Agents and automationLesson 1/8

Lesson 1 of 8 · 15 min

Assistant or agent?

The four levels of autonomy, and what changes when AI can act.

So far, you’ve mostly used AI as an assistant: you ask a question, it answers, and you’re the one who acts. Since 2025, the tools have taken a step further: they can read your email, fill in a spreadsheet, create an appointment, send a message. AI no longer just advises: it acts. That saves a great deal of time, and it also changes the nature of the risk. This lesson gives you the lay of the land before you build anything.

By the end of this lesson, you’ll be able to:

  • tell apart the four levels of autonomy, from assistant to agent;
  • choose, for each task, the lowest level that gets the job done;
  • spot irreversible actions, the ones that always require your approval.

The four levels of autonomy

Level What the AI does Who acts? Example
1. The assistant Answers based on what you give it You You paste an email, it suggests a reply
2. The connected assistant Reads your tools and prepares the work You, after review It goes through your inbox and writes drafts
3. The automation Runs a series of steps set in advance, triggered by an event The machine, following your recipe Every new form is sorted by the AI and filed in a spreadsheet
4. The agent Gets a goal and chooses the steps and tools itself The machine, using its own judgment “Prepare my weekly report”: it looks up the data, calculates, writes

The key difference between levels 3 and 4: in an automation, you drew the path; the AI only handles one step (sorting, summarizing, extracting). An agent draws its own path. It’s more flexible and adapts to the unexpected, but it’s also less predictable.

More autonomy, more gain, more risk

Each level saves more time than the one before, because you step in less. But each level also removes a check: at level 1, nothing happens without you; at level 4, the agent can chain ten actions before you look.

The risk isn’t theoretical. In June 2025, researchers disclosed a flaw called EchoLeak in Microsoft 365 Copilot: a single email containing hidden instructions could lead the assistant, while reading the inbox, to leak internal data, without the victim clicking anything. Microsoft fixed the flaw, but the principle remains: a tool that reads your data and can act can also be manipulated by what it reads. Lesson 7 is entirely devoted to this.

Hence the rule behind this whole course:

The questions to ask for each task

To decide on the right level, ask four questions:

  1. Does the task come up often? A ten-minute monthly task isn’t worth a three-hour automation.
  2. Does it always follow the same path? If so, an automation (level 3) is more reliable than an agent. If every case is different, an agent becomes worthwhile.
  3. What does a mistake cost? A wrong row in a spreadsheet takes ten seconds to fix. A wrong reply to a client, a payment or a deletion doesn’t.
  4. Is the action reversible? That’s the deciding question.

Irreversible actions

Some actions can’t be undone: sending a message, paying, deleting, publishing, accepting a contract, sharing a file outside the company. Once done, they’re done.

The rule is simple: an irreversible action always goes through a human approval. The agent prepares the email, you click “Send.” The automation fills in the order form, you confirm it. Good tools all allow this: a draft instead of a send, an approval step, a confirmation request. You’ll learn to set them up in lesson 6.

The Atelier Boréal case

We’re back with Julie, founder of Atelier Boréal, the Montreal ceramics studio from the course for business owners (a fictional business created for our courses). Her template replies and knowledge base are in place. Now she wants some tasks to happen without her having to start them.

She ran her tasks through the four questions:

Task Frequency Always the same path? Cost of a mistake Level chosen
Filing workshop sign-ups received through a form Every day Yes Low 3: automation
Answering frequent questions Every day Almost Medium: affects the client 2: reviewed drafts
Preparing the Monday update (sales, sign-ups, pending orders) Every week No, the data varies Low: it’s for her 4: scheduled agent
Following up with a restaurant on an unsigned quote Every week No High: business relationship 2: draft, sent by Julie
Refunding a cancelled workshop Rare No High: money 1: Julie alone

Notice that Julie handed no sending and no payment to the machine. Yet she saves most of the time: the daily sorting and the report preparation. That’s the typical profile of a successful automation.

Common mistakes

  • Starting with the tool. You discover an impressive agent, then look for something it could do. Start from your tasks, like Julie.
  • Aiming too high. An agent for a task that always follows the same path is less reliable and more expensive than a simple automation.
  • Automating a poorly defined task. If you can’t explain the task to an intern, the machine won’t do it better.
  • Forgetting the irreversible. An automatic send “because it’s simpler” is the source of most incidents.

Key takeaways

  • Four levels: the assistant, the connected assistant, the automation (your recipe) and the agent (its own path).
  • More autonomy means more gain, but also less control: choose the lowest level that gets the job done.
  • Four questions per task: frequency, fixed path or not, cost of a mistake, reversibility.
  • Sending, paying, deleting, publishing, sharing: always a human approval.

Quiz

1. What is the key difference between an automation and an agent?
2. A task comes up every day and always follows exactly the same path. Which level should you choose?
3. Which of these actions should always go through a human approval?
4. What did the EchoLeak flaw disclosed in June 2025 show?
5. Why did Julie give the Monday update to an agent rather than an automation?
Next