WhatsApp FAQ answerer

What you’ll have at the end

A WhatsApp number that, when customers message it, answers their FAQ in plain language using your knowledge base — and remembers past customers so repeat questions get answered with continuity. About 90 minutes if you’ve got hello-world (§5.1) running.

This is the “small business, no chatbot platform” recipe. Useful for a freelancer, a tutor, a small studio, a consulting practice — anyone who fields the same questions repeatedly and wants the agent to take the first cut.

Why this recipe matters

Most “AI chatbot” products are SaaS — you pay $50–500/month, you put your knowledge in their UI, you live with their privacy posture. This recipe is the self-hosted, your-data-on-your-machine version. Same outcome (FAQ answered automatically), no SaaS dependency, no subscription, your customers’ messages don’t leave your network unless they go to your model provider.

Trade-off: you maintain it. SaaS chatbots are turn-key; this is your responsibility.

What you need

• Hello-world working (§5.1)
• A WhatsApp Business account (different from personal WhatsApp; needed for the official channel)
• A knowledge base — could be a single FAQ markdown file, could be a Notion export, could be your existing website’s FAQ page
• About 90 minutes (most of it is the WhatsApp Business setup)

Step 1 — Set up WhatsApp Business + the channel

WhatsApp Business has its own onboarding distinct from personal WhatsApp. You’ll either:

• Use WhatsApp Business Cloud API (Meta-hosted, requires a verified business; better long-term)
• Use BlueBubbles-style bridging (informal, can use your personal WhatsApp; not for production-public)

For a small business, the official Cloud API is the right path. Follow docs.openclaw.ai/channels/whatsapp — it walks through the Meta onboarding and the OpenClaw pairing.

Once paired:

openclaw channels list
# whatsapp: paired (number: +64xxxxxxxxxx)

Step 2 — Configure DM policy

This is the most important security step in this recipe. WhatsApp FAQ-bot is public-facing — random people will message it.

{
  "channels": {
    "whatsapp": {
      "dmPolicy": "open",
      "allowFrom": ["*"],
      "rateLimits": {
        "perSenderPerHour": 20,
        "perSenderPerDay": 100
      }
    }
  }
}

This is one of the few times we recommend dmPolicy="open" + "*" — for a public FAQ bot, that’s the desired behaviour. But make sure:

1. Sandbox is on for non-main sessions (§6.1 check 4). Public input → adversarial-input-likely.
2. Rate limits are set (above) — limits abuse + caps your model API spend.
3. Tool policy denies dangerous tools for non-main sessions — exec, browser, write outside scope, etc.

Without these three, you’ve handed the agent’s full power to anyone who finds your number. With them, you’ve handed a scoped FAQ-answering capability to anyone — which is what you want.

Step 3 — Load the knowledge base

The agent’s knowledge base lives in the workspace. Put your FAQ content in ~/.openclaw/workspace/faq/:

faq/
  pricing.md
  services.md
  how-we-work.md
  scheduling.md
  policies.md

Each file is markdown with the questions/answers in your own voice. The agent will read these to answer customer questions.

For a slightly more sophisticated setup, use QMD memory (§3.5) — it indexes the markdown semantically so the agent finds the right answer faster:

{
  "agents": {
    "defaults": {
      "memory": {
        "engine": "qmd",
        "indexPaths": ["~/.openclaw/workspace/faq/"]
      }
    }
  }
}

Step 4 — Edit AGENTS.md to give it the FAQ bot persona

## WhatsApp FAQ behaviour

You handle inbound WhatsApp DMs from customers asking about my services.

When someone messages you:
1. Look up the answer in ~/.openclaw/workspace/faq/ (or QMD memory).
2. If you find it, answer in plain language. Quote prices/dates/specifics directly.
3. If you don't know — say so explicitly. Offer to take a message and I'll reply.
4. Never invent details (dates, prices, capabilities) that aren't in the FAQ.
5. If they ask something the FAQ covers but they want detail beyond what's there, offer to take a message.

Tone:
- Warm, professional, concise.
- No corporate jargon.
- Match their formality level — if they say "yo, do you do X?" reply casual; if they say "Hello, I am inquiring about your services," reply polite.

When you take a message:
- Format: NEW MSG FROM <name/number> at <time>: <message + your handoff context>
- Save to ~/.openclaw/workspace/messages.md
- Confirm to the customer: "I've passed this to <my-name>. They'll reply within X hours."

Save AGENTS.md.

Step 5 — Set up the message-review cron

The customer’s “I’ll reply within X hours” requires you to actually see the new messages. Cron the agent to summarise pending messages once a day to your own DM:

openclaw cron add \
  --name "customer-message-digest" \
  --schedule "30 8 * * *" \
  --task "Read ~/.openclaw/workspace/messages.md. DM me on Telegram with new messages from the last 24 hours. Format: 'X new customer messages: <list>'. After DMing, archive the messages to ~/.openclaw/workspace/messages-archive/<YYYY-MM-DD>.md."

Now you get a daily digest of inbound customer questions. Reply by hand to the ones the bot couldn’t answer.

Step 6 — Test as a customer

Pretend to be a customer. From a different phone (or borrow a friend’s), message your business number with FAQ questions:

• “Do you do remote sessions?”
• “What does your pricing look like?”
• “Are you available next Wednesday at 3pm?”

Verify each one:

• Gets a sensible answer if the FAQ covers it
• Gets “I’ll pass this on” if it doesn’t
• Falls into messages.md correctly when handed off

If anything goes wrong, edit AGENTS.md (or the FAQ markdown) and re-test.

What this looks like in practice

Day 1 — set up. Test with friends. Tweak.

Day 7 — you’ve answered ~10 real customer questions, ~3 of which the bot couldn’t handle and you replied to via the cron digest. Bot accuracy is improving as you tune the FAQ.

Day 30 — bot handles ~85% of inbound questions. The 15% you handle manually are the genuinely-novel ones. You spend less time on FAQ, more time on real work.

This isn’t a magic-AI win; it’s a small operational lift that frees up real hours.

Things to try beyond the basics

• Booking integration. Wire a calendar MCP. When a customer asks “are you free Wednesday at 3pm?” the bot checks your calendar and offers slots if free.
• Per-customer memory. Use Honcho or QMD scoped per phone number — repeat customers get continuity (“oh hi, you asked about pricing last week, did you go ahead with that?”).
• Multiple business numbers. Multi-agent routing — different bots for different businesses, same Gateway.
• Voice notes transcription. WhatsApp voice notes → Whisper transcription → agent answers. (Requires more setup; deferred to a future recipe.)

What we are NOT going to claim

WhatsApp Business onboarding has changed multiple times in recent years. The exact Meta steps aren’t reproduced here because they drift fast — verify against WhatsApp Business Cloud API docs before you start. The OpenClaw pairing side is straightforward once you’ve got the Meta side sorted.

Common pitfalls

What to read next

• §5.2 Slack agent that does real work — internal-team analogue of this
• §5.5 RAG over personal docs (planned) — same shape, but for personal knowledge base
• §3.5 Memory — for the QMD setup
• §6.1 Self-hosting checklist — public-DM-bot hardening