Automating a broken process makes it fail faster
The approval that now takes four seconds
A purchase approval used to take two days. It went through three people, two of whom added nothing, and one of whom occasionally caught something expensive.
It has been automated. It now takes four seconds.
The two who added nothing have been removed from the workflow, which is a genuine gain. So has the one who caught things, because the rule she applied was never written down anywhere — she just knew which suppliers to look at twice. Nobody removed her deliberately. She was invisible to the process map, so she was invisible to the automation.
The organisation has bought speed and sold judgement, and the trade will not show up for months.
Automation is an amplifier
This is the whole principle, and it is older than any of the current technology.
Point automation at a process that works and you get more of what works — faster, more consistently, at lower cost. Point it at a process that does not work and you get more of that instead: the same errors, at volume, produced with a confidence and a consistency that makes them harder to spot than when a person was making them unevenly.
A human doing something wrong is usually doing it wrong in a slightly different way each time. That variation is a signal, and people notice signals. A system doing it wrong does it identically every time, which reads as reliability.
The three questions, before anything is automated
- Does this process produce the right outcome today, without automation? If the answer is mostly, find out what the exceptions are before you encode the rule. The exceptions are the process.
- Who currently exercises judgement here, and what do they actually know? This is the question that gets skipped, because judgement rarely appears on a process map. Ask the person, not the document. What they tell you in five minutes is usually the part that matters most and is written down nowhere.
- What happens when it is wrong, and how would we find out? A manual process fails visibly and locally. An automated one fails quietly and everywhere. If the detection mechanism is someone will probably notice, that was adequate at manual speed and is not adequate at machine speed.
What to fix first
The instinct is to automate the process that causes the most pain. Usually that is the wrong starting point, because pain is often a symptom of a process that was never designed — and automating an undesigned process just makes the design decisions permanent and invisible.
Better order: fix, stabilise, then automate. Fixing means the exceptions are understood and the judgement is written down. Stabilising means it produces the right outcome consistently for long enough that you trust it. Only then is there something worth making fast.
That sequence is unpopular because it delays the visible part of the project. It is also the difference between automation that compounds and automation that has to be unwound.
The part nobody enjoys hearing
If a process cannot be described clearly enough to automate, that is not a reason to bring in better tooling. It is a finding about the process.
The clarity that automation demands — what triggers this, what decides it, what happens at the edges — is clarity the organisation should have had anyway. Many automation projects deliver most of their value in the analysis, before a line of code is written, and then attribute it to the software.
One thing to do this quarter
Take the process you are most keen to automate. Find the person who has done it longest and ask a single question: what do you check that nobody asked you to check?
Whatever they say is the part your automation is about to delete. Write it down before you build anything.
CP-003 · v1 · Published 21 October 2026