Do I Need an AI Agent, or Is Plain Automation Enough?

People often ask me: "Mariela, do I need an AI agent?" I don't answer right away. I ask which task is bothering them most, and what exactly happens in it from start to finish. Most questions about agents are really questions about one specific, annoying task.

The dishwasher and the cook

Picture a kitchen. The dishwasher runs the same programme every time. It is cheap, it never gets tired, and it never tastes anything. The cook opens the fridge, sees what is there today, and decides what to make. That takes time and costs more.

Nobody hires a cook to wash plates. And nobody expects the dishwasher to plan the menu. A good kitchen has both, and the owner decides which job goes to which.

AI works the same way:

• Automation is the dishwasher. It follows set steps and runs them identically every time.
• An AI agent is the cook. It decides what to do as it goes, depending on the case.

Anthropic (the company that makes the Claude assistant) describes it much the same way: in "workflows" the program runs along predefined code paths, while agents direct their own process and their use of tools. It also notes that agentic systems often trade latency and cost for better task performance. The cook really is slower and more expensive.

What your real options are

When something in your business eats up hours, you have four paths:

1. Keep doing it by hand. Cheap at first, expensive over time.
2. Hire another person. It works, but that person will repeat the same steps too.
3. Buy a ready-made product with "AI" on the box. It is fast, but it rarely fits your process, and you often pay for judgement where you only needed repetition.
4. Ask for an agent for everything. It sounds modern, but it is like hiring a cook to wash plates.

The fifth path is usually the right one: split the process and give each step the right worker.

What it looks like in our work

We build in n8n (a tool that connects programs and runs steps by rules). Here are three cases, without names.

The morning email for brokers. A real-estate agency wanted its brokers to get an email every morning with new clients, clients who needed a call back, and questions for meetings. My first version used two AI agents: one worked out the strategy for each client, the other built the email. The result was about 90 minutes for one morning's run. Then I saw that the second agent was only pouring text into a ready template. That is dishwasher work. I replaced it with plain code, and the first agent stayed because that step needs judgement. The run dropped to about 18 minutes. The full story, with the technical details, is in "When AI Agents Cost More Than They're Worth".

The agency that was losing clients. At a real-estate agency that had plenty of enquiries but was losing some of them, the answer was the same. AI is used only where judgement is needed: a short picture of the client, a strategy for the conversation, the questions for the next contact, a one or two sentence note in the morning email, and the monthly analysis of results. Everything else runs by rules: the check for duplicate clients is a plain script, and a submitted form is turned into a comment with a fixed layout. One of the rules I give other agencies is: "Automate mechanical tasks, not sales." The whole case is in the article about the agency.

The freight forwarder. There, AI sorts incoming customer tickets: what the person is writing about and how urgent it is. Once a ticket is sorted, something else decides what to do: simple rules, written in advance. And words like "urgent" or "legal" always go to a person, whatever the AI thinks. The cook and the dishwasher work on the same line. The details are in the freight forwarder article.

Why I start with the process, not the tool

In none of these cases did I start with "agent or automation?" I started by asking what happens, step by step. Only once I can see the whole process can I say which part is a dishwasher and which is a cook. That is why analysis is my first piece of work: Process X-Ray shows your process as it really is, and Start here is the way in. Then we choose the right mix. What business process automation actually means is in "What Is Business Process Automation, and Where Do I Start?", and why a process is described before anything is built is in "What is “PDD” and Why It’s Critical When Building Automations?" (a document that describes the process step by step before it is automated).

There is also the question of money. Automation repeats the same steps, and every run costs about the same. An agent thinks on every run, and each judgement is the work of an AI model billed by volume. So an agent is slower and more expensive on every run. That is why, for every step, I ask not only "could an agent do this?" but "is it worth paying for judgement here every single day?" In the morning email, the answer for one agent was yes and for the other no.

The risk runs both ways. Give an agent dishwasher work and you pay more, wait longer, and get a result that is slightly different every time. Give automation a cook's work and it will follow its rules and fail on every unusual case.

What to do tomorrow

1. Pick one task that annoys you and break it into steps. Next to each step write "same every time" or "needs thinking." The first list is often much longer than the second.
2. Ask anyone who offers you an agent: "Which exact step here needs a decision, and what will it cost me per month to run?" If the answer is vague, you probably need automation.
3. Ask for the process to be analysed before you get a price for the tool. The price of a tool you don't yet know you need is not a price.

If you want us to look at your task together, that is where my analysis starts. Start here.