This week I watched two people use the same tool and get completely different results.
Same software. Same access. Same instructions from IT.
One used it to draft an email faster. The other rebuilt half a reporting process around it.
The gap between them was not skill with AI. It was skill with the problem underneath it.
Compute is electricity
Electricity on its own does nothing. It sits in the wire. Somebody has to build the appliance that turns raw power into a washing machine, a fridge, a light in a dark room.
Most people in most companies were handed the electricity and told to work out the appliance themselves.
That is the actual bottleneck right now. Access is rarely the problem. Budget is rarely the problem. Framing a real problem clearly enough that the machine can help solve it is where almost everyone gets stuck.
What the appliance actually looks like
This is the part that usually gets left out of the article, so here is mine, in detail.
I write in public. That writing has rules, most of which I learned the hard way. No client names. No deal terms. No financials. Nothing about my family. Short lines. Never claim experience I do not have.
For a while I would ask the machine for a draft and then fix it. That was faster than writing from nothing. It was not much better than that.
Then I stopped asking for drafts and wrote the rules down instead. There are twenty two of them now. Each one is dated. Almost every one exists because something went wrong once and I did not want it to go wrong twice.
That changed the output more than any model upgrade has.
The rule that cost the most to learn
Rule twenty one exists because of a single sentence.
The system drafted a comment for me that opened with “I have run venue operations for years.”
It is a plausible sentence. It fits my background at a glance. I spent eight years as operations manager of an audio visual and cinema company in the Eastern Province, so cinemas are genuinely in my history.
I have never run a venue floor. Not ticketing, not concessions, not audience flow, not multi site staffing. Anyone who has would have known inside a sentence that I had not.
I killed it, and I wrote the rule the same day. Never draft a first person experience claim that is not grounded in something I have actually done.
Here is the part worth keeping. The model was not malfunctioning. It produced a confident, fluent, well formed sentence that was false. Raw compute does that. It will do it every time, at whatever speed you are paying for.
The appliance is the thing standing between that sentence and my name on it.
Why this is the whole job now
Give someone the framing skill and the same tool suddenly does ten times the work. Withhold it and they get a faster email.
I think of the good version of this as an exosuit. It does not replace the person inside it. It makes an ordinary person capable of things they could not do alone, and once you have worn one, going back feels like a step down.
An exosuit nobody has fitted is just heavy.
Eleven years running operations taught me the part that transfers. A production company built from nothing in Sydney, then eight years as operations manager of an audio visual and cinema company in the Eastern Province. I have carried a payroll. You learn to be suspicious of tools that demo well.
The job
The job is not teaching people to use AI. Most of them can already use it. They used it this morning.
The job is teaching them to see the problem worth pointing it at, and then building the thing that holds the machine to it.
That second half is where I spend my time now. It is also the half almost nobody shows.