The questions your chatbot should never try to answer alone

A chatbot's most important skill isn't answering well — it's knowing precisely which questions it shouldn't touch, and handing those off before it does damage.

The categories that need a hard stop, not a best guess

We build explicit exclusion rules into every chatbot deployment — categories of question the bot is instructed to never attempt, regardless of confidence. Four come up in nearly every industry: anything involving a legal determination (is this covered under warranty, does this violate our contract), anything involving a medical or safety judgment, any request to cancel or modify a financial commitment above a defined threshold, and anything where the customer expresses clear distress or anger. These aren't "the bot might get it wrong" categories — they're "getting it wrong here causes real harm or liability" categories.

How this gets enforced technically, not just as a prompt instruction

Telling the model "don't answer medical questions" in a system prompt is necessary but not sufficient — models can still occasionally comply with a rephrased request that circumvents the instruction. We layer a second check: a keyword and intent classifier that runs before the model generates a response, flagging messages that match exclusion categories and routing them to a human queue without ever generating a bot answer. This means the restriction doesn't depend on the model reliably following instructions every time — it's enforced structurally, upstream of generation.

For regulated industries specifically — healthcare, financial services, insurance — this matters for compliance, not just quality. If a bot in a HIPAA-covered environment gives medical guidance, that's a real regulatory exposure, not just a bad customer experience.

The sentiment threshold that triggers escalation

Beyond fixed categories, we run lightweight sentiment scoring on every incoming message. A single negative message doesn't trigger escalation — customers are often just describing a problem, not upset at the bot. But two consecutive messages with negative sentiment, or specific escalation-signaling phrases ("this is ridiculous," "I want to speak to someone," "this is the third time"), route to a human immediately, with the full conversation history attached so the person doesn't have to ask the customer to repeat themselves.

Where teams get this wrong

The most common mistake is over-trusting a general-purpose model's judgment on where its own limits are. Left to its own instincts, a capable model will often try to be helpful on a borderline medical or legal question rather than declining — because "be helpful" is a strong default behavior. The fix isn't a better prompt; it's not letting the model make that call at all for the categories that matter most.

The second mistake: making the escalation path clunky. If handing off requires the customer to re-explain their entire issue to a human, you've erased the value of catching the escalation early. The handoff needs to carry full context automatically.

How Ndakum approaches it

Defining these boundaries — what the bot should never touch, and how the handoff carries context — is one of the first design conversations in every AI Chatbot engagement, before we write a single prompt.