In August 2025, researchers at the Stanford Digital Economy Lab (Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen) published "Canaries in the Coal Mine?", an analysis of US payroll data with a finding that never quite broke through the news cycle: in the occupations most exposed to AI, employment for people between 22 and 25 sits 19% below where it would be expected. Employment for experienced professionals stayed flat.
In June 2026, SignalFire released its report on tech hiring. New-grad hiring: down 65% at major tech companies, down 76% at early-stage startups, compared with 2019. That is not an accident. It is the visible sign of a problem companies are still struggling to name: if AI now does the entry-level task that used to teach the newcomer, any company that wants to keep growing has to deliberately design how someone learns the job today.
For Brazil and Latin America, Stanford's number is a warning. Here, most companies are still experimenting. There, the shift has already happened.
The risk of nowhere to learn
When AI absorbs the entry-level requests — summarizing, checking data, drafting the first version of a document — the company loses the path it used for decades to train analysts. It used to be one person writing rough drafts that got better round after round, with a team reviewing every submission, giving feedback, and the newcomer taking that day's lesson home.
McKinsey tracked this in 2026. Banks, consultancies, and financial firms redesigned entry-level roles on purpose: computer simulation, annotated routines, concentrated mentoring. The reason is simple: without that, who will review a credit decision two years from now? Who will question the machine's suggestion when it's wrong? Bank of America, for example, kept its intern and campus-hire class at roughly 4,000 people in 2026, matching the prior year, but redesigned the roles around AI and started using simulation to speed up the judgment that used to come only from real routine work, year after year. Because the shortcut exists. Building it is the hard part.
The problem isn't ethical. It's about risk. According to a Gartner survey released in July 2026, at 22% of organizations at least one business leader has already stopped hiring for entry-level roles because of AI. What they didn't say is that these same companies will find out, a few years from now, that they have nobody qualified for the middle ground: no bridge between the automated and the senior executive. Just complex work, and nobody with the depth to question whether the machine got it wrong. It's a dangerous setup: the more the machine gets right on its own, the more blind trust it earns. The fewer people who tested it while they were learning, the fewer people who knew it failed — in the rare situations that, when they do happen, cause real damage.
Why the usual fixes don't work
Three things companies try:
First: "Let's hire more seniors." Result: market-rate seniors cost more, and the team doesn't have the budget to triple headcount. Besides, a senior who only reviews machine output never really becomes a senior.
Second: "Let's roll out online training." Result: much of the class never finishes the course, because a generic video can't replace real practice on a process the company spent years fine-tuning.
Third: "Let the newcomer learn from the AI." Result: learn from something that gets it wrong sometimes and nobody corrects? Learn the house standard when the standard is invisible? A newcomer who only talks to a machine becomes an expert at asking a machine questions, not an expert in the business.
What has to be in place
The environment that trains people rests on four verifiable pillars:
Approved knowledge with sources. Company documents aren't loose notes. They are texts with an owner, a current version, an expiration date, and a link back to the source — a law, a policy, a price table, an internal memo. When the newcomer looks up how to charge for a renewal, they find a document that says "per policy X, from Y, published in Z." They can cite it, check it, even question whether the rule changed in 2026. Visible knowledge is knowledge that teaches. Knowledge with a source is knowledge that doesn't disappear when the person leaves — it stays there, updated, audited.
Two-step review for anything critical. A contract, a termination notice, an opinion sent to a client, a limit exception: none of that goes straight from the machine to the client. It passes through someone who checks the facts, the tone, and the clarity. Then it passes through someone else who approves or sends it back. The second reviewer isn't a supervisor; they're quality control. And the newcomer who watches that review — or who is the reviewer — sees exactly where they were going wrong.
Reuse with permission. Templates, scripts, flows, and prompts that work stay saved for the next person to use. The newcomer building their first pitch learns by seeing five pitches that worked, with permissions set by whoever created them and whoever is allowed to use them, and a record of who changed what and when.
A clear decision trail. Who asked, who approved, when, why, what the alternative was. For agents: what criteria they used, what the context was, whether a human reviewer stepped in. The trail isn't there to blame anyone. It's there so the newcomer reviewing the history can see how decisions get made under uncertainty — which is exactly where someone becomes senior.
That is how Skyller was designed: company knowledge with sources, review and approval matched to risk, reuse with defined roles, and a trail for every action.
From studying alone to reusing what works
Training people right isn't just a risk gain. It's a productivity gain.
When the newcomer just studied on their own, it took three months to get up to speed. When they study material that has a source and goes through review, it takes a month — because they're confirming, not guessing. When they reuse templates, it takes two weeks. When the whole company reuses what one person learned, training doesn't get duplicated. The machine doesn't learn. People learn. The whole company gets faster.
There's a third gain, less obvious. In the shadow-AI scenario — everyone solving things on their own — the AI budget turns into waste: idle licenses on one side, people hitting limits on the other. When approved, reused knowledge exists, AI credits are shared. Teams see their own consumption. They know when to adjust — and, often, when not to buy more AI, because what was actually missing was communication about a standard that already existed.
Three questions to bring to your next meeting
Before deciding whether to hire less at entry level or redesign entry level, answer this with IT and with operations leadership:
-
If someone builds a good credit-request template, who else in the company can reuse it? If the answer is "nobody" or "only people who can code," the template never became knowledge. It became a lost file. Test it: look back over the last three months for someone who copied something that worked, and see if the result was that nobody knew it existed, didn't have permission, or didn't know how to adapt it.
-
When an analyst starts today, how long does it take before they can deliver something critical safely? If it takes eight months because they learn by watching, the company is bleeding money. If it takes eight weeks because there's documentation with sources, that's progress. If it takes eight days because there's an annotated script, even better. That's what the companies winning with AI are actually doing: compressing the learning curve, not cutting the number of people who learn.
-
If the machine gets a recommendation wrong, who will catch it? If the answer depends on "someone reviewing it later," it won't happen. If it depends on "someone reviewing it beforehand, at two levels," it will — one person who understands the task, another who understands the risk. And if the trail behind that review leads to a document explaining the criteria, the next newcomer sees exactly how to question the machine.
