Judgment: the ability to supervise, validate and override AI outputs. IBM’s Institute for Business Value study published September 21, 2026 (1,500 CHROs, 8,800 employees) found 71% of CHROs call it the most essential workforce skill while only 29% of employees rank judgment as important. Davenport and Mittal’s All-In On AI says the human side decides AI outcomes; my fix is a weekly override ledger.

IBM published a workforce study this morning, September 21, 2026, and I read it twice before my first client call: once as the coach who is asked every week what skill to build, and once as the person who builds the systems those skills are supposed to supervise. The headline is a gap, not a trend. The people who design work say one thing matters most. The people doing the work do not rate it.

The numbers come from the IBM Institute for Business Value, working with Oxford Economics, which surveyed 1,500 chief human resources officers across 21 geographies and 8,800 full-time employees across 28 countries between April and June 2026. Seventy-one percent of CHROs identify the ability to supervise, validate and override AI outputs as the workforce’s most essential skill. Only 29% of employees rank judgment as important. Sixty percent of employees worry AI is eroding their skills, and among those who worry, three in four say the erosion has already started, with critical thinking the skill named most often. CHROs agree on the diagnosis: 57% put critical thinking and 48% put human judgment among the capabilities that matter most.

The Override Ledger — the four steps I give clients

  1. Classify before you prompt. For every recurring decision in your week, write whether it is human-led, AI-assisted or AI-executed. IBM found organizations that define workflows this explicitly report 18% lower risk and 20% higher quality. The same labels work for one founder’s calendar.
  2. Log every override. When you reject, edit or reverse a model’s output on anything AI-assisted, record what it got wrong and why you knew. A week with zero overrides is not a week the model was perfect; it is a week you stopped checking.
  3. Promote and demote monthly. A task that produced no overrides in a month moves toward AI-executed. A task that produced a costly one moves back to human-led. The ledger, not your mood, decides the boundary.
  4. Name the human owner out loud. IBM’s employees said that when AI goes wrong, 43% of them take the blame. Put a name next to every AI-assisted decision before it ships, starting with your own.

I reach for Thomas Davenport and Nitin Mittal here because All-In On AI made the same argument three years before this survey did, from the company side. Their finding was that leadership, culture, attitudes and skills affect AI results as much as or more than any technology, and that the organizations winning with AI behave like learning machines, monitoring their models in production and retraining them when accuracy drifts. Nobody applies that discipline to themselves. We deploy a model into our own workflow and never check whether our judgment of it has drifted. Davenport and Mittal also report that 80% of executives they surveyed leaned toward replacing employees rather than retraining them, which they called shortsighted; IBM’s data shows the cost of that instinct landing on individuals as unrecognized checking work. Eighty percent of CHROs say AI creates invisible work, validating recommendations, fixing mistakes, providing context and managing exceptions, and 42% of employees say AI increased their work or made it go unnoticed.

The finding I keep returning to is about safety rather than skill. Where the CHRO at least shares responsibility for deciding which decisions stay human-led, 76% of employees say they feel safe to question or override an AI recommendation. Where HR is merely advisory, that falls to 43%. Where judgment is built into how work is done, 62% of CHROs report growing confidence in AI-enabled decisions; where it is not, 57% report confidence declining. Read that as a founder and the lesson is uncomfortable: your team’s willingness to overrule the model is a policy you set, not a trait they have. If nobody has said out loud that the AI can be wrong and that overriding it is rewarded, the ledger above will stay empty and the mistakes will ship under your name.

The limits, named. IBM sells AI, and a study that concludes companies need better human governance of AI is also a study that concludes companies should buy more of it with IBM’s help. The skills-erosion figures are self-perceived; nobody tested anyone’s critical thinking before and after. Fieldwork ended in June 2026, before the September model releases that changed how confidently these tools answer. And the override skill has its own failure mode: I have watched clients turn validation into reflexive distrust, re-checking outputs that were fine, which is precisely the invisible work the 80% figure describes. The ledger is supposed to shrink over time as tasks earn AI-executed status. If it only grows, you are not supervising, you are duplicating.

What I actually do is smaller than it sounds. My ledger is a single note with three columns, and most weeks it has four or five entries. Two tasks have moved to AI-executed this year: first-pass meeting summaries and formatting a client pre-read. One moved back: a pricing recommendation that a model built on a market assumption it never stated, which I caught only because the override habit made me ask what it was assuming. That one entry paid for the whole practice. Judgment is not a skill you have. It is a log you keep.

Sources: IBM Newsroom — New IBM CHRO Study: AI Puts Critical Thinking at the Center of Workforce Priorities (Sept 21, 2026), IBM Institute for Business Value — CHRO C-suite study (full report and methodology), StockTitan syndication of the IBM release.

Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours