Zendesk Generative AI
Zendesk generative AI writes prose on your behalf, which is useful and slightly dangerous. The difference is entirely in who reads it before the customer does.
The three places Zendesk generative AI shows up
Generative features are scattered across the product rather than sitting in one menu, and they carry very different risk.
Feature names in all three areas have changed repeatedly. Where a name in an old blog post doesn't match what you see in your account, trust your account.
Agent-side generation, the safe bet
This is where generative AI earns its keep fastest, and the reason is structural rather than technical. A bad suggestion in front of an agent costs a second and a click. A bad reply in front of a customer costs a complaint.
The features agents notice most are expansion and summarisation. Typing "refund approved, 3-5 days, apologise for delay" and getting a full paragraph back is a genuine saving repeated fifty times a day. Summarising a forty-comment escalation before you pick it up saves more.
Tone adjustment is the one to watch. It works, and it homogenises. Run every reply through it and your support voice becomes indistinguishable from every other company using the same feature, which is a slow cost nobody notices until a customer mentions it.
The gains are also unevenly spread, which surprises people. On a two-line reply that a macro already handled, generation saves nothing and occasionally costs a few seconds. On a long, technical, multi-part answer it saves minutes. Point it at the second kind and measure there, rather than averaging across a queue where most tickets were never the problem.
Customer-facing generation
A bot that answers from your articles is a different proposition from a bot that follows a flow you drew. It can handle phrasing you never anticipated. It can also assemble two half-relevant articles into a confident answer that is wrong in a new and interesting way.
Two things make the difference. Ground it strictly in your own content, and instruct it to say it doesn't know rather than improvise. Then measure how often it actually says so. A system that never admits ignorance isn't well grounded, it's guessing fluently.
Keep money, policy and anything with a commitment in it out of scope entirely. Those are the answers customers quote back at you three weeks later, and they're usually right to.
One practical point that gets missed: generative answers are only as current as your articles. Publish a policy change on Monday and forget to update the help centre, and the bot keeps confidently quoting the old policy until somebody notices. Put help centre updates into your release checklist, next to the changelog.
The review step people skip
Draft quality is the trap. The output is good enough that agents start skimming, and wrong often enough that skimming eventually costs you.
So make review structural rather than a matter of discipline. Sample generated replies in your QA process specifically, tagged so you can find them. Watch reopen rates on tickets where a draft was used against tickets where it wasn't. Never let a generated reply auto-send on anything involving a refund, a policy statement or an apology with a commitment attached.
A generated reply is a draft written by somebody who has never met your customer and has not read your refund policy.
That framing keeps people honest better than any training deck.
What it cannot know
The model reads the ticket and, where grounding is set up, your articles. That is all.
It cannot see your billing system, your shipping provider, the note somebody left in a spreadsheet or the exception your team agreed verbally last month. Any sentence about this specific customer that you did not feed it is invention, however well written.
The practical version of this rule: generative features are excellent at how you say something and unreliable at what's true. Keep the facts coming from your systems and let the model handle the wording.
There's a data question sitting underneath all of this too. Ticket text accumulates whatever customers pasted in, which over a year includes card fragments, passwords, screenshots containing other people, and occasionally medical detail. Any generative feature reads that text. Redact aggressively, check what your contract says about training, and get whoever owns privacy to look at it once rather than never.
Frequently asked questions
What can generative AI change about a reply?
Three things: Zendesk generative replies draft the answer, Zendesk tone adjustment rewrites what an agent typed, and Zendesk generative answers write directly to the customer in search. Only the last one skips human review.
Which Zendesk plans include generative AI?
Some generative features have sat in the paid AI add-on and others have moved into base tiers over time. The boundary keeps shifting, so check your account and Zendesk's current plan documentation.
Can generated replies send automatically?
Technically you can build workflows that get close. You shouldn't, for anything involving money, policy or a promise. Suggest-only is the correct permanent setting for most of a queue.
Does it use our tickets for training?
Read the current data processing terms rather than trusting a summary. This is a contractual question, and it's one your legal team will ask.
Will customers notice AI-written replies?
They notice sameness more than they notice AI. Tone adjustment applied to everything flattens your voice, and that shows up in feedback long before anyone says the word robot.
Fast replies, twice
Generative drafts make each answer quicker. They do nothing about answering the same question more than once in two tickets.
Start free trial14-day free trial. No credit card required.