AI Glossary / Governance, ethics and safety / Beginner
Chatbot
Definition
A chatbot is a software application that interacts with people through conversational exchanges, usually text or speech. It may follow scripted rules, retrieve prepared information, generate responses with AI or combine these approaches to support a defined set of tasks.
What it means
A chat window describes how the user interacts with a service, not how the service reasons or what authority it has. One chatbot may only display opening hours; another may search internal documents; a third may connect to systems that change bookings.
Not every chatbot uses a large language model, and not every language model is deployed as a chatbot. A rule-based bot can be useful for predictable questions with tightly controlled answers. A generative bot can accommodate more varied wording but needs evaluation of factuality, scope and behavior.
A conversational style can make software feel socially familiar. That does not establish human understanding, emotional experience or reliability. Users need clear information about the bot's purpose, capabilities and how to reach another source of help.
How it works
The application receives a message and determines the next response or action using its conversation state and available information. A scripted system may match an intent and follow a decision tree. A retrieval-based system searches for relevant content. A generative system produces text from its model and supplied context.
Many practical systems combine methods: explicit rules handle account access, retrieval supplies current policy information and generation turns that material into a readable answer. Connected tools may check a booking or submit a request, but those actions need appropriate permissions and confirmation rules.
Conversation history can help interpret follow-up questions, yet memory is limited by the application's design. The bot should distinguish what it knows from what it needs to ask. Evaluation includes complete conversations, failed tool calls, ambiguous requests and hand-offs, not just isolated attractive responses.
Three practical examples
- A website bot answers standard delivery questions using maintained policy content and provides a route to staff for exceptions.
- An internal help bot searches authorized technical documents and points employees to the source used in its answer.
- A booking assistant gathers the date and party size, checks availability and submits a reservation only through the authorized booking workflow.
Think of it this way
Think of a conversational reception desk that helps direct a visitor. Its usefulness depends on the information and services behind it, not only on how pleasantly it speaks.
The analogy does not imply that the bot has a receptionist's understanding or accountability. A fluent reply can be wrong, and a statement that a request is complete needs confirmation from the system responsible for completing it.
Worked example
A response is not a completed action
Consider an invented reservation conversation. A user asks for a room on Friday, the bot identifies the intended date and retrieves three available options. The user selects one, and the bot sends a booking request with a unique request identifier.
The booking service returns a temporary failure rather than a confirmation. A correct next response explains that the reservation is not yet confirmed and offers the defined retry or hand-off path. It should not generate “Your room is booked” merely because that is a plausible final sentence.
If a retry later succeeds, the bot can report the returned confirmation. The sequence separates understanding a request, proposing an option, attempting an action and establishing its completion. Each is a different event that conversation quality tests should cover.
Illustrative case study
Fictional Cedar Tours launches a chatbot for itinerary questions. Early evaluations focus on friendliness and response speed. During a pilot, a customer asks whether an unavailable excursion has been reserved, and the bot produces an enthusiastic confirmation without checking the booking system.
The team changes the workflow so booking statements depend on verified service results. It separates informational answers from transactions, tests unsuccessful requests and makes staff hand-off visible when the bot cannot resolve a case.
It also checks whether source links support the policy claims in generated answers. The case illustrates that a chatbot is an application with operational responsibilities, not simply a text generator placed behind a chat box.
Common misconceptions and mistakes
- Every chatbot is generative AI. Scripted and retrieval-based bots also exist.
- A friendly tone proves understanding. Tone and task reliability are different properties.
- The bot remembers everything previously said. Available history and storage depend on the application.
- A stated action has necessarily happened. Completion must be confirmed by the responsible service.
- More conversation always means better service. Some tasks are easier with a form, a clear link or a human hand-off.
Business use
Businesses should choose chatbot scope around a real user need and maintained information sources. Define which questions it can answer, which actions it may initiate and when it must ask for clarification or refer the user elsewhere. Keep conversation data handling explicit and proportionate to the task.
Measure successful task completion, supported factual answers, appropriate escalation and user effort alongside response speed. Test the actual language and circumstances of the intended audience. A chatbot can improve access to information, but an unclear interface or unreliable action claim can add friction even when the generated wording is polished.
People also ask
Is a chatbot the same as an AI agent?
No. A chatbot is a conversational interface; it may or may not include an agent that plans and acts through tools.
Can a chatbot use current company information?
Yes, if the application retrieves or connects to maintained sources with appropriate access controls. The chat format alone does not provide that information.
Does it learn from every conversation automatically?
Not necessarily. Logging, memory and model training are separate design choices.
Check your understanding
A chatbot writes a booking confirmation after its booking tool returns an error. Has the user received a reliable completion notice?
Answer: No. The message contradicts the action result. The bot should report the unresolved status and follow the authorized recovery process.
Related terms
- Large language model
- Retrieval-augmented generation
- AI agent
- Conversational AI
- dialogue system
- intent classification
- human hand-off
Learn next
Read dialogue system and retrieval-augmented generation. Design a short conversation for one user task, including missing information, an unavailable result and a failed action. Specify what evidence supports each completion message.
More in Governance, ethics and safety
- Business process management - Business process management, or BPM, is the systematic examination, design, operation and improvement of an organization's recurring work processes.
- Business rule - A business rule is a precise statement that defines or constrains an aspect of how an organization operates.
- Calibration - Calibration describes how well predicted values agree with observed outcomes in a defined population and setting.
- Cognitive Bias - A cognitive bias is a systematic pattern in judgment or decision-making that can lead people away from an appropriate evaluation of evidence.
- Column - A column is a named vertical part of a table that holds values for a particular attribute or field across its rows.
- Confirmation bias - Confirmation bias is a tendency to seek, interpret, favor or remember information in ways that support an existing belief or hypothesis.
Sources
IBM, What is a chatbot?: https://www.ibm.com/think/topics/chatbots. IBM, Types of chatbots: https://www.ibm.com/think/topics/chatbot-types. NIST IR 8579 initial public draft, Developing the NCCoE Chatbot: https://csrc.nist.gov/pubs/ir/8579/ipd. The NIST item is identified as a draft implementation report; booking and tour examples are original illustrations.
Previous: Chart parsing Next: Chauvenet's criterion