System note / conversation

A chatbot earns trust by knowing its boundaries.

The model is only one part of a useful assistant. The source set, retrieval rules, failure state, escalation path and review loop determine whether the answer helps.

Short answer: a grounded chatbot answers from an approved source set, shows uncertainty when the source is missing and routes a human when the question needs judgment.

Grounding begins before the model

Start by deciding which documents are authoritative. That might include a service catalogue, policy pages, product manuals, internal FAQs or approved public content. Give each source an owner, a version and a review date. Without that boundary, retrieval can only make an unclear knowledge base easier to search.

Design the conversation around intent

People rarely arrive with the exact label used inside a business. A good assistant recognises the job behind the question: compare options, check eligibility, find a document, book a next step or speak to a person. The response should then answer the immediate question and offer the next useful action.

This is where a small interface can outperform a large prompt. Suggested follow-up questions, a clear source link and a human handoff reduce the amount of guessing the model has to do.

Make failure a visible state

“I could not find that in the approved information” is more useful than a confident invention. Define what happens when retrieval is weak, the question is outside scope or a source is contradictory. For sensitive topics, the handoff should happen earlier and the interface should say what the assistant can and cannot do.

Measure useful outcomes

Conversation length is not the goal. Track resolution rate, unanswered intents, handoff quality, source clicks, repeated questions and user corrections. A shorter conversation that resolves the task is healthier than a long loop that merely keeps someone on the page.

Review a sample of conversations regularly. Look for stale documents, ambiguous policy, retrieval misses and prompts that encourage the assistant to sound certain when it should ask for context.

From prototype to operating system

An assistant becomes part of a real workflow when it can safely read approved information, collect the right context, route the next action and leave an audit trail. That may involve a website, WhatsApp or Telegram interface, but the channel is secondary to the source and handoff design.

Explore the grounded chatbot service, the wider AI systems practice or the search visibility blueprint that helps the underlying knowledge become discoverable.

Design principle: use engagement to create clarity and momentum, never to hide a limitation or pressure someone into a decision.

Continue the thread

Read the companion note on SEO, AEO and GEO, inspect the public project work, or bring a real workflow to the studio.