The vocabulary of AI products has gotten sloppy. Chatbot, assistant, copilot, agent — these words get used interchangeably, and that confusion has real consequences. Companies buy chatbots when they need agents, build agents when they need chatbots, and wonder why neither works well.
Here's the actual distinction, and when each is the right tool.
What a chatbot does
A chatbot has a conversation. You send it a message, it sends back a response. It can answer questions, explain things, generate text, summarize documents, translate languages. It's a sophisticated text processor.
The key property of a chatbot: it doesn't do anything. It produces text. What happens with that text is up to you.
This is genuinely useful. Most of the value people get from ChatGPT is chatbot value — answer a question, write a first draft, explain a concept. The interaction is: human types something, AI types something back.
What an agent does
An agent takes actions in the world. It can read your files, query your database, send emails, update records, run code, browse the web. It operates autonomously through a sequence of steps to accomplish a goal.
The key property of an agent: it has agency. It makes decisions, uses tools, observes the results, and continues until the task is done or it gets stuck.
The interaction is different: human states a goal, agent works toward it, human reviews the outcome. The agent is doing work, not generating text about work.
The practical difference
Same task, two approaches:
Chatbot approach to "summarize last week's sales performance": You copy data from your CRM into the chat. You paste in numbers from your spreadsheet. You ask for a summary. The chatbot writes a summary of what you gave it.
Agent approach to the same task: You say "summarize last week's sales performance." The agent queries your CRM directly, pulls the relevant data, cross-references it with your targets, identifies anomalies, and writes the summary. No copying. No pasting. You get the answer.
The chatbot approach is faster to build. The agent approach is faster to use.
When chatbots are the right choice
Chatbots are underrated for their actual use case: answering questions from structured knowledge.
Customer support — if 80% of your support volume is "how do I do X" questions that have definitive answers, a chatbot on your documentation is the right tool. No agency needed.
Internal knowledge bases — employees asking HR policy questions, IT asking how to configure something, sales asking about product features. Chatbot with good context retrieval handles this well.
Writing assistance — drafting emails, summarizing documents, reformatting content. The human provides the material; the model shapes it.
Exploration and research — when a human is doing the research and wants AI to help think through it. Collaborative, conversational.
Use a chatbot when: the value is in the generated text itself, the task doesn't require persistent action in external systems, and a human will review and act on the output.
When you need an agent
Agents are the right choice when the task involves a sequence of actions that depend on intermediate results.
Automated data pipelines — agent monitors a source, transforms data when new records appear, loads it to a destination. Runs without human intervention.
Research and synthesis — agent searches multiple sources, reads pages, extracts relevant information, and produces a synthesized report. This involves dozens of tool calls, not one.
Workflow automation — agent monitors a trigger (new deal in CRM, new issue in GitHub, new row in spreadsheet) and takes a sequence of actions in response.
Monitoring and alerting — agent checks conditions on a schedule, decides if something needs attention, takes appropriate action or notifies the right person.
Use an agent when: the task requires actions in external systems, the path to completion isn't known in advance, and the work should proceed without human input at each step.
The hybrid middle
Most real business applications are somewhere in the middle. An agent that does most of the work, with a human review step before consequential actions. This is the right default for anything that touches customer-facing data or financial records.
Build the agent to do the work. Put the human in the loop for approval. Automate approval itself once you trust the agent's judgment.
The progression is: chatbot → supervised agent → autonomous agent. Most teams try to jump to autonomous too fast and lose trust in the system when it makes mistakes. Earn trust incrementally.
The one thing to get right
Chatbots should never pretend to have taken actions. Agents should never pretend to be merely answering questions.
The confusion between these two modes is where things go wrong. A chatbot that says "I've sent the email" when it hasn't taken any action destroys trust. An agent that says "here's what you could do" when it should have done it wastes the entire point.
Know which one you're building. Make it behave accordingly.
Published under Field Notes. Not for sale. Share freely.