AI automation: what pays off and what does not

I searched for “AI automation” and looked at the top results. They come from Microsoft, IBM, AWS and ABBYY, plus a couple of “20 best tools” listicles.

Below them, Google shows what people actually ask. One of the four questions is how much it costs. None of the top eight answers it, which figures: they all sell the platform.

So here it is, with the full sum.

The cost you see and the two you do not

When you ask for a quote to automate something, the number you get is usually the tool. It is the smallest of the three.

The first is understanding the process. It is almost never documented. Somebody does it from memory, with exceptions nobody has ever written down, and finding those takes longer than coding them. This phase gets underestimated every single time.

The second is connecting it to your data. An automation that cannot check your stock, your CRM or your invoicing is nicely written text. And that is where the usual surprise appears: the data exists, but incomplete, duplicated, or in a format nobody expected anyone to read.

The third is maintaining it. Every time you change a price, a supplier or a returns policy, the automation needs revisiting. It is not furniture: it lives alongside your operation.

The sum I do before taking a project

Three numbers, and you can estimate them in an afternoon.

How many times a month it happens. If it is five, forget it. If it is five hundred, we keep going.

How long each one takes. Multiply by the above and you have the hours a month it eats today. That is the figure everything else gets measured against.

How alike they are. This is the one that decides. Five hundred times a month with five hundred different variations is not automatable; five hundred times with five variations is.

In plain numbers: a team spending twenty hours a month repeating the same answer has a clear case. One spending three does not — and the project will cost more than it saves for a long time.

What genuinely pays off

Repeated tasks with a pattern and reliable data. Specifically: answering frequent questions from information that already exists, classifying and routing what comes in, extracting data from documents with a stable format, drafting from what is already in a system.

They share one thing: there is a checkable right answer. When that exists, the automation is measurable and correctable.

What does not pay off, however much it looks like it

Anything that changes every couple of months. Automating an unstable process is building on sand: you will spend more time adjusting it than you would doing it by hand.

Anything that depends on judgement. Negotiating, granting an exception, deciding whether this client gets a discount. You can lay the ground so a person decides faster, but the decision does not get delegated.

Anything that is a symptom of something else. The most expensive mistake of the three. If half your enquiries ask where an order is, the problem is not support: it is order tracking. Automating the reply makes the complaint disappear and leaves the problem untouched for another year.

Where to start

With a single task, the most repeated one, and see it through. One automation that works properly is worth more than five half-finished, because the working one earns trust for the next and the half-finished ones create supervision work.

Measure before you start. If you do not know how many hours it consumes today, you will not be able to prove it helped — and that is what whoever signs the invoice will ask.

And define what happens when it fails, because it will. A system that admits it cannot handle a case and hands it to a person is infinitely better than one that answers anything with confidence. I cover that in AI agents: what you learn once they start selling.

The uncomfortable part

Automation does not cut headcount straight away, and anyone selling it that way is usually selling smoke. What it does is take repeated work off people so the same team handles more volume, or spends the time on what actually needs judgement.

When the goal is framed as saving on staff, these projects get measured wrongly and disappoint. When it is framed as capacity — handling three times the volume without tripling the team — the sum works and it holds.

I post working automations on @WhatsMarketing_es.

If you want me to look at which of your processes is worth automating, get in touch. I work from Buenos Aires, originally from Málaga, with clients in Mexico City, Argentina, the rest of Latin America and Spain.