Chatbots that help — and the ones that hurt

Everyone has met the bad chatbot. It pops up uninvited, understands nothing, offers four buttons that don’t match your question, and guards the contact details like a dragon. It was installed to save money, and it does — by driving away the customers who would have cost money by buying things.

And yet: a well-built assistant answering at 2am, accurately, in your company’s voice, is one of the highest-ROI things you can add to a website. Same technology category, opposite outcomes. The difference is buildable.

Why bad bots are bad

Bad bots share a root cause: they’re bolted on, not wired in. A generic widget that knows nothing about your business, connected to nothing, with no plan for failure. It can only deflect — and customers feel deflected, because they are being deflected.

What the good ones do differently

• They actually know your business. Grounded in your real services, prices, process and FAQs — answering from your content, not improvising from the internet. When it doesn’t know, it says so.
• They’re connected. A useful assistant can check availability, book a call, take an order status query, create a lead in your system. Answers without actions are just a search box with manners.
• They hand off honestly. The moment a conversation needs a human — complexity, complaint, high value — the good bot says so, passes the full transcript, and gets out of the way. No loops, no “I didn’t understand that” purgatory. The hand-off isn’t failure; it’s the feature.
• They’re accountable. Every conversation logged and reviewed. You read the transcripts monthly, fix the gaps it exposed, and let its accuracy record earn it more autonomy over time.

The measure that matters

Don’t measure “deflection rate” — that’s the metric that built the bad bots. Measure resolution and capture: questions genuinely answered, leads collected out of hours, bookings made, and the satisfaction of the people who used it. An assistant that resolves 60% of enquiries well and hands the rest to a human gracefully is worth ten that “deflect” 95% badly.

Rule of thumb before you ship one: read fifty test transcripts yourself. If you’d be happy for your best customer to have any of those conversations, ship it. If not, the model isn’t the problem — the integration is.