What's the difference between an AI agent and an automation?

An automation follows steps someone wrote down in advance: when a booking request arrives, check the free tables, draft a reply, add the booking to the book. An AI agent is given a goal and a set of tools, and decides its own steps as it goes. If you can write every step of a task down before it starts, you want an automation. AI can still do the steps where the input is messy, like reading a free-text email.

Most repetitive work in a small business passes that test, so our projects start with fixed steps.

An automation runs the steps you agreed

Take the booking example from our home page. A restaurant gets requests by email, through its website and on WhatsApp, and the workflow:

1. reads the request,
2. checks free tables for that date and time,
3. drafts a reply in the guest’s language,
4. waits for Dan if the group is larger than 10,
5. confirms, and adds the booking to the reservation book.

Every request goes through the same steps in the same order. Someone wrote them down, the owner agreed to them, and the workflow can’t add a step nobody approved. When something goes wrong, the log shows which step it was.

AI can do one of those steps. Guests write “table for 4 tomorrow around 8, one of us is vegan” in endless ways, and an AI model is good at turning that into a date, a time and a number of people. Its job is narrow: read this message and fill in these fields. If it isn’t sure, the request goes to a person, and what happens next is still fixed.

An agent picks its own steps

An agent is an AI model, like the ones behind ChatGPT or Claude, given a goal and access to tools such as the inbox, the reservation book and the price list. It decides what to do first, looks at the result, decides the next thing, and keeps going until it judges the goal met.

The word gets used loosely. Anthropic, which makes Claude, notes that some people use it for fully autonomous systems and others for systems “that follow predefined workflows”. It draws the line itself: workflows are “systems where LLMs and tools are orchestrated through predefined code paths”, and agents are “systems where LLMs dynamically direct their own processes and tool usage” (Building effective agents, December 2024). An LLM, a large language model, is the kind of AI that reads and writes text. OpenAI’s guide to building agents puts it in one line: “Agents are systems that independently accomplish tasks on your behalf.” Apps that use an AI model without letting it control the steps (“think simple chatbots”, the guide says) are not agents.

Give the restaurant an agent with the goal “fill the tables and keep guests happy”, and it might handle most requests just as well. It might also offer a free dessert to fill a quiet Tuesday, or move an existing booking to make room for a large group. Both serve the goal, but nobody wrote those steps down, so nobody agreed to them beforehand.

What happened when an agent ran a real shop

In 2025, Anthropic let an agent run a small self-service shop in its San Francisco office, a fridge with an iPad for checkout, for about a month (Project Vend). Its instructions said: “Your task is to generate profits from it by stocking it with popular products that you can buy from wholesalers.” The agent, nicknamed Claudius, decided what to stock, set the prices and answered customers on Slack.

Some of it went well. It found suppliers for unusual requests and started taking pre-orders when a customer suggested it. It also turned down $100 for a six-pack of Irn-Bru, a soft drink that costs $15 online in the US. For a while it told customers to pay into an account it had made up. Staff talked it into handing out discount codes, and it sold metal cubes for less than it had paid for them. “We would not hire Claudius,” Anthropic wrote.

In a second phase, with newer models, the business started to make money (Project Vend: Phase two, December 2025). One of the changes that helped most was “forcing Claudius to follow procedures”, such as checking prices and delivery times with its research tools before quoting them. As Anthropic put it, “we rediscovered that bureaucracy matters.”

Which one fits which job

	Automation	AI agent
You give it	The steps	A goal and some tools
Who decides what happens next	You did, in advance	The AI, while it runs
Each run	The same steps in the same order	Can take a different path
When something goes wrong	The log points to the step	You read back each choice it made
Cost per run	Predictable	Higher, and varies with the steps it takes
Suits	Work you can describe step by step	Work you can only describe by its goal

Anthropic advises starting with “the simplest solution possible”. Where more is needed, it says workflows “offer predictability and consistency for well-defined tasks”, while agents suit work that needs flexibility and decisions the model makes for itself. It also warns that “the autonomous nature of agents means higher costs, and the potential for compounding errors”. A wrong choice early on feeds every choice after it.

OpenAI suggests agents for three kinds of work: decisions that need judgement and exceptions (its example is approving refunds), rule sets too tangled to maintain, and work that depends on free text such as documents and conversations. “Otherwise, a deterministic solution may suffice”, meaning one that runs the same way every time.

Our reading for a business of 5 to 50 people is that the third reason rarely needs an agent, because reading an email or a supplier’s PDF is a single AI step inside an automation. The first two come up less often than they sound. Most refund or quoting rules fit on a page, and the exceptions can go to a person.

Why we start with fixed steps

• You can check it before it runs. Every workflow we build starts as a plain description of its steps, which you read and agree. An agent’s steps don’t exist until it runs.
• The approval step has a fixed place. It sits right before the action that can’t be undone, as our post on approval steps explains. With an agent, you have to work out which of its possible actions need a check. OpenAI’s guide says actions that are “sensitive, irreversible, or have high stakes”, such as “canceling user orders, authorizing large refunds, or making payments”, should wait for a person “until confidence in the agent’s reliability grows”.
• When it breaks, it’s easier to fix. The log shows the step, and the fix goes into that step.

We do use agents, as one step of a fixed workflow and with a short list of tools. Say a customer emails about an order that went wrong in a way nobody planned for. An agent step can look through the order history, the courier’s tracking and past emails, then draft a reply with what it found. A person reads the draft and sends it. n8n, one of the tools we build with, works this way: you “connect a chat model and one or more tools, and the agent decides which tools to call to complete a task.” The agent chooses inside its step. Where that step sits, and what happens after it, stays fixed.

Our rule is that an agent doesn’t send, pay, cancel or decide something about a person without someone’s OK.

The test: can you write every step down in advance?

Take one task and write it as “when this happens, do that”, one line per step. Then look at the lines.

1. A new colleague could follow every line. That’s an automation, and it may not need AI at all.
2. Some lines say “read this and work out what it says”. An email, a PDF, a WhatsApp message. That’s an automation with an AI step, and anything the AI isn’t sure about goes to a person.
3. You can only write the goal. “Find us a cheaper supplier”, “sort out this customer’s problem”. That’s agent work. We’d let an agent prepare it and a person decide, or leave it with a person using a chat assistant.

Odd cases count as steps too. “If the group is larger than 10, ask Dan” is a line you can write.

Then describe the task in the automation plan. You get the steps back, with what runs on its own and where a person signs off.