TOM McATEE
← All writing

Why putting AI on a process that does not work only magnifies what is wrong with it

The rule was written about automation in 1995. It binds harder now, and almost nobody teaches the order.

The Transformation Office at my university gets a version of the same call most weeks. A manager rings up and says: I have a process that is broken, can you come and put some AI on it?

It is a reasonable thing to ask. It is also exactly backwards, and the reason why has been sitting in print for three decades.

The rule

In The Road Ahead, published in 1995, Bill Gates wrote that the first rule of any technology used in a business is that automation applied to an efficient operation magnifies the efficiency. Then the sentence that everyone forgets: “automation applied to an inefficient operation will magnify the inefficiency.”

Read that again with AI in place of automation. Nothing changes. The rule was not about a particular technology, it was about what happens when you multiply something. Multiplying a good process gives you more of a good thing. Multiplying a bad one gives you more of a bad thing, faster, at scale, and now with a confident explanation attached.

That last part is what makes AI different in degree, though not in kind. An automated bad process produces bad output quietly. An AI-enabled bad process produces bad output and a fluent justification for it. The failure gets harder to see, not easier.

The order nobody teaches

There is an older answer to this, and it does not come from technology at all. It comes from manufacturing, out of the Toyota Production System and the Lean tradition that grew from it, and it runs in three steps.

STEP 01 Stabilise

Same input, same result, twice running. If the output depends on who is doing it, you have a habit, not a process.

Mechanise here →
You get mechanised chaos. The same variation as before, produced faster and in higher volume, now wearing the appearance of rigour.
STEP 02 Standardise

Write down how it is done, and have people do it that way. The step that gets skipped, because stability feels like success.

Mechanise here →
You get mechanised inconsistency. Less variation, certainly, but one person's undocumented preference is now a system everyone lives with.
STEP 03 Mechanise

Now apply the machine. AI is not a new category here. It is the newest layer on a very old sequence.

Mechanise here →
You get the thing you were promised. The multiplication finally works in your favour, because there is something worth multiplying.
The sequence is not mine. It is decades old, out of Lean and the Toyota Production System, and it circulates in several forms. What is mine is the application to the question leaders are asking in 2026.

The uncomfortable part

Here is what leaders do not want to hear. The first two steps are the hard ones, they are unglamorous, and they cannot be delegated to a vendor. Nobody has ever been promoted for stabilising a process. Everybody wants to be the executive who brought AI in.

The good news, and it is genuinely good, is that AI is extremely useful for steps one and two. It is very good at reading a mess of documents and telling you what the actual process appears to be, as distinct from what the manual says. It is very good at spotting where variation is creeping in. It is very good at drafting the standard once you have decided what the standard should be.

So the answer to the manager ringing the Transformation Office is not no. It is: yes, and let us start at the other end. Use it to help you understand the process. Use it to help you write the standard. Then mechanise, and the multiplication works for you rather than against you.

The one question worth asking

Before any AI project in your organisation, ask a single question of the process it is aimed at: if we ran this ten times, would we get the same result ten times?

If the honest answer is no, you are not ready to mechanise. You are ready to stabilise, and that is a different project with a different plan and a much better return.

Three decades on, Gates’s rule is still doing the work. We just keep finding new technologies to prove it with.

← All writing