What is an AI Chatbot?

A properly built chatbot is a retrieval system first and a conversation system second. Every substantive answer traces back to a specific passage in your documentation, product data, or ticket history, and the interface shows where it came from. When retrieval finds nothing sufficient, the system says so and hands off to a person rather than improvising.

Consider a pricing question. An ungrounded chatbot generates a plausible-sounding answer that may or may not match current pricing. A grounded one retrieves the actual pricing page, cites it, and if the plan structure changed last week and the index hasn't caught up, says it can't confirm and routes to a person, rather than guessing confidently.

Retrieval quality is where most of the engineering effort goes, and where most of the answer quality comes from. Hybrid search combining semantic and keyword matching, tuned chunking per content type, and re-ranking all matter more to the final answer than which model does the writing. The model's job is synthesis constrained to what was retrieved, not generation from open-ended knowledge.

• Content pipeline — ingesting docs, help centers, product data, and resolved tickets, with incremental sync so answers track reality, not a stale snapshot.
• Retrieval — hybrid semantic and keyword search plus re-ranking, tuned per content type since a pricing table and a troubleshooting article retrieve badly under the same naive strategy.
• Synthesis — answer generation constrained to retrieved context, with citations attached and explicit refusal when evidence is thin.
• Handoff — escalation on low confidence, detected frustration, or regulated topics, carrying the full conversation and evidence to a human agent.
• Evaluation — a growing set of real questions with known correct answers, run before any prompt, index, or model change ships.

Teams that struggle with chatbot quality almost always under-invest in the content pipeline and retrieval tuning while over-investing in prompt engineering for the synthesis step. A better prompt rarely fixes retrieval that surfaced the wrong passage in the first place.

• Chatbots answer questions in a conversation, grounded in retrieved evidence, with actions layered on top behind explicit confirmation.
• Agents hold a multi-step goal and decide their own sequence of actions to pursue it, with less conversational framing and more autonomy.
• Automation runs a defined workflow, with AI applied at specific decision points rather than directing an open conversation.

The three aren't strictly separate in production. A chatbot's transactional actions are often small automations, and a chatbot that needs to research across several systems before answering starts to look like an agent. Read our AI agents guide and AI automation guide for the other two corners of this spectrum.

• Support and FAQ bots — answer product and policy questions from documentation, escalating what they can't resolve.
• Sales qualification bots — capture structured requirements and book directly into the right calendar.
• Transactional bots — authenticated, account-aware bots that handle scheduling, plan changes, and returns behind confirmation.
• Internal assistants — HR policy, IT support, and sales enablement bots with visibility restricted by role.

Start with support or FAQ. It has the clearest definition of a correct answer and the most existing documentation to ground against, making it the fastest path to a real quality baseline before expanding to transactional or internal use cases.

• Tier-one support deflection for policy and product questions
• Sales lead qualification and meeting booking
• Internal knowledge retrieval across fragmented documentation
• Account-aware order status and simple transactional requests

Containment rate alone is a misleading metric, a bot that frustrates users into abandoning the conversation scores well on it. Measure resolution rate confirmed by follow-up behavior, answer accuracy against ground truth, and satisfaction split between contained and escalated conversations, not raw deflection volume.

• Answer synthesis constrained to retrieved context, with explicit refusal when evidence is thin
• Treating retrieved content and user input as untrusted, prompt injection through either is a live threat
• Escalation on low confidence, detected frustration, and regulated topics
• Transcript retention under policy with PII redaction where required
• No answer path reaching systems the authenticated identity isn't entitled to reach

Every unanswered or badly answered question is a content gap, not just an incident. Feeding it back into the documentation pipeline is what makes the system compound in value over time instead of plateauing at launch quality.

• Is there existing documentation reasonably organized enough to ground against?
• Is there a way to measure answer accuracy against known-correct answers?
• Who receives escalations, and how fast can they respond?
• Is there a process for feeding content gaps back into the documentation?

Three or more clear answers usually means the deployment is ready to pilot. Fewer than that, the highest-leverage first step is often organizing the documentation itself, a chatbot grounded in scattered or contradictory content inherits that inconsistency.