets is often correct judgement — they can tell when a suggestion is wrong. Low acceptance across a whole group is more likely enablement or trust. And uniformly high acceptance from a new agent deserves a QA check, because accepting everything is not the same as reviewing everything.

Use this view for enablement targeting, never for performance ranking. The moment acceptance rate becomes a metric agents are measured on, it stops measuring suggestion quality and starts measuring compliance.

5. The impact comparison

Adoption alone does not justify the investment. The Copilot auto assist dataset conveniently carries ticket outcome metrics alongside the AI metrics — First reply time, Full resolution time, % Satisfaction score, Reopens, Agent replies, Assignee stations, and Group station — so you can build the comparison in one report.

Compare applied-auto-assist tickets against non-auto-assist tickets on:

Reopens and CSAT are the guardrails, and they are the two people skip. A resolution time improvement paid for with a reopen increase is not an improvement; it is deferred work plus a frustrated customer. Same logic as Why fast first reply time does not always mean fast support.

For work outside the auto assist dataset, the recipe Zendesk documents is a pair of calculated metrics on the standard Tickets dataset using the tag Copilot applies:

Auto assist tickets:
IF (INCLUDES_ANY([Ticket tags], "agent_copilot_enabled")) THEN [Ticket ID] ENDIF

Non-auto assist tickets:
IF (NOT INCLUDES_ANY([Ticket tags], "agent_copilot_enabled")) THEN [Ticket ID] ENDIF

That lets you carry the comparison into any Tickets-dataset report you already have, including cloning the Efficiency tab of the prebuilt Support or Messaging dashboard and filtering by procedure.

The selection bias problem — and how to handle it

The comparison above has a real methodological flaw that will be pointed out the moment you present it, so handle it before someone else does.

Auto assist does not fire randomly. It fires where procedures match, which is disproportionately on routine, well-understood requests — exactly the tickets that were already fast. Comparing assisted to unassisted tickets therefore partly compares easy work to hard work, and it overstates the benefit.

Three ways to reduce the distortion, in order of effort:

1. Compare within a narrow slice. Same form, same tag, same priority, same group. Like-for-like beats scale here.
2. Compare before and after within the same procedure’s scope. Take the tickets a procedure now covers, and measure the same population before the procedure existed. This is the cleanest available approach and the reason to snapshot early.
3. Stage the rollout by group. Enable for one group and hold another as a control for a few weeks. Real experimental design, and worth it before a large licence commitment.

Say which method you used when you report the number. A smaller, defensible improvement is far more useful than a large one that collapses under the first question.

How to interpret the patterns

High coverage, low acceptance

Suggestions are firing where they should not. Tighten procedure conditions rather than pushing adoption harder — agents dismissing bad suggestions are behaving correctly, and training them out of it makes quality worse.

Acceptance drops after a strong start

Novelty wearing off, or the easy procedures already covered and newer ones performing worse. Split by procedure and by procedure launch date before concluding that enthusiasm faded.

Edit rate much higher than accept rate

The best problem in the report, because it is cheap to fix. Suggestions are useful but not sendable — usually tone, length, or a missing detail. Compare edited suggestions against your macros; the gap is often literally a macro that was never reflected in the procedure.

Adoption concentrated in a few agents

Normal early, a problem if it persists. Find out what the adopters do differently and whether the non-adopters are seeing suggestions at all — check Auto assists per agent before assuming a behaviour problem.

Resolution time improves but reopens rise

Stop and investigate before expanding. Suggested replies may be closing tickets with answers that do not hold. Cross-check Zendesk Reopened Tickets Report and Zendesk Satisfaction Reason Report — a rise in “the issue was not resolved” is the confirming signal.

Suggestions edited by an app rather than an agent

Suggestion edited by distinguishes Agent, App, and Agent and app. App edits mean an integration is rewriting suggestions, which changes what your acceptance rate means. Worth checking before attributing the number to agent behaviour.

Common mistakes

• Reporting only blended acceptance. The accept/edit/dismiss split is where the fix lives.
• Ignoring the 90-day retention limit. Snapshot monthly from day one or lose the trend permanently.
• Mixing interaction-level and suggestion-level metrics. Two grains, two denominators, irreconcilable in one chart.
• Treating coverage as adoption. Showing suggestions is not using them.
• Comparing assisted to unassisted tickets without controls. Auto assist fires on easier work; unadjusted comparisons overstate the gain.
• Measuring agents on acceptance rate. It converts a quality signal into a compliance signal immediately.
• Skipping reopens and CSAT. Speed without a quality guardrail is not efficiency.
• Scaling procedures before tagging them. Without procedure tags or a custom field you cannot tell which ones work.

What to do when acceptance is low

1. Split accepted, edited, and dismissed — decide which problem you actually have.
2. Rank procedures by dismissal rate and fix or retire the worst.
3. Check whether low acceptance is concentrated in a group or spread evenly.
4. Confirm suggestions are firing on appropriate tickets by cross-checking form, tag, and priority.
5. Read a sample of edited suggestions against your macros to find the wording gap.
6. Verify no integration is rewriting suggestions before blaming agents.
7. Re-measure by procedure after each change, not in aggregate.
8. Keep exporting monthly snapshots so the 90-day window never truncates your evidence.

Where this report fits in your dashboard

• Zendesk AI Agent Performance Report — the bot-facing half of AI reporting
• Zendesk Verified vs Contained Resolution Report — what AI actually resolved
• Zendesk Suggested Article Acceptance Rate Report — the same acceptance logic for knowledge
• Zendesk Macro Usage Analytics — the pre-AI baseline Copilot is competing with
• Zendesk Average Handle Time — where efficiency gains should land
• support metrics dashboard — the hub

FAQ

Which dataset do I use for Agent Copilot reporting? Copilot auto assist for auto assist interactions, Copilot suggestions for similar tickets, merges, quick answers and suggested macros, and Generative AI agent tools for writing enhancements.

Does acceptance rate include edited suggestions? Yes. Zendesk’s % Acceptance rate counts accepted and edited as applied, dividing by all interactions shown. Report the split separately as well.

How far back does auto assist data go? The Copilot auto assist dataset retains 90 days. Export monthly snapshots if you need longer trends.

How do I compare Copilot tickets to non-Copilot tickets outside the Copilot datasets? Use calculated metrics on the Tickets dataset with the agent_copilot_enabled tag to split the two populations.

Can I see which procedures perform best? Yes, via the Auto assist procedure attribute. Add procedure-specific tags or a custom field during setup to make the reporting readable.

Should I set an acceptance rate target for agents? No. Targeting acceptance converts it from a measure of suggestion quality into a measure of compliance, and you lose the signal that tells you which procedures to fix.

See whether AI help is actually speeding up your Zendesk queue - start free