Zendesk Auto Assist

Zendesk Auto Assist is assistance aimed at the agent rather than the customer. It reads the ticket, proposes the next move, and waits to be told yes.

Start free trial All guides

What Zendesk Auto Assist is, before the product name

Zendesk ships an agent-side assistance layer inside the workspace. Auto assist is the part that goes beyond drafting text: it follows procedures you have written, works out which one applies to the ticket in front of the agent, and proposes the next step in that procedure.

Names, tiers and capabilities in this area move quickly, and Zendesk has renamed AI features more than once. Treat this page as a description of how the category behaves and confirm the current feature name, plan requirement and add-on cost in Zendesk's own documentation before you buy anything.

What has stayed stable is the shape. Suggestion, then human, then action.

Procedures are the whole game

A procedure is your process written in plain language: how to handle a refund request, what to check before issuing a replacement, when to escalate a suspected outage. The assistant reads those and follows them.

Which means the output quality is a direct function of how well you write them. Vague procedures produce vague suggestions.

Write for a competent new hire. Not for a machine, and not for someone who already knows the answer. Name the checks, the order, and the decision point.

One procedure per outcome. A single document covering refunds, exchanges and goodwill credit will get applied to the wrong ticket sooner or later.

Include the refusal cases. What shouldn't happen matters as much as what should. Say when the answer is no.

Keep them where you maintain them. A procedure that drifts out of date silently is worse than no procedure, because now it's being followed.

Version them. When a suggestion goes wrong you want to know which version of the document produced it. Blaming the model for faithfully following something you changed on Tuesday wastes everybody's afternoon.

What it drafts

Broadly three things, and they carry different levels of risk.

Replies, assembled from ticket context, your help centre and the relevant procedure. Low risk, because an agent reads every word before it goes out.

Summaries of long threads, which is the feature agents notice first when they inherit a forty-message ticket at 4pm on a Friday.

Proposed actions: look up an order, apply a macro, set a field, ask a clarifying question. This is where the interesting boundary sits, because an action changes something.

Worth being blunt about what it isn't. It isn't a chatbot talking to your customer, and it isn't a rules engine with deterministic outcomes. It's a suggestion layer sitting between your written process and the person doing the work, which is why it fails softly. The bad outcome is a wasted second, not a wrong answer already sent.

Where the human stays in the loop

The useful mental model is a spectrum from suggest to act. Drafting sits at the suggest end and can be adopted broadly with very little governance. Anything that touches money, entitlement, account state or a third-party system belongs at the other end, behind an explicit approval.

Set that line deliberately rather than accepting a default. Ask two questions of each proposed action: if this fires wrongly, does the customer see it, and can we undo it? A wrong draft is invisible and free. A wrong refund is neither.

Agents also need a way to decline that is faster than accepting. If dismissing a bad suggestion costs three clicks, people will accept things they shouldn't, and you'll have built a system that manufactures agreement.

Measuring it without fooling yourself

Suggestion acceptance rate is a vanity metric. It goes up when suggestions get blander and when agents get tired.

Measure handle time on the ticket types where assistance is actually used, and put CSAT next to it on the same chart. If handle time falls and satisfaction holds, it works. If both fall, agents are shipping drafts they should have edited, which is a coaching problem rather than a tooling one.

Watch reopen rate too. A fast wrong answer looks brilliant in a throughput report and shows up a day later as the same customer, writing again.

Give it a fair trial length while you're at it. The first fortnight measures novelty rather than value: some agents accept everything, others refuse on principle, and neither behaviour survives a month of ordinary work. Look at week five.

FAQ

Frequently asked questions

Is auto assist part of Copilot?

Yes. Zendesk auto assist Copilot features are the agent-facing suggestions, and Zendesk agent Copilot auto assist reads the ticket against your Zendesk procedures to propose the next step. Zendesk AI suggestions elsewhere in the product are simpler than that.

What is Zendesk auto assist?

The part of Zendesk agent-side AI that follows procedures you have written, proposes the next step on a ticket and drafts the reply, with a human approving before anything is sent or actioned.

Does auto assist reply to customers on its own?

The design point is that an agent approves. Fully autonomous customer replies are what AI agents do, and that's a separate feature with separate scoping. See AI agents.

Which plan includes it?

AI capability in Zendesk is tier-dependent and some of it carries additional cost. Naming and packaging change often, so get an all-in quote for your seat count rather than relying on a comparison table you read last year.

What do we need in place before turning it on?

Written procedures and a help centre worth reading. Assistance drawn from thin documentation produces confident nonsense, and no amount of model quality fixes an empty knowledge base.