Zendesk AI Agent Knowledge Base: Prepare and Maintain Sources

Prepare a Zendesk AI agent knowledge base by choosing approved sources, making each answer clear and complete, and testing representative customer questions before relying on it. Then assign ownership for gaps, source changes and failed answers. The useful measure is whether the agent can support an answer from current content, not how many articles you connected.

fenbs can hold the editorial decisions, knowledge-gap tasks and lessons from those tests. It does not connect to Zendesk or supply answers to your customer-facing agent. It gives the people and supported assistants maintaining the sources a shared record of what needs to change.

How do you prepare a focused first collection?

1. Choose a limited set of customer questions you want the agent to answer. Identify which questions should go to a person instead.
2. Find the current approved article for each answer. Name a content owner and resolve contradictory versions before connecting the collection.
3. Rewrite incomplete or ambiguous instructions so a reader can follow them without guessing. Add the conditions that change the answer.
4. Select the intended knowledge sources for your agent, using the controls available in your Zendesk setup. Verify that the expected content is actually available.
5. Run coverage tests with ordinary wording, exceptions and missing-information cases. Keep the failed cases as correction work rather than treating publication as completion.

Zendesk’s getting-started guide↗ (opens in a new tab) puts source preparation before agent creation and describes selecting knowledge sources. Use its current instructions for your setup; menus and available options can vary. The preparation process here remains useful whether your sources start in the help center or another supported location.

For example, a fictional appliance company might start with setup and cleaning questions for one product family. Leave warranty exceptions and account-specific requests with the appropriate review process until their source coverage is ready. A narrower scope makes failures easier to understand and repair.

What makes an article suitable for AI answers?

Write the direct answer near the beginning, then explain the conditions and steps. Name the product or policy explicitly. A passage saying “this also applies to it” may be understandable on the full page but unclear when retrieved separately.

Zendesk’s content optimization guidance↗ (opens in a new tab) explains that knowledge is split into smaller passages for retrieval. It recommends focused, self-contained content, familiar wording and clear headings. It also advises removing duplicate or conflicting information and adding written explanations for images.

• State the question the article answers, using the words a customer would naturally use.
• Keep each procedure in order, with prerequisites before the action they affect.
• Put an exception beside the rule it qualifies, using the relevant product or policy name.
• Include essential instructions in text so a diagram is not the only place to find them.

Try a simple editorial check: give someone one section without the surrounding page and ask what they should do. If they cannot identify the product, condition or next action, improve the source before changing the agent’s tone.

How do you check coverage before customers depend on it?

Create a small test set from the question families in scope. Use sanitized or fictional examples, not copies containing customer details. For every question, write down the source that should support the answer and what a correct response must include.

Read each answer against the source. A plausible answer is not enough if it changes a required condition, invents a step or gives instructions for a similar product. Review the explanation and any source reference together. Repeat the same questions after an editorial change so you can compare results.

New to planning the responsibilities of an automated assistant? AI agents for beginners provides that background; keep this test focused on the knowledge your Zendesk agent should use.

How do you turn a failed answer into a useful correction?

Classify the failure before editing. The right article might be missing, the connected copy might be outdated, two sources might disagree, or the answer might misread a correct passage. Those are different problems. Rewriting the final response repeatedly can hide a source problem instead of fixing it.

For each failure, capture a sanitized question, the expected answer, the approved source and the correction needed. Assign a person to review the change. After the correction reaches the knowledge collection, rerun the original question and nearby cases that could be affected.

Record the reason when a person chooses between conflicting policies. That decision should not survive only in a test chat. In fenbs, Decisions and rules, tasks and lessons keep that reasoning with the maintenance work; see how it works.

If you need help separating retrieval from answer generation, the AI agent architecture guide covers those components. Designing a document workflow yourself is a different project, covered in the n8n knowledge base guide.

How do you keep sources current after launch?

Give each source a change trigger and an owner. Product instructions, eligibility rules and contact routes can become wrong before the next planned editorial review. Include a knowledge check in the work that changes those facts, and verify the update in the agent afterward.

Zendesk says its agents use imported or synchronized knowledge rather than searching live internet information for each reply. Its external knowledge explanation↗ (opens in a new tab) describes supported source types. Check the update behavior of the source you actually connect; editing the original page and updating the agent’s available knowledge are separate checkpoints.

Track repeated failures as lessons: what happened, what the team learned and what to do differently. Keep reviewing time in your operating plan; the AI agent costs guide discusses costs beyond the model call. Maintenance is part of keeping the answers useful.

How to do this with fenbs

1. Create a free account and add a project for your knowledge maintenance work.
2. Add tasks for the gaps found in coverage testing. Include source links and sanitized examples of the expected behavior.
3. Record human approvals in Decisions and rules. Save reusable findings in Lessons learned, including what to check after the next source change.
4. Connect a supported assistant through Integrations. Ask it to read the project context before helping with corrections and save the next step in Where we left off afterward.

fenbs is not the right fit for hosting your help center, synchronizing Zendesk sources or operating customer support conversations. Keep those functions in your support system. Use fenbs when the people maintaining that system need the same decisions, lessons and next steps.

The context engineering hub covers related ways to supply reliable information. Start with one question family, one approved source for each answer and one repeatable correction loop.