Nobody at your company is going to become an AI expert. That is not the problem you think it is.
Every rollout is sold the same way. Buy the licenses, run the training, send the announcement, and wait for the organization to become AI-native.
Then six months pass and nothing is faster.
The licenses got used. The dashboard says adoption is up. Somebody in operations is genuinely doing three people's work now. But the invoices still take the same number of days, the same reports still get built by hand every Monday, and the backlog looks exactly like it did in January.
This is not a failure of execution. It is what happens every time.
Watch any organization thirty days after it puts an AI tool in front of its people and the same shape appears.
A small handful become genuinely good at it — usually the ones who pushed for the tool in the first place. They were already using AI on weekends, for fun. They now move at a speed that makes their managers uncomfortable.
A larger group uses it occasionally and badly. They paste things in. They accept whatever comes out. They get a little value and create a little risk.
And most people never open it again after the first week.
Fig. 01 — The barbell. Illustrative of the split reported across large rollouts, not a measurement of your organization.
The tool was not the problem. The shape is not a sign of poor execution either. It appears in a fifty-person firm and in a five-thousand-person company, and it appears when the rollout is done well.
There is a gap in front of you and a chasm behind you. If you use a model most days and have opinions about how to use it, you are already far ahead of the median employee — the one who opened a chatbot twice in two years, decided it was useless, and never came back. Every AI strategy deck assumes that person catches up. They do not. Each new release raises the skill needed to use the tools well, so the chasm gets wider, not narrower.
Most writing about this problem is aimed at enterprises with five thousand employees, where at least a cohort of naturals emerges. Ten percent of a big number is still a real number of people.
Run the same arithmetic on a forty-person law firm, a family real estate office with nine people in the back office, or a carrier with a dispatch team of twelve.
In organizations this size there may be no internal AI expert at all — and if there is one, they are already the most overloaded person in the building. The strategy of "empower your power users and let it spread" has nothing to spread from.
For most organizations under a few hundred people, the human-adoption path is not the hard road. It is not a road at all.
Anyone can open a model, type four words, and watch something happen. The output may be right. It may be wrong in a way that takes three weeks to surface.
Doing it well is a craft. It means knowing what to hand a model and what to never hand a model. It means noticing that the thing you just did twice should become a written instruction the system reads every time, instead of a prompt you retype. It means knowing which part of a process needs judgment and which part is simply rules that should have been code years ago. Most of all, it means reading the work before you accept it.
Give the same task to two people and the difference shows in the first five minutes.
Pastes the request in. Takes the first answer. It looks right, so it goes out.
Nobody checks which parts came from the record and which parts the model filled in.
Three weeks later somebody finds the one sentence that was invented.
Names the files to use and the ones to leave alone. Has standing instructions the model follows on every run.
Reads the output against the source, catches one stray line, fixes it in a sentence.
Half the work, and it holds up.
Fig. 02 — Same tool, same seat, different outcome.
At least half of any organization never becomes person two. That is not a character flaw. It is a craft they did not sign up for.
The natural objection is that these are just bad rollouts. Ours would be better.
Run the arithmetic instead. In a large deployment, a small slice of seats consumes most of the usage. Now suppose the training worked and everyone used it like the top slice. Consumption does not rise by a few points. It rises by a multiple, because heavy use is what skill looks like.
The success case and the budget disaster are the same event. That is a strange thing to build a plan around.
Vendors build for the frontier user, because the frontier user is loud, flattering, and easy to impress. Every release adds capability and raises the skill needed to hold it. Nothing in that roadmap makes the median employee better.
Inside your own building the incentive is no better. Your best AI user's advantage is the gap. They finish in a third of the time. Why would they hand that away for free?
So the popular fix — teach everybody to prompt — is a small part of the job. The larger part is deciding which work should never touch a model and which work should run without a person at all. That answer is different at every company, and you only get it by watching how the work really moves.
Here is the reframe that makes the problem tractable. Stop asking how to get people to use AI. Ask which work should stop touching a human at all.
Your bookkeeper does not want a chatbot. She wants the invoices coded. Your paralegal does not want to learn prompting. He wants the deadline calendar to be right. Your property manager does not want an assistant to converse with. She wants to know which certificates of insurance expire next month, before they expire.
Nobody is in the market for a tool that helps them do the work. They want the work done.
That is a different kind of build entirely. Instead of software waiting for a human to think of using it, you get a system that already knows the operation — its files, its terms, its procedures, its deadlines — and does the repetitive work on its own, on a schedule. People stop producing the work and start approving it.
The person who was afraid of AI never has to touch it. They just find their morning easier.
A property portfolio where every lease, rider, and amendment is searchable in plain English, every renewal and insurance expiration is tracked automatically, and the arrears letters are drafted before anyone sits down.
A practice where deadlines calculate themselves, intake documents assemble themselves, and thirty years of the firm's own work is retrievable in seconds.
A prosecutor's office where statutory disclosure deadlines are computed on every case, and no attorney is ever surprised by one.
A carrier where the exceptions surface themselves instead of being discovered on Thursday.
In none of these does an employee need to become good at AI. The system knows the operation. The people make the decisions.
This does not mean training is pointless. It means training has a different purpose than you were sold.
You cannot know who your naturals are until you train everyone once. Treat that session as a diagnostic, not a cure. Then give the people who emerge a place to post what they build, so one person's method becomes everybody's default instead of a private advantage. Status is the incentive that buys back the edge they would otherwise keep.
No new window, no new habit, no prompt. The agents run on a schedule inside the systems your people already open, and the person's job becomes approving, correcting, or rejecting work that is already done. This is the path most of your organization is on, and it is the one that moves the numbers.
Fig. 03 — Two paths, sized to who is actually on them.
Stop reporting adoption. It is a yes-or-no question standing in for something you actually care about, and it will tell you everything is fine while nothing improves.
The published numbers already show the split. McKinsey's 2025 survey found that 88% of organizations use AI in at least one business function, while only 6% attribute more than 5% of earnings to it. MIT's GenAI Divide report puts it harder still: a small minority of pilots produce real value, and most show no measurable effect on the books.
Both of those can be true at once, because adoption counts logins and value comes from work that stopped being done by hand.
Report this instead: what share of the work is still fully manual, what share is assisted, and what share runs on its own with a human approving it.
That number moves only when something real changes. It cannot be inflated by a training session or a license count. And it is the only one that has ever shown up in anyone's margins.
Figures attributed above belong to their publishers: McKinsey State of AI (2025) and the MIT GenAI Divide report. We cite them as published and do not restate them as our own measurements.
The rollout split in Fig. 01 is illustrative of a widely reported pattern. It is not data from your organization, and we do not present it as a study.
We do not sell licenses and we do not run prompt workshops. We build a Company Brain: a private system that holds what your operation knows and does the work on it, inside your own environment, with your people deciding everything that matters.
Brief Us →