Agent, Assistant, Chatbot, Automation: The Words People Mix Up

The vocabulary in this field is deliberately loose. Vendors benefit from the ambiguity, and buyers pay for it — by purchasing one category while expecting another.

The five categories

Automation

Fixed steps, fixed order, triggered by an event. When an invoice arrives, extract these fields, put them there, notify that person.

For a vendor-side perspective, Monitask also has a page on workforce optimization.

Deterministic. Same input, same output, every time.

No model involved necessarily, and increasingly one is bolted on for a step like classification.

When it is the right answer: the process is stable, the steps are known, and you want predictability. Which is most business processes. See when an agent is the wrong answer.

Chatbot

Answers questions, usually from a defined knowledge base.

No actions in real systems, or a very short list of them.

Modern ones use a model and retrieval, which makes them substantially better than the scripted ones of a decade ago and does not make them agents.

When it is right: high-volume repetitive questions with answers that exist somewhere.

Assistant

Helps a person do a task, with the person driving.

Drafts, summarises, suggests, explains. The output goes to a human who decides what to do with it.

No autonomy — it does not act unless asked, each time.

When it is right: individual productivity, where judgement stays with the person. This is where most people's actual daily use sits.

Agent

Goal-directed, tool-using, iterative, with bounded autonomy.

Takes actions, observes results, adjusts. Handles cases nobody scripted.

When it is right: the task varies enough that a fixed script fails, and there is a real tolerance for occasional wrong answers. See what an agent actually is.

Multi-agent system

Several agents with different roles, coordinating.

Heavily discussed, much less often working in production. Each handoff loses context, error compounding applies across agents, and debugging is substantially harder.

Treat claims here with particular scepticism and ask what is running rather than what is architected.

The words that signal nothing

"Agentic." Now attached to products that have not changed. Ask what it does when the first attempt fails — if the answer is "returns an error", it is not agentic.

"AI-powered." Could mean anything from a language model to a rules engine written in 2014.

"Autonomous." Almost never means what it says. Ask specifically which decisions it makes without a person.

"Intelligent."

"Copilot." A branding convention, not a category. Usually an assistant.

"Orchestration." Real term, and it describes a technical arrangement rather than a capability level.

How to tell which one you are being sold

Four questions, in a demonstration or a sales call.

What can it actually reach? Which systems, with what permissions. If the answer is vague, the integration work is yours. See integration is the hard part.

What happens when the first attempt fails? Automation stops. A chatbot apologises. An agent tries something else. This single question separates the categories faster than any specification.

Which decisions does it make without a person? The honest answer describes the autonomy level. See human oversight.

What does it log? Agents that take actions need action logs. If the vendor talks only about conversation history, actions are probably not being recorded. See logging and audit trails.

Why buying the wrong category costs money

Buying an agent when automation would do: you pay per call for a loop, accept non-determinism, and inherit governance and security work — for a process that was stable and scriptable. This is the commonest expensive mistake.

Buying automation when the task genuinely varies: it breaks on every case the script did not anticipate, and someone handles the exceptions manually forever.

Buying an agent when you needed an assistant: you built autonomy into something where a person was going to review every output anyway. For broader independent background, see NIST AI Risk Management Framework.

Buying a chatbot and expecting actions: it answers questions about the thing instead of doing it.

The useful mental test

Ask what the thing does with a situation nobody anticipated.

Automation fails. Predictably, and you can catch it.

A chatbot says it does not know.

An assistant produces something and waits for you.

An agent tries. Sometimes well, sometimes not, and it tries.

That difference is the whole category boundary, and it is also the source of both the value and the risk.

A note on how the words will change

This vocabulary is unstable and will shift. "Agent" already covers a wider range than it did eighteen months ago, and the drift is toward inclusion — more things get called agents over time, not fewer.

The underlying categories are more stable than the labels. Fixed script, question answering, human-driven help, autonomous action with tools. Those distinctions will outlast the current words.

Check the date on this article, and when the terms have moved, use the four questions rather than the labels.

The short version

Five categories: automation, chatbot, assistant, agent, multi-agent. Sold under one label, with different costs and risks.

"What happens when the first attempt fails" separates them faster than any specification.

"Agentic" attached to an existing product usually means nothing.

Buying an agent where automation would do is the commonest expensive mistake — you pay for a loop and inherit governance work for a process that was stable.

And the labels will drift. The four questions will not.