Assistant, workflow, or AI agent — how to choose the level of autonomy

Three levels, one question

The phrase "AI in the company" hides three very different things. They differ not in technology but in who decides on the next step.

• An AI assistant waits for a command, carries it out, and hands control back. A human decides on the next move.
• An agentic workflow runs a predefined sequence of steps. The order is fixed and known in advance.
• An AI agent plans which steps to take itself, uses tools, and repeats the loop until it reaches the goal.

The question for a rollout isn't "which one is the most cutting-edge," but: how many decisions do I want to hand to the machine, and how costly is its mistake.

Assistant — the human leads

An assistant is a model invoked on demand. It summarizes a document, drafts an email, classifies a ticket. After each response, it's the human who judges the result and decides what comes next.

This level is the safest, because the decision loop never closes without a human. It works where tasks are short, the context shifts, and judging quality calls for human judgment. The downside is that it doesn't scale on its own: you get as many responses as the commands you issue.

Workflow — steps set in advance

An agentic workflow chains several model and tool calls into a fixed, designed sequence. For example: receive a ticket, classify it, extract the data, save it to a system, notify the owner. Every step is known before it runs.

A key ingredient here is tool use — the model's ability to call a specific tool (a search engine, an API, a database) at a designated point. In a workflow, it's a human who designs when and which tool gets used.

A workflow delivers repeatability and scale while staying predictable: since the path is fixed, it's easy to test and monitor. Its limit is rigidity — if a situation falls outside the designed cases, the workflow won't invent a new route.

Agent — the machine plans the step, the human sets the limits

An agent differs in one thing: it decides on the next step itself. It's given a goal, breaks it into tasks, picks tools, checks the result and — if needed — tries another way. It's a loop, not a single call.

This gives the most flexibility for tasks you can't script step by step in advance. It also carries the highest risk: an agent can head in the wrong direction for several steps before anyone notices. That's why an agent's autonomy is autonomy within the successive steps of a task, not a withdrawal of supervision.

The control stays a human in the loop: a person sets the limits on what it can do, approves sensitive operations (a payment, a dispatch, a change to customer data), and can stop the loop. An agent doesn't replace a person — it takes over the repetitive steps, while decisions about outcomes stay with the team.

A quick comparison

Criterion	Assistant	Workflow	Agent
Who picks the next step	Human	Designer (fixed path)	Model in a loop
Autonomy	Low	None (fixed path)	High, within set limits
Supervision	On every response	Path monitoring	Limits + approval of sensitive steps
Risk of a silent error	Low	Medium	Higher
Predictability	High	High	Lower
Typical use	Drafts, summaries, classification	Fixed processes: tickets, reports	Tasks with no fixed path, e.g. multi-step research

Operator's rule: start with the lowest level that solves the problem. Raise autonomy only where repeatability and scale genuinely call for it — and never at the cost of supervision.

How to choose the level

Reduce the decision to three questions, in this order:

1. How costly is a mistake? The more expensive the error (money, customer data, compliance), the closer you stay to an assistant and the more approval points you add, whatever level you pick.
2. Can the path be scripted in advance? If yes — a workflow gives you scale and predictability. If the steps depend on the situation and can't be planned ahead — that argues for an agent.
3. How often does this recur? Single, variable tasks stay with an assistant. Repeatability justifies automation.

In practice, many rollouts are a mix: an assistant where human judgment matters, a workflow for fixed processes, an agent for narrow multi-step tasks — always with a human in the loop as the control layer. The level of autonomy isn't a goal in itself; it's a tool matched to the cost of a mistake.