Zendesk AI Chat

Three quite different capabilities get sold as Zendesk AI chat. The decision that matters isn't which model, it's which conversations you let it near.

Three things called Zendesk AI chat

Separate them before you evaluate anything, because they fail differently and they cost differently.

Scripted bot flows. A branching conversation you draw yourself. Deterministic, predictable, and boring in the good way. Covered in the chat bot guide.
Generative answers from your content. The model reads your help centre and answers in prose. It's impressive when your documentation is good and embarrassing when it isn't, which makes it a documentation project wearing an AI badge.
Agent assist. No customer sees it. It drafts replies, summarises long threads and suggests articles while a human stays in control. The least discussed of the three and often the best return, because the downside of a bad suggestion is an agent ignoring it.

Which of these your account includes depends on tier and add-ons, and that has changed more than once. Check the current Zendesk docs against your plan before you design anything around a specific capability. More on the wider family in AI agents.

What is genuinely safe to automate

Anything where the answer is the same every time and being wrong is cheap.

Where is my order. A lookup, not a judgement. The highest volume, lowest risk automation there is, assuming you can reach the data.
Password and access basics. Predictable, tedious, and nobody on your team enjoys handling them.
Opening hours, policy facts, shipping windows. Documented answers with exactly one correct version.
Collecting context before a human arrives. Product, version, order number, account. The bot doesn't resolve anything and the agent starts with everything.

That last one is the most undervalued use of AI on chat. A conversation arriving pre-qualified saves an agent two minutes and saves the customer being asked for things they already typed, which is the complaint people actually make about support.

What to keep it away from

A short and non-negotiable list.

Anything involving money. Refunds, credits, billing disputes, cancellations. A model that confidently offers a refund it shouldn't have has cost you the money and the trust in one move.

Anything involving an apology. When somebody is genuinely upset, a generated expression of regret reads as contempt. Route it to a person immediately, and build an escalation path that reacts to tone rather than waiting for the customer to ask three times.

Anything where the policy is ambiguous or recently changed. Generative answers are only as current as the content they read, and your help centre is always slightly behind your actual policy. That gap is where the confidently wrong answers live.

And anything legal, medical or safety related, which should not need saying and does.

The handoff decides everything

Customers don't judge your bot on its best answer. They judge it on what happened when it couldn't help.

Three rules. The exit to a human has to be visible at every single step, not buried behind three failed attempts. The transcript must travel into the ticket, so nobody's asked for their order number twice. And if agents are offline, say so plainly and give a time, rather than transferring into silence.

A bot with no obvious exit is the fastest way to turn one conversation into a complaint on another channel, and that channel is usually a public one with a star rating attached.

Measuring it without fooling yourself

The vendor metric is containment: conversations the bot handled without a human. It's a genuinely useful number and it's also the easiest one in support to flatter.

A conversation where the customer gave up counts as contained. So does one where they got a wrong answer and emailed you an hour later. Neither of those is a success and both look identical in the report.

So pair containment with three things. Satisfaction on bot-handled conversations specifically, not blended with agent conversations. The rate at which a bot conversation is followed by a human contact from the same person within 48 hours. And a weekly read of ten actual transcripts, chosen at random, by somebody who cares.

That third one isn't a metric and it'll teach you more than the other two combined.

A sensible order to roll it out

Agent assist first, because it improves replies without exposing anything to a customer. Then scripted flows on your top three or four intents, measured against the transcripts you already have.

Generative answers last, and only after somebody has audited the help centre articles the model will be reading. Buying the feature before fixing the content is the single most common way these projects disappoint, and the disappointment gets blamed on the AI rather than on the documentation.

FAQ

Frequently asked questions

Is a Zendesk AI chatbot the same as an AI agent?

Not quite. A Zendesk AI chatbot follows a flow you drew; Zendesk AI agent chat resolves without one. Zendesk chat automation covers both, plus the routing rules underneath.

Does Zendesk AI chat work on live chat or messaging?

Bots and AI features live on the messaging channel, which means the modern web widget, the mobile SDKs and social channels rather than the older session-based chat product.

Should the bot answer first or offer a human first?

Bot first is fine as long as the exit to a human is visible from the opening message. Hiding it until three failures is how you generate complaints.

Can AI answer from our help centre articles?

Generative answers read your published content, so the quality of the answers is the quality of your documentation. Fix the articles before you buy the feature.

What containment rate is realistic?

It depends entirely on your mix, and any specific figure quoted at you is marketing. Measure yours, and pair it with satisfaction and follow-up contact rate before you believe it.

When the bot does not land

A customer who gives up on a bot writes in again through another channel. Ticket Merger matches the pair and merges them.

Start free trial

14-day free trial. No credit card required.