Companies have turned around on how they get AI skills. In 2024, two in three company leaders planned to hire AI talent from outside rather than train their staff. In 2026, organizations are about three and a half times more likely to train existing people than to hire in the areas they call strategic. The reason is what an AI skill is worth once everyone can learn it at the same speed.
A head start in AI lasts a few months
Experience with a tool is only worth something while the tool stays the same. AI tools change every few months, so what an outside specialist learned about last year's version is partly out of date by the time they start, and they still have to learn the company's systems before they can use any of it. By then the people already there have been using the newest tools on real work for months.
Guardant Health, a cancer-testing company, sees it this way. Much of its code is already written by AI, its engineers learned as they went, and it has not hired an outside AI specialist yet. A new hire's edge over its own team would be weeks.
The newcomer cannot bring the business
What does not wear out is knowing the business. This fits what we wrote in September about company know-how becoming AI's scarcest resource.
For example, a hospital group's finance team could have AI draft the monthly explanation of why each department's costs moved. Any capable person can set that up in a few days. But the draft is only useful if someone can read it and see at once that the radiology number is wrong, because a scanner contract renewed in that month. The accountant who has been there eight years sees that in seconds. A newcomer needs a year of closes to see it. If five people each spend two days a month on those explanations, a reviewed draft could give back most of ten working days every month, and only someone who knows the numbers can safely review it.
The insurer Liberty Mutual works on the same logic. Its technology staff have been there more than 11 years on average, and it builds in-house the AI skills closest to its own data and its own decisions. It buys what is widely available on the market.
A course is not training
Most company AI training stalls, because one-off sessions do not stick. A certificate shows that someone sat through a course. What counts is whether they now do a piece of their work differently, and that needs the tools, the permission and a real problem to try them on. Companies that train and then leave people to find a use for it will see little change, and they will wrongly decide that training does not work.
Hiring shrinks, and becomes a way to teach
Outside hiring does not disappear. It gets smaller and more specific: a few people whose full-time job has been AI for two years, who also know a field like cybersecurity or bioinformatics, and who are hired to teach the staff, not to do the work alone. A specialist whose knowledge stays in one head is a poor hire in a field where knowledge expires quickly.
For the person being trained, this makes the choice simple. If your employer offers AI training, the people who get ahead will be those who bring one real task from their own week, do it with AI and show what changed. That is something a newcomer cannot copy, because it depends on knowing your work.