How to set up a WhatsApp AI agent

Before you start

• Your policies and FAQs as documents
• A clear scope for the agent
• Escalation rules and a person behind them

The steps

1. 1

   Gather the material it will answer from

   Returns policy, price list, prospectus, shipping table, opening hours. Upload PDFs, DOCX or text, or point it at your own site. It answers from this, not from general knowledge.

2. 2

   Give it one job and a voice

   A narrow agent is a better agent. "Answers questions about our courses" beats "helps customers" every time, and it makes the failure cases obvious.

3. 3

   Write down what it must never attempt

   Clinical advice, legal positions, exceptions to your own policy, anything involving money changing hands on a promise. Scope these out deliberately rather than hoping.

4. 4

   Attach the skills it needs

   Order lookup, calendar availability, a custom API action against your own endpoint. An agent that can only talk will send every real request to a person anyway.

5. 5

   Test it in the playground

   Run real questions and read what it retrieved as well as what it said. If it answered well from the wrong document, that is a problem you will only see here.

6. 6

   Go live in test mode first

   Test mode answers only your own number and your team's. Leave it there for a day of real questions before switching it on for customers.

Where this usually goes wrong

• Uploading everything you have

  A knowledge base full of outdated price lists makes the agent confidently wrong. Upload what is current, and re-upload when it changes.

• No escalation rule

  Hand-off is the design, not a failure state. The measure of a good agent is what it does with what it should not answer.

• Skipping test mode because it looked fine in the playground

  The playground is your questions. Test mode is theirs.