To filter spam and time-wasters out of WhatsApp and Instagram DMs, use three layers. Meta's own tools hide or block the worst senders. An AI answers routine questions from your notes. Clear rules then send refunds, complaints, urgent cases and ready buyers to a person.
Below are the settings for each layer, a test list you can run, and the limits. We read Meta's pages and Instant Reply's code on November 14, 2026.
See your AI answer your customers
The three layers at a glance#
| Layer | What it catches | Where it runs |
|---|---|---|
| 1. Meta's tools | Scam requests, words you pick, people you block | The Instagram and WhatsApp apps, and the WhatsApp API |
| 2. AI for routine questions | Prices, hours, location, abuse and floods | Your inbox tool |
| 3. Handoff rules | Refunds, complaints, urgent cases, hot buyers | Your inbox tool, then your team |
Layer 1: Meta's own tools#
Instagram: Hidden Words#
Instagram's Hidden Words page explains how to hide comments and message requests that contain words you choose.
- Open Settings in the Instagram app.
- Under How others can interact with you, tap Hidden Words.
- Tap Manage custom words and phrases.
- Add words, phrases, numbers or emojis. Split them with commas.
The same page lists two more filters. One hides comments with common offensive words, and it is on by default. The other moves message requests that may be scams into a separate folder. That one is off by default, so turn it on.
Note: the page says your custom list does not filter message requests on the web. We read it in Arabic, so the English menu names may differ a little.
WhatsApp Business app: block and report#
Open the chat, open the menu and tap Block. WhatsApp's help page says a blocked number can no longer call you or message you.
Reporting sends WhatsApp up to the last five messages from that person. WhatsApp does not tell them they were reported. The page also mentions an option to block large numbers of messages from unknown accounts.
WhatsApp API numbers#
If your number runs on the WhatsApp Business Platform, Meta has a Block Users API. Meta's page sets three rules.
- You can only block users who messaged you in the last 24 hours.
- One request can block up to 1,000 users.
- The block list holds up to 64,000 users.
Instant Reply does not call this API for DMs. To block a sender on an API number, use the API directly or ask your provider.
Layer 2: let the AI take the routine questions#
Many inboxes get the same few questions again and again: price, hours, location and what is free this week. An AI can answer those from the notes you save. You choose how much it does on its own.
- Approval: every AI reply waits as a draft for a person to send.
- AI can hand off: the AI answers and calls in a person when it needs one. This is the default.
- Full auto: the AI sends every reply.
Before the AI sees a message, Instant Reply runs a filter called Troll Shield. It is on by default. It stops three kinds of message.
- Abuse, such as common swear words aimed at you.
- Threats, such as "I will hack you".
- Floods: one character typed 10 or more times in a row, or a 40-character message in capital letters.
A stopped message gets one fixed reply and no AI answer. A sender who sends more than 30 messages in 60 seconds gets no more replies in that minute.
On comments, Troll Shield matches are hidden and never answered. Once three of one person's spam or abusive comments are hidden, Instant Reply blocks them. On Facebook it blocks them on your Page. On Instagram it adds them to its own block list, so their next comments are skipped.
Layer 3: send the right chats to a person#
| Message | What happens |
|---|---|
| "Talk to a person", "real person", "live agent", or keywords you add | A short holding reply. The chat goes to your team |
| A refund ask, a disputed charge, a scam claim or a threat to complain | No AI answer. A holding reply, an urgent ticket and an alert to the owner |
| Words about self-harm | No AI answer. One fixed line that points to help, and the chat is flagged for a person |
| Urgent symptoms, such as heavy bleeding or chest pain | A fixed safety reply with your urgent number. Your team is alerted |
| The AI is unsure (below a 0.70 confidence bar by default) | It asks a short question, or sends a holding reply and opens a ticket |
| A lead turns hot or very hot (Growth and up) | One alert to the owner per chat, by email or Slack, and by WhatsApp if you set up alerts |
The refund and complaint rule matters most for spam-heavy inboxes. Angry or money messages never get an AI guess. A person sees them first.
How the AI spots a ready buyer#
On Growth and above, each chat gets a lead temperature: cold, warm, hot or very hot. The code looks at signals in the chat. Examples are a price question, a booking question, contact details, how urgent the person sounds, and whether they say they want to buy.
Chats with no buying signal stay off your Customers page. A chat joins it once it looks warm or hotter, reaches a buying stage, or needs support. So browsers still get answers, but your team's list shows the buyers. The lead management page shows what each card holds.
Who gets the chat after a handoff#
On Growth, you get 5 seats and assign chats by hand. On Pro ($349 a month), lead routing can assign them for you. It picks an online teammate with room, using each person's share and max leads.
If nobody is free, the chat waits in a queue. A long wait alerts owners and admins. Routing has no rules by region, language or product. If you need sales and support split by topic, assign those chats by hand. See the lead routing page for the settings.
Test your filter with these messages#
Send these from a test phone, not a real customer's. Each result comes from the Instant Reply code.
| Send this | Expected result |
|---|---|
| aaaaaaaaaaaa | Troll Shield sends one fixed reply. No AI answer |
| I want a refund | No AI answer. A holding reply and an urgent ticket |
| Can I talk to a real person? | A holding reply. The chat goes to your team |
| I have heavy bleeding | The fixed safety reply. Your team is alerted |
| What time do you open? | The AI answers from your notes |
For a longer list of hard cases, use our WhatsApp sales bot test pack. Clinics can pair this with appointment booking in the chat.
The limits#
- Troll Shield's word list is in English. The flood checks work in any language.
- Instant Reply does not block WhatsApp senders for you. Use WhatsApp's own tools.
- Routing does not split chats by topic, region or language.
- Lead temperature needs Growth or above.
What we could not confirm#
- How much spam each layer catches. We have no measured rate, so we give none.
- The exact English menu names for Hidden Words. We read Instagram's page in Arabic.
- Whether Meta's Block Users API works on numbers that also run the WhatsApp Business app. Meta's page does not say.



