Zendesk Artificial Intelligence
Zendesk artificial intelligence is a dozen features, several rebrands and a sales page that treats them as one product. They are not one product.
One useful way to cut Zendesk artificial intelligence
Forget the marketing groupings. Sort every AI feature by who sees the output, because that single question predicts the risk, the effort and the payback better than anything else.
Output a customer sees fails in public. Output an agent sees fails privately, in front of somebody qualified to throw it away. Output nobody sees, a field value used for routing, fails invisibly and cheaply. That ordering is also, conveniently, the order you should turn things on.
It's a better lens than the usual one, which sorts features by whether they're "generative" or not. That distinction matters to engineers and almost nobody else. What matters to you is who gets hurt when the thing is wrong, and how quickly you find out.
The portfolio at a glance
Names change. The categories have been stable.
| Piece | Who sees it | What it does |
|---|---|---|
| AI agents | The customer | Handle a conversation end to end and resolve it without a human |
| Copilot | The agent | Drafts replies, summarises threads, suggests next steps |
| Automatic triage | Nobody | Writes intent, language and sentiment into ticket fields |
| Knowledge generation | The writer | Drafts and rewrites help centre articles from solved tickets |
| Quality and workforce tools | The team lead | Scores conversations at scale, forecasts staffing |
The naming history, briefly
Worth knowing, because half the documentation you find online predates the current names.
Answer Bot was the original article-suggestion feature. Most of what it did lives on inside the bot and article recommendation features, under different branding. Advanced AI arrived as a paid add-on bundling triage and richer generation. Copilot is the agent-facing assistant. AI agents is the current name for autonomous resolution, and it covers both the bot you build yourself and a more capable tier above it.
Zendesk has also put an umbrella name over the whole thing more than once, usually framed around resolution rather than around tickets. Umbrella names change faster than features do. Ignore them and look at the feature list.
For the customer-facing side in detail, see Zendesk AI agents. For the agent-facing side, Copilot.
What's included and what costs extra
Here's the honest answer: it depends on your plan, your contract date and which add-ons you hold, and the boundary has moved repeatedly as features graduated from the add-on into base tiers.
Anyone who tells you confidently which tier includes which AI feature is quoting a snapshot. Open Admin Center, look at what your account actually offers, and check the current plan comparison on Zendesk's site. That takes twenty minutes and it beats a stale table every time.
The Advanced AI add-on guide covers the commercial side in more depth.
The order to switch things on
Not the order the sales process suggests.
Teams that reverse this order buy an expensive bot, point it at four thin articles, and conclude AI doesn't work. The bot was fine. The sequence wasn't.
The prerequisite nobody sells you
Every feature in that table reads something. Copilot reads the ticket and your articles. AI agents read your articles. Knowledge generation reads your solved tickets. Triage reads the incoming message.
So the quality ceiling of the whole portfolio is set by two things you already own: your help centre content and the state of your ticket data. Neither is on any price list, and both take longer to fix than the procurement cycle takes to complete.
The practical implication is uncomfortable. If your help centre has forty articles, twelve of which are out of date, buying AI gets you a very fast way to distribute wrong information. Spend the first month on content, tags and forms. It isn't what anybody wants to hear in a kickoff meeting, and it's the single highest-return month of the project.
A cheap test: pick your ten most common questions and try to answer each one using only your published articles. However many you can't answer is your real backlog. Do the same exercise with a colleague who doesn't work in support, because they'll get stuck in the places your customers get stuck and you won't.
The ticket data side is less obvious and matters nearly as much. Inconsistent forms, tags applied by three different conventions and a custom field half the team ignores all make every downstream classification worse. Tidy that before you measure anything, or you'll spend a month blaming a model for your own metadata.
Frequently asked questions
What is Zendesk AI, in one answer?
What is Zendesk AI: an umbrella over a dozen features. The customer-facing Zendesk AI tools are bots and generative answers, and the internal Zendesk AI features are triage, intent detection, summaries and macro suggestions.
Is Zendesk artificial intelligence one product?
No. It's a portfolio spanning customer-facing bots, agent assistance, automatic ticket classification, content generation and quality tooling, with different gating and different pricing.
Which AI features come free with my plan?
That depends on tier, contract date and add-ons, and the line has moved repeatedly. Check Admin Center and the current plan comparison rather than any published list.
What happened to Answer Bot?
The capability continues inside the bot and article recommendation features. The brand name was retired as the AI agent naming took over.
What should a small team turn on first?
Language and intent classification for routing, then agent-side summaries. Both are cheap, both are safe, and both teach you what your queue really looks like.
The problem no model solves
None of these features notice that two tickets in the queue are the same request from the same person.
Start free trial14-day free trial. No credit card required.