Zendesk AI chatbot setup

Design the ticket before the conversation

Teams often start by writing the bot's greeting. We start with a different question: when the bot fails, what should the agent see? A chat that arrives as an empty ticket titled 'Conversation with Web User' wastes the first two minutes of an agent's time.

This table becomes your build specification. It also exposes problems early, such as an order-number field that exists in three variants across brands.

What we configure

We have no published case study on Zendesk, so this is the usual scope and not a client story.

1. Intent list from real ticketsWe cluster 500 to 1,000 recent tickets by what the customer wanted, and pick the ten or twelve intents that make up most of the volume.
2. Content in GuideFor each intent, one article that states the rule and the next step. Old, overlapping articles are merged or unpublished.
3. Automated answers and flowsGenerative answers for open questions, and fixed flows for steps where precision matters, like a return request.
4. Actions and dataLookups to your store or system through an API or the Zendesk integration points, read-only to begin with.
5. HandoffBusiness hours, queue routing, the context note, and what the customer is told while they wait.
6. Testing and releaseReplay past tickets, fix content, then enable one channel and one brand first.

For the general approach, see AI customer service automation. If you are weighing a different helpdesk, our Intercom Fin page covers that route.

Gotchas

• Billing by automated resolution. Zendesk's AI agent pricing has been based on resolutions, and plan structures change. Check the current terms, and define what counts as resolved in your reporting.
• Guide quality. If your articles disagree with each other, so will the bot. Content cleanup is the biggest part of most projects.
• Channels differ. Messaging, email and help center search behave differently. Decide where the bot lives and test each channel separately.
• Multi-brand setups. Each brand may need its own articles, tone and routing. Plan for that from the first week.
• Triggers colliding. New triggers can fire alongside existing ones and reassign tickets unexpectedly. We list your current triggers and put the new rules in order.
• Privacy. Chats can contain personal data. Check retention settings and update your privacy notice to say an AI agent answers first.
• Customer expectations. Offer a clear way to reach a person at any point.

How we measure it

Before launch we record four baselines from your last quarter: first reply time, tickets per day by intent, share solved in the first touch, and customer satisfaction on chat tickets. After launch we report the same four, plus the bot's own numbers.

• Automated resolution rate per intent, not one blended figure.
• Handoff rate and the reasons for handoff, which show which articles to fix.
• Reopened tickets after a bot resolution, a sign the answer was wrong.
• Time to first agent reply on escalated chats, which should fall if the context note works.

A weekly review of twenty random conversations catches problems the numbers miss.

Timeline

Multi-brand or multi-language accounts take longer. The fixed quote after our call shows the phases that apply to you.