Agent, assistant, automation: three words, three different things you can buy
The three terms describe three different levels of autonomy, and the difference decides who is accountable, what it costs to run, and whether the EU AI Act applies. A definition of each, with the test that tells them apart.
Three words get used interchangeably in AI sales conversations, and they describe three genuinely different products with different price tags, different failure modes and different legal treatment. Buying one while believing you bought another is the most expensive mistake available in this category.
This is a definition of each, a test that tells them apart in one question, and what changes on the invoice and in the org chart depending on which one you have.
Automation: a fixed path, no judgement
An automation executes a predetermined sequence of steps whenever a defined trigger fires. Same input, same path, same output, every time. There is no model deciding anything, and that is the point.
A form submission that creates a CRM record, assigns an owner by territory, and posts to a channel is an automation. So is a nightly job that deduplicates contacts by exact email match. The logic was written by a person, and it will keep doing exactly what that person wrote until someone changes it.
Automations are the cheapest thing in this article to build and by far the cheapest to run, because nothing is billed per token and nothing has to be evaluated for quality. They fail loudly: when the path breaks, it breaks visibly and immediately, which is a feature.
Their limit is also structural. An automation cannot handle a case its author did not anticipate. Feed it something unexpected and it either errors or does the wrong thing confidently. Most of what companies call “our AI project” is, correctly, this, and there is nothing wrong with that.
Assistant: judgement on request, no autonomy over timing
An assistant applies a model’s judgement to a task a person asks it to do, in the moment they ask. It has latitude over how it answers and none over whether or when.
The chat window is the obvious form, but the shape matters more than the interface. A button in your CRM that drafts a follow-up email from the deal record is an assistant. A model that summarises a call transcript when you click “summarise” is an assistant. In every case a person initiates, the model produces, and the person decides what happens next.
This is the category most companies actually adopt first, and its economics are the ones most often misjudged. An assistant costs per use, and the usage comes from people, so the bill scales with adoption.
95 % of enterprise AI pilots produce no measurable result.
MIT, via Forbes. An assistant nobody has a reason to open on a Tuesday is indistinguishable from no assistant at all.
Agent: decides when to act, and which steps to take
An agent decides for itself when to act and which sequence of steps to take, in pursuit of a goal, using tools it can call. Nobody presses a button.
The clearest formulation of the boundary comes from Anthropic’s engineering write-up on building effective agents, which separates two architectures that get sold under one word:
Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.
That sentence is the whole distinction. If the path is written down in advance and the model fills in the blanks, you have a workflow with a model inside it, closer to an automation than to an agent, whatever the invoice says. If the model chooses the path at runtime, you have an agent.
Agents are the most capable and the most expensive to operate, and the expense is not mostly tokens. It is supervision. Something that decides for itself needs a way to tell whether its decisions are still good, which is why an agent without an evaluation harness is not a product but a liability with a good demo.
The test, in one question
Who decides that work should happen right now?
- A rule someone wrote → automation
- A person, in the moment → assistant
- The system itself → agent
That question also sorts the accountability. With an automation, the person who wrote the rule owns the outcome. With an assistant, the person who accepted the output owns it. With an agent, the organisation owns it, and “the model chose to” is not a position anyone wants to hold in a customer escalation.
Why the distinction changes the price
The three categories fail differently, and the failure mode drives the operating cost far more than the licence does.
| Automation | Assistant | Agent | |
|---|---|---|---|
| Trigger | A rule | A person | Itself |
| Fails | Loudly, at once | In the open, one output at a time | Quietly, over weeks |
| Billed by | Runs | Usage | Usage plus supervision |
| Needs evaluation | No | Rarely | Always |
The third row is where budgets break. An automation that breaks stops. An assistant that gets worse produces one visibly weaker draft, which a person catches. An agent that gets worse keeps running: a model gets deprecated, a field gets renamed, a source changes format, and the output degrades by degrees while every dashboard stays green. The first person to notice is usually a customer.
Which is why the ongoing cost of an agent lands at roughly a quarter to 40 % of the build cost per year, and why anyone quoting only the build number is quoting half the project.
Why the words matter legally, not just commercially
The EU AI Act’s definition of an AI system turns on exactly this axis. It describes “a machine-based system that is designed to operate with varying levels of autonomy”: autonomy is in the definition itself.
A pure automation with no model in it is not an AI system at all and is outside the regulation. An assistant and an agent both are, which since 2 August 2026 means the Article 50 transparency obligations apply wherever they interact with people or generate content, regardless of risk class, and with no assessment that can end in “does not apply to us”. We separated what came into force from what was postponed, because most vendor material still has the two the wrong way round.
Being able to say which of the three you operate is therefore not pedantry. It is the first line of the classification, and the classification decides the rest (auf Deutsch).
What to do with this
Take your list of AI initiatives and put one of the three words next to each. Two things usually fall out.
Most entries turn out to be automations, and several of those do not need a model at all, a fact worth discovering before paying per token for deterministic work. And the one or two genuine agents on the list turn out to be the ones with no named owner and no way of knowing whether they still work, because they were scoped as projects rather than as systems.
Fixing that ordering is worth more than any tool decision on the page.
Frequently asked questions
Is a chatbot an agent?
Usually not. A chatbot answers when spoken to, which makes it an assistant. It becomes an agent when it decides on its own to act (opening a ticket, scheduling a call, updating a record) without waiting to be asked.
Is an “agentic workflow” an agent?
By Anthropic’s distinction, no: if the steps are laid out in code and the model fills in the blanks, it is a workflow. The word “agentic” in vendor material almost always describes this middle case, which is fine and often the right architecture: it is just not what “agent” means.
Which of the three should we start with?
Whichever matches the task. Deterministic, high-volume, well-understood work should be an automation, because it is cheaper and more reliable. Reserve agents for work where the right next step genuinely varies and a person cannot be in the loop each time.
Does an automation fall under the EU AI Act?
Not if there is no AI system in it. A rule-based sequence with no model is outside the definition in Article 3. Add a model that classifies or generates, and it is in scope.
Do we need an evaluation harness for an assistant too?
Less urgently. An assistant’s output passes a person before it does anything, so degradation is caught in the open. An agent has no such checkpoint, which is why measurement is non-optional there.
Sources: Anthropic, “Building effective agents”, AI Act Article 3, definitions, AI Act Article 50, MIT Media Lab Project NANDA via Forbes.
Strategy Call