Knowledge base

AI agent or regular automation?

Everything is called AI these days, and that makes choosing hard. This is the decision framework I use myself before I build anything. It tells me when an AI agent is needed, when fixed rules are enough, and why that second option is the right answer more often than the sales brochures let on.

The difference in plain language.

Regular automation works with fixed rules, as in if this, then that. An order comes in? Book the amount in the accounting with the right VAT. Someone signs up through a form? Send a confirmation and put the details in the system. The rules are set in advance, and the system does exactly the same thing every time.

An AI agent does something else. It reads, weighs and writes. It interprets an email whose content isn't known in advance, understands what a customer is actually asking, or writes a reply in plain language. Fixed rules can't do that, because you can't capture every possible sentence in a rule ahead of time.

So the difference is in the nature of the work, not in how modern it sounds. And that comes with a simple test.

Three questions per process.

Ask these three questions for each process. If the answer is yes one or more times, you need an AI agent. If it's no all three times, fixed rules are enough, and they're cheaper and more predictable.

  • Does something need to be read?Does text come in whose content isn't known in advance, like an email, a review or a question through the site? Fixed rules can search for words, but they can't understand what's written.
  • Does something need to be weighed?Does the step call for a judgment? Is this a complaint or a compliment, is it urgent or can it wait, does it belong with the bookkeeping or the scheduling? Weighing isn't if-this-then-that.
  • Does something need to be written?Does a reply have to go out in plain language, in a tone that fits the situation, such as a response to a review or an answer to a customer question? A template feels like a template.

Watch out for the trap the other way around. Many processes feel like reading work but aren't. An invoice with a fixed format, an export that's the same every week, a confirmation email. An agent is overkill there.

Two real-world examples.

Both run at Inge de Bruijn Zwembad, and they show the difference clearly.

01

Review management: an agent, with a person in charge

Every Google review has to be read (what does it say), weighed (compliment, question or complaint) and answered in a tone that fits the business. Three yeses, so an AI agent. But nothing goes out without a staff member's approval. The agent does the reading, and the person has the final say. See the review case study.

02

Register to bookkeeping: regular automation

Every transaction has an amount, a VAT rate and a category. The rules are fixed and rarely change. AI has no business here. Fixed rules do this work exactly the same way every day, auditable and without surprises. That's exactly what you want when money is involved.

Honestly, agents cost more.

An AI agent costs more to build, because besides the integration you also have to set up and monitor the quality of the reading and writing. And it costs more to run, because every run uses computing power from a language model, while a fixed rule runs almost for free. An agent is also never as predictable as a rule, so important decisions need a human in the loop.

That's why my advice is usually to start where fixed rules are enough. That's often where you get the most time savings for the least money, and it's how you build the data flows an agent will need later. The agent comes after that, only in the places where something really has to be read, weighed or written.

Not sure about a specific process? The free site scan gives you three concrete automation opportunities for your business, and the automation page explains what building costs and how it works.

Which of your processes are rule work and which are reading work?

That's exactly what I answer for each process in the free process review. What do you automate with fixed rules, where does an agent pay off, and what do you simply leave to a person?

No newsletter spam. At most one email a month, and unsubscribing is just a reply.

Walk me through one process