AI agent vs chatbot, copilot and RPA: what is the difference?

A chatbot is the oldest of the three: scripted answers, or a model answering from a knowledge base, with at most a fixed hand-off to a person. It is predictable because it does nothing. A copilot adds a model that works beside a person inside the tool they already use: Microsoft 365 Copilot drafts an email or answers from the files the person can already open, and the person reads, edits and sends. Predictability stays high because every action is still a human action.

An agent moves the decision into the model. OpenAI’s definition is "systems that independently accomplish tasks on your behalf", and its exclusion is just as useful: applications that integrate a model "but don’t use them to control workflow execution" are not agents. Microsoft’s, on the Entra Agent ID page, is an application that "attempts to achieve a goal by understanding its environment / context, making decisions and acting on them autonomously using available tools". Both turn on the model picking the next step and the software carrying it out.

Microsoft drew the line between its two words in October 2024: "Copilot is how you’ll interact with these agents", with agents ranging "from simple prompt-and-response to fully autonomous". In practice, Microsoft 365 Copilot is the assistant people talk to, Agent Builder inside it creates declarative agents from SharePoint and connector knowledge with no external actions, and Copilot Studio is the usual place an agent gets actions: connectors, MCP servers, agent flows, computer use and other agents; developers can also add API and MCP plugins to declarative agents. Read-only plugin calls in Copilot skip confirmation after the first consent; calls that modify data ask the user each time.

RPA automates a user interface or an API with a recorded script. Workflow automation does the same with a chart of steps and connectors. Both fail the same way, by breaking or routing to a person when the input does not match what the script expects. Copilot Studio’s documentation makes the contract explicit for agent flows: they "are deterministic", and "the same input always produces the same output". That property is the reason to keep them: an auditor can replay a flow, and a finance controller can sign off on one.

Agentic process automation is the industry’s name for the combination. The agent sits in front: it reads the email or the document, decides what the case is, calls the flow or the bot for that case, and handles the exception itself or hands it to a person. UiPath launched its platform for this on 30 Apr 2025, designed to "unify AI agents, robots, and people on a single intelligent system", with UiPath Maestro as the orchestration layer. Copilot Studio reached the same place from the other side: its standard harness runs deterministic agent flows, and its GitHub Copilot harness runs goal-driven agents that call those flows as tools.

Computer use is the part of RPA that changed most. Instead of a recorded script, a model looks at a screenshot, decides where to click or what to type, and the harness executes it in a browser or a virtual machine. It can handle interface changes a script would break on, at a cost: each screenshot is roughly 1,000 to 1,800 input tokens on Claude, Copilot Studio bills per step, and the vendors warn about prompt injection through whatever appears on screen. Copilot Studio computer use went generally available in May 2026, Claude’s left beta on 19 Aug 2026, and Gemini’s followed on 30 Oct 2026. Use it for systems with no API, behind an allow list, with a person confirming consequential actions.

The confusion is understandable because every agent contains a generative model. The difference is what the model’s output is used for. In generative AI the output is the product, an email, an image, a summary or a block of code, and a person decides what to do with it. In agentic AI the output is a decision, which tool to call with which arguments, and the software acts on it. The same model can do either; the loop around it is what changes.

A model on its own is generative AI; it becomes an agent only with instructions, tools, memory and a loop around it. In its chat form, ChatGPT is a model wrapped in a chat interface with a person in charge of the actions. OpenAI’s agent platform moved fast in 2026: the Assistants API shut down on 26 Aug 2026, the visual Agent Builder shuts down on 30 Nov 2026, and the Agents API with its managed harness entered public beta on 10 Sep 2026.

Vendors blur these words, and Gartner’s analysts, as reported by Reuters on 25 Jun 2025, estimated that only about 130 of the thousands of vendors claiming agentic AI were real, and predicted that more than 40% of agentic AI projects would be cancelled by the end of 2027 over escalating costs and unclear business value. The word they used for the rest is agent washing.