AI won't fix a broken process. Fix the process first.

Every week I talk to companies that want to automate a process with AI. They have picked the right candidate: it is repetitive, it is time-consuming, and the team is tired of it. But when I ask them to walk me through the process step by step, the conversation quickly gets uncomfortable. “Well, normally it goes like this, but sometimes Kris does it differently.” Or: “That step isn’t really documented anywhere, we just know.” Or the classic: “The spreadsheet is the process.”

These are not AI problems. These are process problems. And putting AI on top of a broken process does not fix anything. It just makes the mess run faster.

Faster garbage is still garbage

AI is extremely good at following patterns. Give it a clean, well-defined workflow, and it will execute it reliably and quickly. Give it a workflow full of exceptions, undocumented workarounds, and steps that depend on one person’s memory, and it will automate all of that chaos with impressive efficiency. The wrong version of the form gets processed instantly. The exception that Jan always catches by eye slips through every time. The workaround that was supposed to be temporary three years ago becomes the permanent, automated default.

The result is not a more efficient operation. It is an operation that breaks in new, harder-to-debug ways, because the errors now happen at machine speed instead of human speed.

Five signs your process is not ready

Before you hand any workflow to AI, run through these five questions. If more than two apply, the process itself needs work first.

1. Different people do it differently. Not slight style differences, but meaningfully different steps or shortcuts. If three people handle the same process in three different ways, there is no single process to automate.

2. Exceptions outnumber the rule. The “standard” path covers maybe 40 percent of cases. Everything else is a judgment call, a workaround, or a phone call to someone who knows. AI needs a dominant pattern. If the pattern is “it depends,” you are not ready.

3. One person is the process. If it lives in someone’s head and they are the only one who can do it correctly, you do not have a process. You have a dependency. Automating it means encoding one person’s habits, including the ones that should have been questioned years ago.

4. The inputs are inconsistent. Orders arrive in five different formats. The “same” data field is filled in differently depending on who enters it. Before AI can process inputs reliably, the inputs need a minimum level of structure.

5. No one can draw it on a whiteboard. If you ask the team to sketch the process and they cannot agree on what happens after step three, that disagreement will not resolve itself inside a piece of software.

The cleanup does not have to take months

Here is the good news: fixing a process before automation is not a six-month consulting project. For a typical SME workflow, it takes days, not months. A practical approach:

Map it together. Get the two or three people who actually do the work into a room for an hour. Have them walk through the process from trigger to output. Write down every step, including the workarounds. Do not clean it up yet. Just capture reality.

Find the forks. Where does the process split into different paths depending on who is doing it? Those forks are decisions that were never made. Make them now: which path is correct? Document the answer.

Kill the workarounds. For each workaround, ask: does the root cause still exist? Often it does not. The workaround was created for a system limitation that was fixed two years ago, or a client requirement that no longer applies. Remove what you can. Formalize what you cannot.

Standardize the inputs. If orders or requests come in five formats, reduce it to two. If a form field is ambiguous, rename it. Every hour spent here saves ten hours of AI debugging later.

Write it down. One page, not twenty. A simple numbered list of steps that anyone on the team can follow without calling a colleague. If the process cannot fit on one page, it is probably two processes pretending to be one.

What you gain before AI even starts

What I find surprising about this step, and what makes it worth the effort even if you never automate at all: the cleanup itself already saves time. A company I worked with recently discovered that their invoice handling had 14 steps, four of which existed purely because two departments did not trust each other’s data. Removing those four steps saved roughly 6 hours a week, before any technology was involved.

Another team found that their order intake had three different email templates for the same product category, created by three different people over the years. Standardizing to one template cut handling time by a third, because nobody had to figure out which format they were looking at anymore.

Process cleanup is not overhead before the real work. It is often the most valuable part of the entire automation project.

Then, and only then, bring in AI

Once the process is clean, documented, and agreed upon, AI becomes dramatically easier to implement. The inputs are consistent, the logic is explicit, and the team can tell immediately whether the AI is doing it right or wrong, because they all agree on what “right” looks like.

This is exactly the sequence we follow in every AI-waardescan. Before we talk about tools or models, we map the process with the people who run it, identify the mess, and clean it up. That diagnostic is step one for a reason: it determines whether automation will create value or just create speed. If you suspect that some of your workflows need a cleanup before AI can touch them, that is a perfectly valid reason to start the conversation. Knowing what to fix is the first step toward building something that actually works.