Your people completed the training. Do you know if it worked?

Enterprises have invested heavily in AI training over the past three years. The programmes are well-intentioned. The completion rates look healthy. The certificates are prominently displayed. And the underlying assumption — that completing training produces competence — goes largely unexamined.

The research does not support that assumption.

The BCG-Harvard study of 758 consultants, published in Organization Science in March 2026, found that AI training alone showed no statistically significant performance advantage over simple tool access. The consultants who received training and those who simply received the tool performed indistinguishably. The variance in outcomes was driven by individual proficiency — a characteristic that training completion cannot detect.

Section AI's 2026 Proficiency Report, which combined surveys with hands-on skill testing across 5,000 knowledge workers, confirmed the pattern at scale: 97% of the workforce are using AI poorly or not at all. Among them, employees who had completed AI training programmes still scored only 40 out of 100 on proficiency assessments. They remained firmly in the “experimenter” category — capable of basic prompting but unable to reliably evaluate AI output, identify when AI use is inappropriate, or manage the risks of AI-generated content in professional settings.

Aalto University researchers published findings in Computers in Human Behavior (February 2026) that upend the assumption behind self-assessment. In two studies with approximately 500 participants, they found a reverse Dunning-Kruger effect: higher AI literacy correlated with greater overconfidence, not better self-calibration. Participants using ChatGPT overestimated their correct answers by 4 points out of 20 — a gap larger than the actual performance improvement from using AI. Financial incentives for accurate self-assessment did not correct the bias.

The industry data confirms it at scale. 79% of tech workers admit to pretending they know more about AI than they do (Pluralsight 2025). 81% profess confidence in their AI skills, but only 12% have significant hands-on experience. And 64% of workers pass off AI-generated content as their own (Salesforce 2024) — a behaviour that self-assessment by definition will not reveal.

Kruger and Dunning's foundational 1999 research found that bottom-quartile performers rate themselves at the 58th–62nd percentile on average. The gap is structural, not motivational — people lack the very skills needed to recognise their own deficiency. In the context of AI, where outputs appear fluent and authoritative regardless of their accuracy, this metacognitive blind spot is particularly dangerous.

Completion tracking misses the problem. Self-assessment misrepresents it. Standard LMS quizzes — built on Classical Test Theory where all items count equally — lack the precision to detect it. The difference between an LMS quiz and psychometric proficiency measurement is the difference between a pop quiz and a medical board exam: one checks recall, the other measures whether you can practise. Performance-based psychometric assessment measures the capability that matters — with known precision, calibrated difficulty, and scores that mean the same thing regardless of which questions were asked.

AI proficiency measurement applies the same psychometric science that has been trusted for 60 years in the highest-stakes assessments — from graduate admissions to medical licensing to military selection — to the specific question of how effectively professionals use AI in their work.

The approach differs from completion tracking in three fundamental ways. First, it accounts for question difficulty. Answering a hard question correctly reveals more about proficiency than answering an easy one — a principle that percentage-correct scoring ignores entirely. Second, it produces comparable scores across different test forms. Because ability is estimated independently of the specific items asked, two employees who take different versions of the assessment receive scores on the same scale. Third, every score includes a confidence interval — an explicit estimate of how precise the measurement is, preventing managers from over-interpreting small differences that may be noise.

The result is organisational visibility that completion metrics cannot provide: which teams can deploy AI effectively today, which need targeted development, where the risks of unsupervised AI use are highest, and whether training investments are producing measurable change over time. The board presentation shifts from “87% completed the programme” to “Advisory is at Competent level, Tax is Developing, and here is where the next quarter's investment should go.”