AI chat widgets on B2B sites are easy to install and easy to get wrong. The failure mode we see most is not "the model is dumb" — it is that the knowledge base never received the sentences sales already uses on calls: pricing floors, scope boundaries, regions served, and disqualification rules. The assistant then invents friendly nonsense.
What belongs in the knowledge base first
Prioritize extractable facts over brand essays:
- Who you serve (and who you do not)
- Starting budgets / fee ranges with conditions
- Onboarding steps and typical timelines
- Documents required from the client
- Regions and languages supported
- What happens after a lead submits a form
If a human SDR would refuse to quote without a call, teach the bot to say the same — and to collect the fields that make the call useful.
Structure beats dumping PDFs
Chunk content by question, not by PDF page. Each chunk should answer one intent: "minimum monthly ad budget", "do you open TikTok accounts for games", "which countries for crypto merchants". Include the canonical URL of the public page so the bot can cite it.
Refresh chunks when pricing or policy changes. A stale minimum budget in the KB is worse than no budget at all, because prospects screenshot it.
Qualification logic
Good assistants ask 3–6 questions before deep product talk: company type, target markets, monthly budget band, timeline, existing ad accounts. Route high-fit leads to calendar or WhatsApp; route low-fit to a resource email. This is ordinary sales process, automated.
Safety rails
- System prompt: never invent fees; if missing, say you will have a human follow up
- Block list: competitors' libel, illegal verticals you do not serve, credential requests
- Human handoff visible within two turns when confidence is low
- Log prompts and answers for weekly review
A starter overview of the product side sits at overseas marketing automation / AI sales assistant. Brand home: PandaWM.
Success metric
Not "number of chats". Use: qualified conversations per week, form completion rate, and % of answers that match the approved KB (spot-check 20 transcripts weekly). Hallucinated pricing should be a P0 defect.
Updated October 2026.
Top comments (0)