Something strange is happening with AI in the workplace. Companies spend heavily on licences, pilots and dashboards, but the real learning happens elsewhere: in the evening on the sofa, where someone asks ChatGPT for a birthday speech, or at the kitchen table, where a child puts forward a maths problem. That is where people grow comfortable with the technology. At the office, it is mostly something to hold meetings about.
That difference explains why most corporate AI investments deliver so little. The solution lies where executives rarely look: at home.
The uncomfortable number
Research from MIT showed last year that roughly 95 percent of corporate AI investments produced no measurable result (MIT NANDA, 2025). The cause was not the technology; the models work fine. The problem lay in what the researchers called the learning gap: the technology gets in, but people never turn it into a skill.
Access to AI does not make anyone skilled on its own. In one study, experienced software developers were allowed to use AI in their work. They believed they were working faster, but in reality they were 19 percent slower (METR, 2025). They had the tool, but not yet a feel for it.
That feel comes from doing, from hundreds of small attempts in which someone notices what works and what is nonsense. Those attempts largely do not take place at the office.
Where the practice hours are
OpenAI published figures on how people actually use ChatGPT. In a single year, use outside of work grew from 53 to over 70 percent of all conversations, faster than use for work (OpenAI/NBER, 2025). For every time someone uses AI for a report, that same person uses it twice for the life alongside it: a recipe, a tricky letter to the council, a dispute with the insurer.
The family is therefore the largest training ground we have. In the United States, 81 percent of parents use AI for parenting tasks (Lurie Children's, 2026). In the Netherlands, nearly 90 percent of secondary school students use AI for school (Scholieren.com, 2025). Within a household it also spreads in two directions: parents show their children something, children show their parents (Zhang and Cagiltay, 2025).
A work account that locks at five o'clock cuts off exactly the part where the least practice happens, and leaves untouched the 70 percent where most of the experience is gained.
We have seen this before
The computer went exactly the same way. In the late 1990s, researchers in a designed experiment handed out free computers for home use. Who received one was decided by lottery, so that cause and effect could be measured. The people who got a computer at home became demonstrably more skilled than the group without, a difference of around 17 to 19 percentage points (Fairlie, 2012). Not because anyone taught them, but because they had the device in their hands every day.
That same study produced a second finding. When the researchers looked at salaries years later, they found no demonstrable effect on income (Fairlie and Bahr, 2018). The skill was there, but the value did not arrive on its own: it only emerges with direction and guidance. Giving people an account and hoping for the best is therefore too easy.
The risk no one sees
Objection: letting people work at home with business-sensitive AI sounds like a data breach waiting to happen. The alternative is not a world without risk. It is risk the employer cannot see.
It is already happening. Employees use AI extensively out of sight of the organisation, and companies have an estimated less than a tenth of the actual usage in view (Awareways, 2025). That usage runs through free consumer accounts, outside any data processing agreement, while liability under the GDPR stays with the employer. At Samsung, employees leaked source code to ChatGPT in 2023 by simply pasting it into the text box (Samsung, 2023).
At home that risk is greatest and least regulated. That is also where the children are, the chatbots that pose as a friend, and the emotional dependence that is causing growing concern. UNICEF, the European AI regulation and Dutch organisations such as Netwerk Mediawijsheid arrive at the same line: guide, rather than ban or leave free (UNICEF, 2025; EU AI Act, art. 4 and 5).
But isn't AI free? Anyone who wants to can type chatgpt.com and start. That is true, and it is exactly what makes it so easy to do it wrong. The free version often puts people on the smallest model, precisely the type that failed most often in our own exam test. Anyone who practises only with that does not learn the strength of AI but its shortcomings, and soon concludes that it is all a bit overblown. An employer who steps in gives people access to the better models and the room to try several side by side, because which model is best differs by task.
Because anyone who thinks a random free chat service is safer than a data breach had better read the fine print first. Take use.ai, a service that is also promoted in the Netherlands. Everything you type is collected, in their own words 'the content of the prompts, questions, messages, or other data you enter', including sensitive information you put in yourself. That data goes to the United States, where you agree to 'lawful access requests by government authorities'. Anonymised input may also be used to improve their models. That is not an incident that leaks by accident, that is the business model.
A sponsored, sovereign family account can offer that guidance, a ban cannot. A European-hosted variant, where the data stays under the GDPR and in the Netherlands, brings home use under the same arrangements as at the office, instead of chasing it off to the free American variant. The Dutch government bans its own civil servants from consumer AI because the data processing cannot be properly arranged. That ban is the proof that the gap is there.
Employers already pay for home life
An employer contributing to something in the private life of staff: it has been happening for years. Companies sponsor tuition schemes that extend to employees' children, they arrange childcare, they provide budgets for sport and health. The thinking is always the same: some things are too important to stop at the front door. An employee whose life at home is in order performs better and stays longer (Gallup, via Compt, 2025). AI literacy is the newest skill in that list, after reading, arithmetic and learning to handle the computer.
One historical warning sharpens the difference. A hundred years ago, Henry Ford doubled wages, but sent inspectors to his workers' homes to see whether they were living decently (The Henry Ford, 1914). It was experienced as meddling and abolished within a few years. The model now on the table is the opposite. No inspector who comes by, no view into what people do at home. The account belongs to the employee and the family and stays theirs, even after a move to another job.
What this means for the Netherlands
The Netherlands is not a frontrunner with AI at work. There are many meetings about it and it is rolled out slowly. At the same time, the entire workforce is practising in the evening, unseen and without guidance. There lies an opportunity that no internal AI programme can match: making the usage that already exists safe and visible, and extending it to the people around the staff.
One caveat belongs here. That daily home use also makes employees better at the office is historically recognisable but not yet firmly proven for AI. That does not argue for waiting: the company that starts now provides the proof itself and sets the standard.
The question for executives is not whether their people will use AI; they already do. The question is whether the employer contributes to the place where the real learning happens, or whether the investment stays limited to the 30 percent at the office.