AI and the workforce have the same blind spots

Every AI rollout comes with the same internal pitch: We’ll move faster, and we’ll stay accurate, because humans will review the output. It’s a two-part promise. Most organizations have spent the last two years measuring the efficiency of this, but not many have seriously tested the second part.

New research suggests we should be questioning whether humans are holding up their side of the bargain.

Two bets, one plan

The logic is intuitive enough. AI handles execution with speed, consistency and at scale. Humans handle judgment, context, skepticism and error detection. Together, you get faster output and a built-in check on work quality. It’s become a standard part of every AI governance plan.

What it assumes is that the human reviewer can actually do the job the plan assigned them.

Cangrade mapped 71,747 Gen Z and Millennial skills assessments against the five soft skills that appear most consistently in AI-era job postings, drawn from an analysis of 200 AI-related roles across industries, seniority levels and functions. The five skills AI-augmented roles required for success were remarkably consistent: Communication, strategic thinking, critical thinking, attention to detail and creative problem-solving.

How well the younger workforce measures up against those skills is where the plan starts to show cracks.

What the data shows

Not everything in the findings is a concern. Gen Z and Millennials scored 14% above average in Communication, ranking 8th out of 40 measured competencies. That’s a genuine strength and a well-timed one. As AI handles more content generation and information retrieval, human communication shifts toward higher-value work like aligning stakeholders, translating AI outputs into decisions and managing the coordination that AI augmentation actually increases. The incoming workforce is well-positioned for that shift.

Strategic thinking landed just under average, 1% below baseline, ranking 24th out of 40. For most roles, that’s sufficient. But there are two scenarios where it isn’t. For senior positions where setting direction is the primary responsibility, it warrants direct assessment rather than assumption. Secondly, as analytical tasks are increasingly outsourced to AI, the human role in strategy shifts from information gathering to high-level decision-making. While AI excels at identifying patterns and processing data at scale, it can’t weigh competing priorities, set a clear direction, or grasp contextual meaning. As that analytical work is outsourced to AI, an “average” level of strategic thinking is the minimum.

The skills that matter most for work quality and accuracy are further down the list.

Where the plan falls short

Critical thinking ranked 37th out of 40 competencies, 18% below average. Attention to detail ranked 36th, 17% below average. Creative problem-solving ranked 29th, 10% below average.

Those competencies are exactly what ensuring work quality depends on. Critical thinking questions a confident-sounding answer before accepting it. Attention to detail finds the error in AI output before it moves on. Creative problem-solving reframes questions or proposes novel solutions that fall outside the pattern-recognition capabilities of the AI’s training data. If the people reviewing AI output are weak in all three, the review part of the plan isn’t working.

And the pattern is persistent. The critical thinking gap has held across two years of data and an 113% increase in sample size. It’s time to examine the plan more closely.

The speed is real. The quality isn’t guaranteed

The efficiency AI brings to organizations is irrefutable. Output volumes go up. Turnaround times come down. Those numbers show up in dashboards and are easy to point to at a board meeting.

But the work quality piece is harder to measure. In most organizations, it isn’t measured at all. The assumption is that the human review step makes AI-augmented output as accurate as, or more accurate than, what the team produced before. But that assumption only works if reviewers actually catch errors.

Large language models don’t flag their own mistakes. They produce a plausible answer with the same confidence whether it’s correct or not. The human reviewer is the error-detection mechanism. When that mechanism is weakest in exactly the competencies error detection requires–critical thinking, attention to detail and creative problem-solving–the partnership breaks down. Reviewers become more likely to default to the AI’s suggestions because generating a genuinely better alternative requires the very capabilities that are in short supply. The result is a partnership that appears complementary but fails in practice.

Because AI accelerates every workflow, any wrong or incomplete answer that passes through a weak review cycle moves further and faster than ever before.  As automated systems get embedded in more processes, the volume of output needing verification explodes. And so does the pressure to approve quickly rather than scrutinize closely. The speed is visible on the dashboard. The work quality gap is invisible until something goes wrong: A factual error that ships, a strategic decision fueled by unverified data, a hiring recommendation that was never properly scrutinized. By the time the exposure surfaces, it’s already cost something.

Where this matters most

Not every AI-augmented role has the same stakes. The gap in critical thinking and attention to detail matters more in some contexts than others, and it’s worth being precise about where.

Think about it in terms of two variables: How consequential a missed error is in a given role, and whether the team in that role has actually been assessed for the competencies required to catch one.

Where consequences are high, and assessment hasn’t happened, that’s where your organization is the most exposed. Output review, quality assurance, analysis that feeds decisions, AI-assisted hiring – these are the types of roles where errors that get through have real downstream impact.

Where consequences are lower, average critical thinking is adequate. Not every workflow requires a rigorous review of every AI output. A missed error in an internal brainstorm is different than a missed error in a client deliverable. The risk is proportional to what a missed error actually costs.

To minimize your exposure, map AI-augmented roles by consequence, then assess the teams in the highest-consequence positions directly, before assuming those capabilities exist because the headcount does.

Building the quality the plan assumed

Treating human review as a skill that has to be deliberately built is what makes the second half of this AI governance step work.

Start with measuring the right competencies before assigning oversight responsibility. Resumes and unstructured interviews don’t reliably surface critical thinking or attention to detail. The variation in these competencies across candidates is significant, and it’s only visible through direct assessment.

Extend it to team design. A team that pairs strong critical thinkers with strong communicators covers more of the human-AI collaboration requirements than a group of generalists who are adequate across the board and strong at nothing in particular. The gap at the individual level becomes less consequential when teams are built to account for it.

Align roles to actual skill profiles rather than assumed ones. Competency requirements vary across AI-augmented roles. In high-stakes roles, critical thinking and attention to detail are paramount. In roles where AI handles the data and humans are responsible for interpreting it, creative problem-solving matters most. Match the oversight responsibility to the actual, measured abilities of your workforce, not assumptions.

And separate the two metrics that keep getting conflated. Track speed and work quality as distinct outcomes. If the only number coming out of AI deployment is throughput, there’s no way to know whether the second bet is paying off, or quietly failing.

The organizations that get this right won’t be the ones that moved fastest. They’ll be the ones that verified what they assumed.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?