Automation and agents: who lays the tracks and who drives

If you already have the data and know what to do with it, you don’t need an agent. You need a well-built automation.

I say this because we are in the middle of an AI agent boom and almost nobody is making that distinction. Plenty of companies say “we have AI” when what they have is a ChatGPT subscription the team uses to ask questions. They haven’t automated a single task, nobody works faster, and their data and processes are exactly where they were.

Automation only works if the data already exists

Make or Zapier do one thing very well: pick up an event in A, map the fields and fill them in B. Connecting a form to the CRM, for example. You can add an AI step in the middle to reason or classify, but the underlying logic stays the same: if A happens, do B.

That’s what sits behind many “AI strategies”: a workflow that sends a birthday coupon at 9:00 sharp. It’s a very punctual flowchart, and that’s fine. What’s out of place is calling it AI.

That logic has a condition people often forget. You need a starting data point. Inside the flow you can search, cross-reference or modify information, but always starting from something you already have.

Its virtue is that it never improvises: it does the same thing, well, every time, and it costs far less than an agent. That’s also its nightmare. Doing the wrong thing perfectly, ten thousand times.

An agent earns its place when someone has to go and look

An agent comes in when the work is research: reading several sources, connecting tools to each other (today, through MCP servers), understanding the context and writing something that didn’t exist before. Reports, presentations, personalised messages, or the follow-up with a customer after a conversation in which they explained their problem, which is never the same as the last one.

That’s where automation falls short, especially with personalisation. “Hi {first name}, I see you work at {company}” no longer counts as personalisation. Everyone has received it so many times that they read it for what it is: a template. Automation personalises with the contact’s name. The agent reads the situation.

Well prepared, with clear rules, the right sources, your tone and the skills it needs, an agent can do the work of a junior. Sometimes that of someone a bit more senior.

The problem starts when it isn’t prepared. If you let it try things out, or force it to finish a task without giving it all the information, it gets creative in the worst way: it invents data, puts words in your mouth you would never say, or reaches for the most worn-out cliché. The “shall we grab a coffee in Madrid?” that lands in so many inboxes comes from there, from the most common thing it has seen. Its nightmare is doing something brilliant that nobody approved.

With automation you’re the architect. With agents you’re the boss.

Before asking for “AI in marketing”, put the house in order

If a CEO tells me on Monday “I want AI in marketing”, the first thing I say is that, put like that, it’s too abstract to start with. Using AI for social media, for press, for the website or for sales presentations have nothing in common. Each task needs something different, and the answer depends on what you want to achieve and how much you want to automate.

My rule for deciding is simple. If I have all the data sources and only need to set the rules and map the fields, I use an automation: it’s cheaper and it simply follows the rules of the game. If the work is creative, draws on several sources and requires research, I bring in an agent.

That rule has a cost the demos don’t show. An agent needs preparation before it does anything useful, and if you’re not going to do that work, you’re better off sticking with automation.

If I were deciding today, I’d do three things:

• Split the tasks. The ones that fit an automation on one side; the ones that need research or judgement on the other.
• Get the foundations in order before the first agent. Review how the data is structured and write the brand voice and brand book, so what it produces sounds like the company and not like the internet average.
• Put a person in charge. An agent is one more employee and, like any employee, it needs a manager above it who reviews what it delivers. You can’t take everything it produces as good.

Paying for a ChatGPT subscription doesn’t make you a company with AI. It makes you a company with a chat.