New research on early-career professionals found something most managers do not expect: employees who got measurably better at their jobs with AI and employees who got measurably worse were using the identical tool. The difference was not critical thinking scores, AI literacy, or domain knowledge. It came down to one specific, teachable behavior.
The skill is orchestration: treating AI as a collaborator you have to direct, check, and revise rather than a tool you hand a task to and wait on. Employees who frame the problem, verify the output against what they already know, and revise across multiple rounds get measurably better results. Employees who accept the first draft do not, and that gap shows up in real performance data, not just in survey opinions.
It is not aptitude. A 2026 field study of 523 early-career professionals from KPMG and researchers at the University of Texas at Austin found that critical thinking scores, prior AI literacy, and domain knowledge did not predict which employees got better with AI and which got worse. What predicted the outcome was whether someone treated the AI as a collaborator to direct and check, or as a system to hand work to and accept back.
The researchers, writing in Harvard Business Review, split participants into three groups based on how their AI-assisted work compared to a human baseline on the same tasks. Just over half, 50.1 percent, became what the KPMG and UT Austin study calls AI Amplifiers, beating the baseline outright. Another 25.8 percent were AI Delegators, whose results roughly matched it. The remaining 24.1 percent were AI Apprentices, whose output fell below what an unassisted person would have produced working alone.
That last figure deserves attention from any leader rolling out AI access company-wide. Nearly a quarter of early-career workers in this study did worse with an AI tool in hand than they would have done without it. Access is not the same as capability, and for a quarter of the sample, access alone measurably reduced it.
| Group | Share of participants | What they did with AI |
|---|---|---|
| AI Amplifiers | 50.1% | Framed the problem, directed the AI, checked output against real domain knowledge, revised across multiple rounds |
| AI Delegators | 25.8% | Accepted output close to the first draft, with limited direction or review |
| AI Apprentices | 24.1% | Tried to critique the output but chased the wrong issues, or accepted it with little verification |
Amplifiers did three things consistently. They framed the problem for the AI instead of accepting how the tool framed it first. They anchored the output against domain frameworks they already knew, rather than trusting it on faith. And they revised the result across multiple passes instead of shipping the first version. Apprentices, by contrast, often tried to critique the AI's work but chased irrelevant details instead of the errors that mattered. Delegators mostly skipped that step altogether and moved on.
Not exactly. Unrestricted access without any coaching on how to direct and verify AI output tends to produce the Apprentice pattern more often than the Amplifier pattern, and a near-even split of the workforce already agrees some kind of guardrail makes sense. The fix is not a ban. It is teaching the orchestration behavior on purpose instead of hoping new hires discover it on their own.
A CNBC and SurveyMonkey survey of 1,686 US workers, fielded in the third quarter of 2026, found the workforce split almost down the middle on how junior staff should use AI. Forty-two percent said junior employees should be allowed to use AI with clear guidelines and limits. Thirty-five percent said AI should be prohibited for junior-level workers outright. Another 14 percent wanted supervision and approval required, and just 9 percent supported free, unrestricted use.
That split is not indecision. It is a reasonable response to watching junior employees who have not built judgment yet default to the Apprentice pattern, accepting AI output without the framing and verification that make it useful. A blanket ban avoids that risk, but it also blocks the only path to becoming an Amplifier, which requires practicing the orchestration behavior on real work, with some structure around it.
The same survey found usage itself is uneven inside most companies. Thirty percent of workers use AI daily, 34 percent use it occasionally, and 37 percent say they never use it at work at all. Most organizations are not managing one AI adoption problem. They are managing a wide range of exposure under one roof, from employees who have never opened the tool to employees already running a meaningful share of their week through it with no structured coaching in either direction.
PwC's 2026 Global AI Jobs Barometer, built from more than a billion job postings across 27 countries, shows employers are already paying for the orchestration behavior even if they cannot name it on a job posting. The wage premium for workers with AI skills reached 62 percent this year, up from 57 percent in 2025, and the spread runs as high as 118 percent in the highest-paying sectors.
The clearest signal is in entry-level hiring. Roles most exposed to AI are now seven times more likely to require judgment and leadership skills that used to be reserved for senior staff, and these seniorized entry roles have grown 35 percent since 2019 while other entry-level postings have shrunk. Employers are not asking new hires to know more. They are asking them to exercise more judgment, sooner, because that is exactly the behavior separating Amplifiers from Apprentices.
The same report found headcount growth at the most AI-exposed companies outpacing growth at the least AI-exposed companies, 52 percent versus 36 percent against a 2018 baseline. Companies leaning into AI are not shrinking their entry-level workforce because of it. They are growing it faster, but mostly for the entry-level workers who can operate at a more senior level of judgment earlier than previous cohorts had to.
That shift changes what hiring and onboarding should screen for. A resume listing AI tools someone has used tells you almost nothing about which of the three groups they fall into. If you are building a program to evaluate and develop that judgment across your team rather than leaving it to chance, Securafy's AI governance and security services are built around exactly that gap, assessing how your people actually use AI today, not just whether they have access to it.
The caution behind those restrictive attitudes has research behind it too. A 2026 study in the journal Cognitive Processing compared AI-assisted and manual task performance in 120 young adults and found that the AI-assisted group needed significantly less mental effort to reach their answers. On its own that sounds like a win. The detail that matters is how they got there.
The researchers identified two distinct patterns inside the AI-assisted group: complete delegation, where participants handed the task to the tool and accepted the result, and selective, strategic engagement, where participants used AI as a scaffold and stayed involved in the reasoning. Which pattern someone fell into determined whether AI functioned as a replacement for their own thinking or a support for it, the same fork in the road the KPMG and UT Austin study found between Apprentices and Amplifiers, reached through a completely different research method.
That convergence matters more than either study alone. Two separate research teams, using different populations and different methods, landed on the same conclusion. The behavior around the tool determines the outcome, not the tool itself and not how capable the person using it already was on paper.
Once orchestration is the skill that matters, the practical question becomes how to teach it deliberately instead of hoping new hires stumble into it. A few practices show up consistently in organizations that move junior staff toward the Amplifier pattern.
None of that requires banning AI or waiting for someone to figure it out on their own. It does require knowing where your team actually sits before you decide what to fix.
For a security-conscious or regulated business, that individual-level picture matters as much as the productivity case does. An Apprentice-pattern employee is not only producing weaker work. They are the person most likely to paste client data into a tool without checking where it goes, or push AI output into a regulated workflow without the verification step a reviewer would expect. That is a data handling and access control problem as much as a training gap, and it is worth knowing where those gaps sit in your environment before an incident finds them for you. Securafy's cybersecurity assessment is built to surface exactly that kind of exposure at the team level, not as a single company-wide score.
If you are evaluating vendors or tools as part of that response, the questions worth asking go beyond feature lists. Securafy's cybersecurity buyer's guide walks through what to verify on security and compliance grounds before you commit budget to any platform your team will be feeding company data into, AI tools included.
The gap between an Amplifier and an Apprentice is not talent. It is a behavior you can teach, measure, and build into how your team already works, starting with the people you have right now.
If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. Securafy AI University gives your people role-based AI training with security built into the material, not bolted on afterward.
If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.