Best practices

A great agent is mostly a function of great data and clear guardrails. These practices come from what consistently separates agents that resolve 70%+ of tickets from ones that frustrate customers.

Structure your knowledge for retrieval

• One topic per source. Split a sprawling FAQ into focused sources. Retrieval works better on tight, single-topic chunks than on one giant page.
• Lead with the answer. Write support content answer-first; the model surfaces the first relevant lines.
• Remove noise. Navigation, cookie banners, and marketing fluff dilute retrieval. Point the crawler at content pages, not your homepage.
• Keep it current. Re-crawl when policies change. Stale data is the most common cause of wrong answers.

Pin high-stakes answers with Q&A

Anything involving money, eligibility, or legal commitments should be a Q&A pair, not left to paraphrase. Q&A sources take priority over crawled text, so the exact wording you approve is what customers get.

Set clear guardrails in the prompt

Your system prompt should tell the agent what to do when it doesn't know, and what it must never do:

Answer ONLY from the provided knowledge. If something isn't covered,
say you'll connect the customer with the team — never invent policy,
prices, order details, or shipping dates. Keep replies short and friendly.

Choose the right model

Faster, cheaper models handle the bulk of routine questions well. Reserve a larger model for agents that reason over complex policies or long context. See Models & model choice for the trade-offs and per-model credit costs.

Iterate from real conversations

1. 1
   Watch the logs

   Review real conversations weekly. The questions customers ask reveal the gaps in your data.
2. 2
   Fix the data, not just the prompt

   Most wrong answers are a missing or unclear source. Add a Q&A pair and move on.
3. 3
   Track resolution rate

   Use Analytics to measure how often the agent resolves without a human, and where it escalates.

What's next