5. Go live

Don't flip the agent on for every customer at once. Release it to a small, controlled set of numbers first, watch how it behaves on real conversations, and fix what you learn. That's what the allowlist is for.

How the allowlist works

The agent's audience is controlled by the ai_audience setting:

• ALLOWLISTED_ONLY — the agent responds only to consumers whose numbers are on the allowlist. Everyone else is handled as if the agent weren't there.

• EVERYONE — the agent responds to all consumers (that's full rollout).

Turn on a controlled launch

1

Add your test numbers to the allowlist

curl -s -X POST https://waba-v2.360dialog.io/agent_config/allowlist \
  -H "D360-API-KEY: <YOUR_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{ "phone_number": "[phone]" }'

Add to allowlist

2

Enable the agent and set the audience to allowlist-only

curl -s -X PUT https://waba-v2.360dialog.io/agent_config/settings \
  -H "D360-API-KEY: <YOUR_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "rollout": { "enabled": true },
    "ai_audience": "ALLOWLISTED_ONLY"
  }'

Update AI agent settings

Now only the numbers you added will talk to the agent.

Iterate

Treat the controlled launch as a feedback loop:

1. Message the agent from allowlisted numbers on real WhatsApp.

2. Watch the replies — on the phone, or by logging the standby webhook events. Is it grounded, on-brand, and handing off when it should?

3. Refine skills, FAQs, and business info based on what you see.

4. Expand the allowlist gradually — add colleagues, then friendly customers.

5. Repeat until you're confident.

Reminder: agent_test won't return replies under ALLOWLISTED_ONLY. During the allowlist phase, validate by messaging from allowlisted numbers on WhatsApp rather than the test endpoint. See Test your agent for why.

You can monitor the allowlist phase over webhooks. While the agent holds a chat, consumer messages and agent replies arrive on the standby field of your number's webhook — so you can log and review real conversations instead of reading them off the phone. See Webhooks.

Have your human-handoff path ready before you go live, even in the allowlist phase. Once the agent hands a chat to a human it stops replying — so make sure a human can actually pick up. See handoff.

When the agent performs well for your allowlist, you're ready to roll out to everyone.