AI doesn't fix a bad process. It scales it

AI doesn’t fix a bad process. It scales it.

Automating a bad process with AI doesn’t improve the process. It makes the error happen faster, at greater scale and with less visibility. And that last part is the most dangerous.

The most common case

A company wants to cut customer service costs. It adopts a generative AI chatbot. It feeds the model the knowledge base it already had, months out of date.

The chatbot answers fluently and confidently. And wrongly.

The customer who got the wrong answer comes back. Now more irritated, and through another channel. Service costs go up, because more volume escalates to humans and every contact is harder.

The technology worked perfectly. The problem came before the technology.

Why AI makes visibility worse

In a bad manual process, the error has a face. An agent gives a wrong answer, someone notices, the supervisor corrects it, the knowledge base eventually gets updated. The process is slow, but the error surfaces.

In a bad automated process, the error is spread out and silent. Thousands of answers a day, each one plausible, none reviewed. The signal only shows up in aggregate, weeks later, in an indicator that rises for no obvious reason.

The automation removed exactly the people who, without knowing it, worked as quality control.

The sequence that works

Order matters more than the tool:

1. Understand the process in real depth. Not the flowchart version. What actually happens, exceptions included.
2. Find where the process fails, and why. An outdated knowledge base? An ambiguous rule? Data that never arrives?
3. Fix the process. Often this step alone solves much of the problem that motivated the AI in the first place.
4. Only then automate the improved process. With measurement defined up front, so you know whether the automation improved anything.

Whoever jumps straight to step 4 isn’t saving time. They’re postponing the bill.

AI is a multiplier

The simplest way to think about it: AI is a multiplier.

Applied to a bad process, it produces problems at scale. Applied to a good process, it produces results at scale.

It’s the same logic that applies to any automation, with one difference: generative AI produces output that looks right. A spreadsheet with a wrong formula usually gives you a strange number. A model with a wrong knowledge base gives you a convincing answer.

The question before any adoption

The question that usually opens an AI project is “how do we implement it?”.

The one that should open it is different: which specific process, one that already works well today, gets even better with automation?

If the answer doesn’t come right away, the problem is earlier in the chain. And no model, however good, will solve that for you.