A good chatbot is not a prompt behind a speech bubble. It needs a knowledge source, session logic, explicit answer boundaries, a verifiable human handoff, and metrics for actual quality. It does not replace your team; it handles recurring questions, gathers context, and passes difficult cases over in a structured way.
System components
- Channel & identity: website widget, customer portal, Slack, or Teams; anonymous, authenticated, or role-based depending on the data and audience.
- Knowledge access: RAG over approved website content, manuals, wiki pages, or product data. Sources are shown, and missing evidence triggers a question or handoff.
- Model layer: suitable hosted APIs or local models through Ollama/vLLM. Routing and fallbacks can treat simple, confidential, or especially demanding requests differently.
- Conversation state: only the context needed for the dialogue is retained; retention, deletion, and handoff are defined deliberately.
- Tools: availability checks, ticket creation, or CRM handoff use narrow, server-validated functions — not freely interpreted model text.
- Escalation: the team receives a summary, conversation context, retrieved sources, and the concrete reason for handoff.
What gets tested
- answer quality on real frequent and difficult questions
- grounding and correct behaviour when no reliable source exists
- prohibited topics, prompt injection, and attempts to retrieve confidential content
- handoff rate, drop-off points, and genuinely resolved requests
- latency, cost, and behaviour under concurrent conversations
- regressions after content, prompt, retrieval, or model changes
Typical use cases
- lead qualification with structured CRM context
- customer questions about services, prices, delivery times, or processes
- internal helpdesk over approved documents
- pre-sales guidance with an appropriate human handoff
What’s included
- define conversation goals, knowledge boundaries, tone, and escalation rules
- build a RAG pipeline or structured data connection
- integrate the chat frontend and authentication
- configure model access, session state, tool functions, and fallbacks
- set up an evaluation set, tracing, and usage/quality metrics
- document content and operations workflows plus 30 days of post-launch support
Clear boundary
No phone bot, legally binding advice, or unsupervised high-impact decisions. Unclear, sensitive, or critical requests must hand off to people in a controlled way.
After discovery, usually 4–8 weeks depending on the knowledge base, integrations, frontend, and agreed evaluation and approval depth.