Every company has the same story. The new software went in. Training happened. A recording went out. Three weeks later the same questions are being asked at the same desks, and the person who actually knows the answer is the one sitting nearest.
This is not a discipline problem. It is not that people didn't pay attention. It is that the way most companies teach software is out of step with how people learn anything, and the bill for that gap is bigger than anyone puts on a slide.
Learning happens in the hands, not the ears
In 2014 a team at the University of Washington pooled 225 studies comparing two ways of teaching the same material: a lecture, or something that made the student do the work during the session. The result was not close. Students in lectures were one and a half times more likely to fail. Exam scores were about six percent higher when people had to act instead of listen. The effect held across every subject and every class size.
That was undergraduates and chemistry. But there is no reason to think an account manager learning an invoicing tool is wired any differently from a nineteen-year-old learning titration. The finding is about people, not about students.
The second half of the story is forgetting. Hermann Ebbinghaus mapped it in 1885 and a 2015 replication found the same curve: most of what you take in without using it is gone within a day or two. A forty-five-minute webinar on Tuesday about a form you will first open next Thursday is, from memory's point of view, almost designed to fail.
Put those two findings together and you get a plain rule. If the learning does not happen at the moment of doing, most of it does not happen.
Adaptive means the teaching changes, not the learner
"Adaptive learning" gets used loosely. Here is the useful meaning. A fixed course shows everyone the same forty slides in the same order. An adaptive one watches what the person is trying to do, meets them there, and stops once they can do it.
Three things follow from that.
The unit is the task, not the course. Nobody needs to know the whole system. They need to create a job code, right now, correctly. Everything else can wait until it is the thing they need.
Explanation fades as competence grows. The first time, you say why. The second time, you say less. The third time, you say nothing, because they have it. A guide that keeps explaining is not thorough, it is in the way.
The company's version is the only version. Generic help can tell you what a field does. It cannot know that your finance team reconciles against the estimate number, or that "brief" means one specific thing in your studio. That local knowledge is exactly what people ask each other about, and it is exactly what the vendor's documentation leaves out.
Why this matters more now
Two things changed in the last few years. Companies now buy more software, faster, and swap it more often. And a large share of what they are buying is AI tooling that has no fixed "correct" path through it at all. There is no manual for a blank prompt box. The only way anyone learns to use it is by using it, seeing what comes back, and adjusting.
That makes the old model worse, not just slower. A recorded walkthrough of a tool whose interface changes monthly is out of date before it is watched. And when the tool is an assistant rather than a form, there is nothing to walk through. The person has to do the thing. The question is only whether someone is beside them while they do.
What it costs to get this wrong
Gallup's onboarding research puts the ramp-up time for most roles at around twelve months. Only twelve percent of employees strongly agree their organisation does a good job of onboarding. Between a third and two thirds of new hires leave within their first year. Those numbers are not about software, but software is where a great deal of that year goes: figuring out where the button is, asking, waiting, asking again.
Say a fifty-person team spends twenty minutes a day on questions that a nearby colleague answers. That is over sixteen hours a day, and the cost is doubled because it takes two people. Nobody reports this number because it does not appear anywhere. It is not a line item. It is just the shape of the work.
Sometimes, though, the cost does appear, and it is worth looking at those cases because they show what the invisible version is made of.
Revlon, 2018
In February 2018, Revlon switched its Oxford, North Carolina plant onto a new SAP system. The plant could not keep up. Orders to major retailers went unfilled for months. By the end of the year the company reported roughly 64 million dollars in lost net sales and another 54 million dollars spent fixing the disruption. Its auditors flagged a material weakness in financial controls. Shares fell more than ten percent on the news, and shareholders sued.
The software worked, in the sense that it ran. What did not work was the people who had to operate it on day one and could not yet do so at the speed the plant required. That gap between "installed" and "used" was the whole loss.
Lidl, 2011 to 2018
Lidl spent seven years and an estimated 500 million euros building a new SAP-based merchandising system. More than a thousand employees and a three-figure number of consultants worked on it. It went live in Austria, Northern Ireland and the United States. Then in July 2018 the board cancelled it and went back to the in-house system it was meant to replace, saying the goals could not be met "at reasonable cost".
Reporting at the time pointed at one detail. Lidl's whole business was built around purchase prices. SAP's retail software assumed retail prices. Rather than change how the business worked, the company tried to bend the software to match, and the cost of that bending is where the money went.
Two lessons, one from each. Revlon shows that a system people cannot yet run is a system that is not there. Lidl shows that "how we do it here" is not a detail you can train away. It is the thing you have to teach.
What to do instead
Move the teaching to the moment of use. Not before, in a room, and not after, in a recording. Beside the person, on the actual screen, when they are stuck.
Teach the company's version. Whatever your process docs say, whatever the person at the next desk would have said, that is the content. The vendor's manual is not.
Let the guide get quieter. The measure of good teaching is not how much help was given. It is how many people no longer need it.
And count that. Not questions asked, not videos watched. How many people, out of how many, can now do the task on their own. That is the only adoption number that means anything, and it is the one almost nobody tracks.