Handy Harry: a chat assistant that qualifies leads on its own

Across several websites, enquiries came in through a standard contact form: a name, an email address and a short message. That's where the real conversation had to start, because a colleague still had to figure out what the visitor actually needed, how urgent it was, and what situation was behind the request. Visitors asking a question in the evening or over the weekend got no response at all until someone was available.

The challenge wasn't simply to bolt on a chatbot. A chat assistant had to behave like an experienced intake specialist: calmly asking follow-up questions without being pushy, never promising prices or guarantees, and knowing exactly which brand it represented on each site — without ever mixing up companies from the same network. On top of that, the system had to keep working even without an active AI connection, so a technical hiccup would never leave a visitor stranded.

We built a lightweight chat widget that shows the same recognisable icon on every website, with a clear “Direct contact” label on hover. Behind the button sits an AI agent that gets its own knowledge source per website: a central, continuously updated knowledge document with that specific brand's services, frequently asked questions and safe conversation boundaries. The agent recognises which piece of context is still missing from the conversation and asks a targeted follow-up question about exactly that, instead of working through a fixed script.

The first version went live with one simple promise: never leave a visitor empty-handed. Without an active AI connection, the widget automatically fell back to a clean, fixed question so a lead was always captured. From there, the setup improved in small steps: first the connection to a real language model, then a richer knowledge document per brand.

During rollout we ran into a subtle bug where the knowledge retrieval structurally leaned toward one brand. It only became visible once multiple brands were running on the same underlying technology — a good lesson that shared knowledge sources always need to be tested per brand, not just as a piece of technology in isolation.