replybek.Docs

Get started

Best practices.

What makes an AI agent answer well, and the mistakes that make it answer badly.

An agent is only as good as what you teach it. These habits make the biggest difference.

Give it clear knowledge

  • Write facts plainly. "Delivery to Samarkand costs 45,000 so'm and takes 2 to 4 days" works better than a page that says "see our delivery terms".
  • Keep one fact in one place. If two pages give different return periods, the agent may quote either. Fix the page, then re-sync it.
  • Cover the questions people really ask. Start from your inbox, your phone calls and the unanswered questions list, not from what you think customers ask.
  • Add Q&A answers for the tricky ones. A short Q&A answer is the fastest way to fix a question the agent gets wrong.

Keep instructions short

Instructions shape how the agent answers, not what it knows. A few clear rules work best:

- Keep answers to two or three sentences.
- When someone asks about exchanges, mention that they're free.
- Refund requests over 1,000,000 so'm go to a teammate.

Avoid putting facts like prices in the instructions. They belong in knowledge, where the agent can cite them and you can update them in one place.

Test before you publish

Ask the Playground the ten questions you hear most, in Uzbek, Russian and English. Read the sources it used. If an answer is wrong, fix the knowledge first, then the instructions.

Leave a way to a person

Keep Offer human handoff on, and make sure someone watches the inbox. Turn on email alerts so nobody waits when the dashboard is closed. Customers trust an AI agent more when they know a person is one tap away.

Review every week

Spend fifteen minutes a week on Analytics:

  1. Answer the top unanswered questions.
  2. Read a few conversations that were handed to a person, and ask whether the agent could have answered with better knowledge.
  3. Check that the website sources re-synced and no pages failed.

Connect systems with care

When you add integrations, give the agent only the actions it needs. Make everything that changes data a change action, so the customer always confirms it, and test with sample data before you enable it.