A station can hold every AI license it needs and still get almost nothing out of them. That is usually not a technology failure, and it is usually not a people failure either. It is a gap between two things that look the same on a status report and are not the same at all: making a tool available, and making people able to use it.
I ran a workforce assessment on AI use at a US nuclear utility. Awareness was not the problem. People knew AI was coming, because leadership had said so clearly and more than once. Most of them wanted to use it. What they did not have was anyone showing them how, and anyone checking back afterward.
The five parts, and which three cost something
The Prosci ADKAR model breaks an individual change into five parts: awareness of the need, desire to take part, knowledge of how, ability to perform, and reinforcement to make it stick.
Most utility AI programs put nearly all their effort into the first two. Communication plans, town halls, an executive video, a license count in a monthly report. That is real work and it is not wasted. It is also the two cheapest parts, because both can be done to a population rather than with one.
Knowledge, ability and reinforcement are where the effort actually sits. Knowledge means sitting with a procedure writer and showing her how to ask a model for what she needs. Ability means she does it herself on real work, badly the first few times, with someone available when it goes wrong. Reinforcement means somebody comes back in six weeks and asks whether she is still doing it.
None of those three scale the way a town hall scales. Each one takes time from people who are already fully committed. That is the whole reason they get skipped.
Why a healthy report and a failed rollout look identical
This is how a station ends up looking fine on every measure a program manager reports and still has almost nobody using the tool six months in. The measures being reported are the ones that were easy to move.
Licenses issued. Training completions. Awareness survey scores. Every one of those can go up while the number of people who used the tool on real work last week stays near zero. The same thing has happened with enterprise learning platforms for years, in and out of this industry. The seats were bought, the content was there, and almost nobody opened it. Buying a license is a procurement event. Adoption is a change in how the work gets done, and the two are not measured the same way.
The awkward part is that the reported measures are not wrong. They are accurate answers to questions nobody needed answered.
The second try is harder than the first
There is a cost to getting the order wrong that does not surface for months.
Make a tool available before anyone has been taught to use it, and a share of the workforce will try it, ask it something the way they would ask a search box, get a mediocre answer, and conclude the tool does not work. Those people are then harder to reach than the people who never tried at all. You are no longer building awareness. You are arguing with somebody’s firsthand experience.
Wanting to use it also has a shelf life. The people who are willing to try right now, while AI is new and the organization is talking about it, will not stay willing indefinitely. That window is what a license-first rollout spends without noticing.
What to do differently
Four changes, in the order I would make them.
Measure use, not access. Add one number to whatever you report today: how many people used the tool on real work in the last week, by department. If that number is not instrumented, the program has no feedback loop, and everything else here is guesswork.
Put the help where the work is. Knowledge and ability close at the point of use, not in a library someone has to remember exists. A short job aid inside the workflow, or a person on the floor for two weeks, does more than a recorded course that runs longer than the task.
Name the follow-up before you start. Pick the date, put a name on it, and give that person the authority to change the rollout based on what they find. Reinforcement with no named owner is a line in a plan.
Design the pilot around the people who do not want it. Pilots get staffed with volunteers, which is exactly why pilot results do not predict a fleet rollout. If the tool survives a skeptic doing real work, you have learned something you can use.
None of this is unfamiliar to a nuclear organization. It is what we already do for a plant modification: training, qualification, a period of increased oversight, and an effectiveness review afterward. The discipline is in the building. It just does not get pointed at software, because software arrives through procurement rather than through the modification process.
That gap is wider at utilities that run lighter programs than a nuclear station does. It is the same gap, and the answer already exists a few miles up the road in an industry that worked it out under a regulator.
Where this is heading
The NRC is preparing for AI in both licensing reviews and licensee programs, and publishes its work on the subject openly. The regulatory questions are getting clearer every year. The adoption questions are the ones each station will still be answering on its own, and the answers to those are not in the tools.
So if you are running an AI rollout right now, the useful question is not whether people are aware. It is which of the five parts is weakest at your station, and whether anyone owns it.
Talk it through
I spent 15 years inside a nuclear station running the corrective action program, performance improvement and emergency preparedness. I am an ANSI N45.2 certified lead auditor and a Prosci certified change management practitioner. The work I do now builds AI and data analytics into performance improvement programs rather than alongside them, and it is described on the services page.
If you want to walk through where your own rollout is weakest, or what to measure instead of license counts, get in touch through the contact page. Half an hour is usually enough to tell where the gap is.