A chatbot that answers confidently and wrongly costs more than no chatbot. The whole design problem is knowing where the assistant’s knowledge ends.
Grounded in your content, not the internet
The assistant answers from your help centre, your product docs and your past tickets. Every answer carries a link to the passage it came from, so a user can check it and an agent can correct it. Content it has never seen is a handoff, not a guess.
The handoff is the feature
We build the escalation path first: when confidence drops, when the user asks twice, when the topic is billing or cancellation, the conversation goes to a person with the full transcript attached. Nobody has to repeat themselves.
Sales assistants are a different job
Support wants resolution. Sales wants qualification. A pre-sales assistant asks the three questions your team would ask, books the call and writes the summary into the CRM. That last part is where the value is, and it is CRM work as much as model work.
What we measure
- Resolution rate, not deflection rate
- Satisfaction on assisted conversations against unassisted ones
- Handoff quality: how often the human has to start over
- Cost per conversation, tracked from day one
Deflection alone is a vanity number. It counts people who gave up.
Voice, tone and the brand
The assistant is a public surface. It gets the same tone rules as the rest of the site, which come out of strategy and branding rather than out of a default system prompt.
Launch and after
We ship behind a flag, to a slice of traffic, with a kill switch. Then we read the transcripts weekly and fix the top failure. That loop is support and evolution, applied to a feature that changes every time the underlying model does.
Next: LLM development services, or the whole AI map.