I have watched intelligent systems arrive four times and I have watched a market get a technology completely right and completely wrong at the same moment. This is what I took from it.
A computer only cares about what you say. It never cares what you mean. People are the exact opposite.
For most of computing history that gap was obvious. You typed the command wrong and the machine said bad command or file name and sat there, unbothered, waiting. The failure was immediate, visible, and clearly yours.
Language models closed that gap from the wrong side. They answer what you appear to mean, fluently and confidently, and they never tell you when they guessed. This is why so many capable people try these tools, get a mediocre result, and quietly decide the whole thing is overhyped. They said one thing and meant another. The system answered the thing they said. Nobody told them that is what happened.
The tool did not fail. The handoff did. Almost every practical AI skill worth teaching is a skill in closing that gap: saying what you mean precisely, grounding the request in real source material, and checking what comes back. None of it is technical. It is a communication discipline, which is why the people who get good at it are frequently not the engineers.
I have watched intelligent systems arrive four times. Collaborative computing, which promised that if we could get people into the same digital room the work would improve. Knowledge management, which promised to capture what the experts knew before they retired. Expert systems, which promised to encode judgment into rules. And now this. Three of those promised more than they delivered, and all three left something durable behind once the noise cleared.
Running through the middle of them, I watched the web arrive. I was inside a Fortune 500 company the year the Internet Tidal Wave hit, installing Mosaic on corporate machines and being asked to build an intranet before most of us could define the word. I went straight into the dot-com world that a few years later became dot-bomb for a lot of good people.
That one taught me the most, because its lesson is counterintuitive. The internet was not overhyped. If anything it was underestimated. What was overhyped was the timeline and the business models. The technology won completely and most of the companies betting on it still went under. Both of those were true at the same time, and the people who could hold only one of them made bad decisions in both directions. The believers bought everything. The skeptics missed the entire thing.
So what I look for now is the part that will still be here in ten years, and I build on that instead of on the announcement.
Nobody was ever afraid of a spreadsheet. Nobody worried that using a database made them look lazy, or replaceable, or foolish for trusting it. AI arrives carrying all of that. An individual wonders whether using it counts as cheating. A team waits for someone else to go first, because the risk is local and the credit usually is not. A leader is asked to sponsor something they cannot measure and would have to defend if it failed publicly. None of that is irrational. It is what thoughtful people do around a system that gives confident answers and shows no work, and it does not respond to a mandate. It dissolves when someone they trust shows them.
Human in the loop has become a checkbox, and in practice it produces someone clicking approve on output they are not qualified to evaluate. What holds up is a named subject matter expert with the standing to overrule the system, working from answers grounded in real, governed sources rather than in a model’s recollection. Grounding and expertise together. Either one alone is theater. This matters most where I work, because our output can end up in a due process hearing, an eligibility appeal, or an audit. A commercial AI error costs money. In the public sector it can be subpoenaed.
You cannot explain an output if you cannot explain its input. Data lineage gets treated as the boring prerequisite to the interesting work. It is not a prerequisite. It is the entire basis on which anyone will believe the answer, and in public institutions it is the difference between a decision that survives review and one that does not. Repeatability follows from the same discipline: the first implementation of anything is a project, the tenth is a product, and the acceleration everyone wants comes from that discipline rather than from the model.
If the only responsible way to use AI requires a twenty-person data team and a six-figure platform, then only well-resourced districts will ever use it well and everyone else goes without or uses it badly. Right-sizing is not an efficiency concern. It decides who gets to participate. Any framework that cannot scale down is not a framework. It is a product for large customers.
Nearly all public conversation about AI in schools is about students cheating. It is an understandable question and it is the least interesting one available.
The larger opportunity sits with the adults. School systems run on people buried in reporting: the data coordinator reconciling submissions against three incompatible systems, the special education administrator assembling compliance documentation, the counselor who cannot get a straight answer about which students are off track because the answer lives in four places.
Educators should not have to fight their data in order to help their students. I have spent twenty-five years on that sentence and I have not gotten tired of it yet.
It is with the analyst reconciling a report, the consultant drafting a deliverable, the program staffer answering the same question for the ninetieth time. If a meaningful share of an organization recovers even a few hours a week, that shows up in capacity and in margin. It also shows up in whether people stay, because the hours this gives back tend to be the worst hours of the job.
I use these tools every day in real work and recover somewhere between eight and twelve hours a week doing it. That is not a demonstration figure. It is what happens when someone becomes deliberate about how they ask, what they ground the request in, and how they check the result. I am still getting better at it. I think being a good operator is a craft worth taking seriously and worth teaching plainly, and I would rather show someone than tell them.
Every serious problem I have watched an organization have with AI was a relationship problem before it was a technology problem.