What is the difference between a chatbot, a workflow and an AI agent?

The three terms describe different parts of a system. A chatbot describes how a person communicates with software. A workflow describes how work moves through a sequence. An agent describes who or what chooses the next action.

Google Cloud defines AI chatbots as applications or interfaces that hold natural-language conversations and generate responses to text or voice input. A chatbot may answer questions without changing any external system. It can also sit in front of a workflow or agent.

A workflow is a sequence that moves from a trigger to an outcome. In tools such as n8n, a workflow is represented by connected nodes. Each node performs a defined operation, and the connections determine how data moves. An AI model may classify text or write a summary inside the sequence without controlling the whole sequence.

OpenAI's practical guide distinguishes agents from LLM features that do not control workflow execution. In an agent, a model manages at least part of the work: it can select a tool, assess a result, decide whether the goal is complete and hand control back when it cannot proceed safely.

Is a chatbot an AI agent?

A chatbot is not automatically an AI agent. Conversation is an interface, while agency is a control pattern. If a system receives a question and returns an answer, the system can be useful without making a plan, choosing a tool or carrying a task across several steps.

A chatbot becomes the front end of an agent when the system can act beyond the conversation. For example, a support chatbot may collect an order number. A workflow validates its format and retrieves the order. An agent then reads an unusual complaint, checks approved policies, chooses which evidence to gather and proposes a resolution. A person approves a refund before money moves.

When is an automated workflow the better choice?

Use a deterministic workflow when the business already knows the correct path. A form submission can be validated, stored and routed by explicit rules. An invoice can be generated from approved fields. A notification can be sent when a threshold is crossed. These processes benefit from consistency more than judgement.

OpenAI recommends checking whether a deterministic solution is sufficient before building an agent. Agents are more appropriate when the work requires contextual decisions, rules have become difficult to maintain or the input is largely unstructured. Adding an agent to a stable process increases the testing surface, latency and possible failure paths without necessarily improving the result.

A workflow is usually enough when:

• the trigger, inputs and expected output are known;
• important branches can be written as explicit conditions;
• the same conditions should lead to the same action;
• failure must be easy to reproduce and audit;
• an LLM is only needed for a bounded step such as extraction, classification or drafting.

When does an AI agent earn its complexity?

An agent earns its complexity when the correct next step depends on information discovered during execution. The model may need to interpret a request, select one of several data sources, compare conflicting evidence, retry a failed approach or decide that human judgement is required.

The agent should still operate inside a designed process. Tools define what it can access. Instructions define its role. Guardrails restrict risky actions. Logs preserve evidence. Exit conditions stop the run when the goal is reached, the attempt limit is exceeded or the system cannot continue safely.

An agent may be justified when:

• inputs arrive as emails, documents, images or free-form requests;
• exceptions are common and hard to express as a manageable rule tree;
• the choice of tool depends on the meaning of the case;
• the system must adapt after a tool returns missing or unexpected information;
• success can be evaluated and unsafe actions can be blocked or approved.

The diagram's text equivalent is simple: use a chatbot for natural-language interaction; use a workflow when the sequence can be predefined; use an agent only for the part that must select actions from changing context. A system may move through all three layers without giving the agent control over every step.

How can the three approaches work together?

A hybrid design separates interaction, execution and judgement. Consider a supplier-onboarding process. A chatbot answers common questions and collects missing details. A workflow validates tax fields, creates folders and routes documents. An agent reads non-standard documents, identifies inconsistencies and prepares a risk summary. A procurement manager approves the supplier before the workflow updates the system of record.

This architecture keeps routine operations visible and testable. The agent receives only the tools and data needed for the judgement step. A wrong summary cannot create a supplier because the approval and final write remain separate controls.

How do you choose the architecture step by step?

1. Define the outcome. Name the business result, the owner and the evidence that proves completion.
2. Map the current process. Separate stable steps from decisions, exceptions and human judgement.
3. Choose the interface. Add chat only when conversation improves data collection or access to the result.
4. Automate the fixed path. Keep validation, permissions and predictable integrations deterministic.
5. Bound the agentic step. Give the model the smallest useful set of tools, explicit instructions and clear exit conditions.
6. Protect consequential actions. Require human approval before external communication, deletion, financial changes or other hard-to-reverse actions.
7. Evaluate the complete process. Test normal cases, missing data, conflicting inputs, tool failures, unsafe requests and escalation.

Criteria for success

• The system completes the defined outcome, not merely a convincing response.
• Every external action has an owner, permission and audit trail.
• The agent stops or escalates when evidence is missing or risk exceeds its authority.
• Deterministic steps produce repeatable results under the same conditions.
• The hybrid system performs better on agreed tests than the simpler baseline it replaces.