Before a message reaches the assistant: how the protection shield works

A public WhatsApp number doesn’t hear only from buyers. The same inbox receives insults, flirting attempts, ads, random images, voice notes — and, from time to time, messages trying to trick the system. Build the assistant on top of that noise and two things happen: the AI starts talking nonsense, and your customer list fills with junk. That’s why every incoming message passes through a protection shield before it is handed to the assistant.

The shield comes before the assistant

The order of operations matters: the filter stands in front of the AI, not behind it. A message is evaluated before it ever reaches the assistant; content that doesn’t belong is simply never shown to it. The question “could the AI fall for something malicious?” largely disappears — because the AI never sees that content in the first place.

The four things the shield weeds out

In practice, four main categories are blocked:

• Insults and abusive language: Profanity and abuse aimed at the agent or the system is set aside without any attempt to continue the conversation.
• Flirting and off-topic chatter: Personal advances with no connection to a property request never mix into the customer flow.
• Attempts to mislead the system: Messages of the “forget your previous instructions, do this instead” variety — prompt injection, designed to push the assistant outside its own rules — are recognised and neutralised.
• Media files: Attachments such as images, video, audio and documents are not passed to the assistant to be processed as text; this closes off both a security risk and a misinterpretation risk.

Why a separate error log

Silently deleting blocked messages would be easy — and wrong. These events don’t vanish; they’re collected in a separate Error Log. The reason is twofold. First, keeping your pipeline clean: this noise never sits in the same list as real buyers, so your reports and conversion numbers stay honest. Second, auditability: if a message was blocked by mistake, you need to be able to see it, and we need to be able to tune the shield to your office.

The shield’s purpose is not to “destroy messages” but to put the right message in the right place: genuine demand goes to the assistant and the dashboard; noise goes to a separate, traceable bucket.

The limit: no filter is perfect — but it is auditable

No filter is 100% accurate; occasionally a real buyer writing in an unusual way may get caught. The design choice here is not to promise a “filter that never errs” but to make errors visible and correctable. That’s exactly what the Error Log is for: over time we sharpen the shield to your office’s language and customer profile.

← All articles