A business chatbot is software that accepts a question in natural language and returns an answer grounded in content or systems the operator has approved. An AI agent, on Sigi’s AI agent development page, is different: it completes multi-step tasks through tools and APIs with controls, approvals, and auditability. Copilots and generation features sit on generative AI development. Retrieval-heavy assistants sit on RAG assistants.
Decide the job before picking a model
- FAQ deflection: answers from a reviewed knowledge base. Fail closed when the answer is not in the corpus.
- Authenticated help: the bot may read account or order state through APIs, with the same permissions a support agent would have.
- Handoff: collect context and open a ticket or chat with a human. Many “chatbots” should stop here.
- Agent work: create or update records, trigger workflows, or take payments. That is the AI-agent product, not a widget on a marketing site.
A build sequence that stays evaluable
- Write the allowed questions and the forbidden actions. If the bot must never invent a price or a clinical instruction, that is a product rule, not a prompt afterthought.
- Choose retrieval before “fine-tune everything.” Sigi’s RAG page exists because many business chatbots are search-plus-generation over owned documents.
- Put a human fallback on day one. Unanswered questions are a dataset, not a failure to hide.
- Measure answer faithfulness against the corpus, not only user thumbs-up. Cost and reliability work is the subject of Sigi’s LLM optimization page.
- Promote tool use only after the Q&A path is boringly correct. Sigi’s agent page starts with workflow mapping, then a small prototype, then phased rollout with permission boundaries.
What has to be designed besides the chat window
- Source of truth: help centre, policy PDFs, product catalog, or live APIs — named, versioned, and reviewable.
- Identity: anonymous marketing chat versus signed-in support with audit logs.
- Languages and tone: ServiPR is bilingual because the market is; a bot should match how customers already write, but that is a requirement, not a ServiPR-AI claim.
- Regulated content: Allfor Care shows how UK care data is permissioned. A care chatbot that quotes clinical records would inherit those access rules. Sigi has not published such a bot.
Related Sigi reading
Capability pages: AI agents, generative AI, RAG assistants, and LLM optimization. For a regulated records system rather than a bot, see CQC and UK GDPR for care software. For product cost context (mobile, labelled estimates), see how much it costs to build a mobile app.