Enterprise AI

The Next Enterprise AI Crisis Is Critical Thinking

By Ken Yao, Head of Partnerships at TryHackMe

AI adoption isn’t the challenge anymore. Most enterprises already have employees using the tools, licensed models in place, and usage numbers that keep climbing. The real risk now lives in the moment after AI produces an output:  who checks it, how, and what they decide to do next. 

Workers are increasingly accountable for outputs they can’t independently validate. That’s not something organizations have prepared for. For the past two years, organizations have taught professionals how to generate work with AI. Now, training needs to evolve beyond prompting and into evaluation. 

Judgment is already deferring to the machine 

As more employees turn to AI tools to get their work done, they’re also handing over their thinking.  Nearly two-thirds (64%) of workers believe AI recommendations can be as helpful as their own judgment, and among security professionals the pattern is starker: 76% rely on AI for tasks they don’t fully understand as part of their normal workflow [1]. That shift in trust is dangerous. People question AI less often, so the problems it introduces or misses are more likely to go unnoticed. 

Discernment is eroding because people have only been taught to prompt. This is not some fringe effect of a few over-reliant employees. It’s the direct, predictable result of what training does and does not teach. 

New data from TryHackMe’s AI Skills Illusion report, a Censuswide survey of 1,000 US office workers, shows how wide that gap has become [1]. Seventy-three percent of workers say AI training is available at their job, but fewer than half, 41%, believe that training has actually prepared them for a changing work environment. Companies have rolled out the courses, and workers still don’t feel ready. 

The dependence underneath those numbers runs deep. Forty percent of the US workforce say they’d struggle to perform their job effectively without AI tools, and 43% admit they regularly use AI to complete work they don’t fully understand [1]. Neither of those is a character flaw. Both point to training that teaches people how to ask AI for the right answer but never teaches them how to tell if the answer that came back is actually right.  

Understanding verification as a distinct skill 

Verification literacy is the piece missing from most training plans. In practice, it means a worker can look at AI-assisted work and answer whether it’s technically correct, whether it’s safe to act on, and whether they could explain the reasoning behind it if asked. What it requires is domain knowledge: understanding the work well enough to catch the tool when it’s wrong. That makes it a fundamentals problem, because nobody can verify an answer in a domain they never learned. 

A security analyst reviewing an AI-flagged threat checks whether the behavior actually matches known attacker tactics, rather than accepting the flag because the tool surfaced it. A finance analyst reviewing an AI-generated forecast, meanwhile, traces the assumptions behind a number back to source data, rather than accepting a chart that merely looks plausible. 

Building the habit, and actually measuring it 

One practical fix is to separate generating work from validating it, at least during training. Have employees work through a set of problems without AI first, then compare their own reasoning against what the tool comes back with. Wherever the two diverge, you’ve found the part of the curriculum that needs work. 

Spot-checking helps too, and it does double duty against the fear driving shadow AI. Instead of only reviewing AI-assisted work after something goes wrong, leaders can periodically ask an employee to walk through the reasoning behind a specific output, picked at random, after the fact. When review is routine and blame-free, people stop hiding how they used the tool. 

Review has a blind spot, though. It only covers what the model surfaced. The harder problem is the false negative: what a model waves through without raising a flag, because no human team can review at the speed or volume a model operates at. Training has to cover both directions: catching AI when it’s wrong, and noticing what it quietly let through.  

But building these routines is only one part of the equation. Organizations need to know how to measure it too. Most teams measure AI by adoption rates and usage volume. The next step requires looking at the other side of the coin. How many wrong outputs were produced and, critically, how often an employee can catch an incorrect response. It’s a metric missing from most dashboards, and the one that will give leaders a clearer view of just how exposed the company truly is. 

 The test enterprises haven’t started studying for 

None of this asks companies to slow AI adoption or treat the technology with suspicion. Adoption has already happened. What it asks is to start naming verification as its own skill, one that gets taught and measured separately from how well someone operates the tool. 

The companies that come out ahead probably won’t be the ones running the most advanced AI stack. They’ll be the companies where people can look at AI-assisted work and say, with evidence, why it holds up, or catch it when it doesn’t. That kind of capability won’t show up in an adoption report. A leader usually finds out it’s missing only after something has already gone wrong, when the question becomes who actually understood the work well enough to have caught it first. 

Getting people onto AI tools was always going to be the easy part. Teaching them what to trust, and how to prove it, is the harder assignment most enterprises haven’t begun yet. 

Ken Yao is Head of Partnerships at TryHackMe, a hands-on cybersecurity training platform used by more than 8 million learners worldwide. He spent nearly 15 years as a cyber practitioner working with government agencies and regulated enterprises, and has since held CISO roles at global enterprises.

References 

[1] TryHackMe. (n.d.). The AI skills illusion report [Survey of 1,000 US office workers, conducted by Censuswide]. [https://tryhackme.com/business/resources/reports/shadow-AI-report] 

 

Related Articles

Back to top button