Companies report that AI saves employees 40 to 60 minutes a day. At the same time, 80 percent of companies report no measurable productivity gain at the company level. The space between the two is the time people spent thinking, interpreting, and questioning. Who owns that time now?
OpenAI's 2025 report on enterprise AI use describes the saving of 40 to 60 minutes a day as a central success figure.1 An NBER study from February 2026 that surveyed nearly 6,000 managers in the US, the UK, Germany, and Australia found at the same time that over 80 percent of companies report no measurable effect on either employment or productivity.2 PwC's 29th Global CEO Survey, with 4,454 leaders, showed the same pattern: 56 percent saw neither higher revenue nor lower cost from AI.3 Deloitte's ROI of AI study found that only 15 percent of companies report significant measurable ROI even though 85 percent have increased their AI investment.4
Several analysts, among them Faros AI and McKinsey, describe the phenomenon as an AI productivity paradox. The gains are real at the task level and dissolve at the company level.
The usual explanation is downstream bottlenecks, training gaps, or measurement problems. Those explanations are partly correct and structurally incomplete. Part of what gets measured as time saved is time that used to go to interpretation. When the measurement system counts that time as a cost, it counts out the thing the competence was made of.
What the interpretation gap is
The interpretation gap (tolkningsmellanrummet) is the cognitive room that opens when something stays unclear long enough to force a person to engage with it. The hesitation, the reconsideration, the moment where a phrasing resists and demands a position. That is where professional judgment is exercised. It is also the first thing to go when AI systems are optimized for flow.
The time measured as saved is often time that used to go to interpretation. When the measurement system counts that time as a cost, it counts out the thing the competence was made of.
What disappears and shows up as gain
Three mechanisms shrink the gap without anyone deciding to. Auto-complete in decision documents sets an anchor before judgment has formed. Sycophantic assistants remove the resistance that used to make the user think again. MIT published in February 2026 a mathematical proof that a model trained to agree can push a rational user toward incorrect conclusions through confirmation alone.5] AI summaries replace reading the original and deliver an interpretation that is already made.
Each of these reads as productivity in a KPI built on time saved. Together they form an architecture where interpretation never quite has to happen. Gerlich's 2025 study, with 666 participants in the UK, found a correlation of 0.72 between AI use and cognitive offloading, and minus 0.75 between offloading and critical thinking.6] The figures sit in the range counted as very strong.
Who owns the interpretation gap
The question gets concrete when you hold it against the org chart. Four functions have a possible claim on the responsibility, and none of them carries it today.
The CIO and CTO own the tool choice. They are measured on adoption, time saved, and integration. The measurement system makes it rational for them to maximize the very things that shrink the gap.
HR and L&D own AI literacy under Article 4 of the EU AI Act. In practice the mandate is often read as teaching employees to prompt efficiently. Responsibility for what happens to the organization's collective interpretive capacity rarely lands here in earnest.
The line and the employee see the consequences first. Decisions grow more homogeneous, divergent readings rarer, reconsideration later and weaker. The rhetoric places the responsibility here, critical thinking is your job, while the mandate over the AI architecture sits elsewhere.
AI governance and compliance own risk. Cognitive offloading rarely appears in their taxonomy. The risks are data leakage, bias, discrimination, law. That an organization's collective interpretive capacity erodes is a slow risk that few frameworks catch.
None of these owns the interpretation gap today. That is the problem itself.
The time someone has to own
The interpretation gap is at once a capacity question, an architecture question, and a flow question. So it cannot sit with one function. It has to be carried by three working together.
L&D owns capacity. That means AI literacy that goes deeper than tool knowledge and builds employees' ability to notice when the interpretation gap is shrinking, understand what it costs, and know how it is defended. That is what Article 4 makes possible when it is taken seriously.
AI governance owns architecture. That means design choices about where auto-complete is allowed and where it is held back, which decisions require the human to form the question before the model answers, which documents may be summarized and which are to be read. These are technical choices with cognitive consequences, and they need to be treated as such.
The line owns flow. That means decision processes designed with explicit friction at the points where judgment is the value being delivered. Adversarial prompting, mandatory source review, structured points of reconsideration. It costs time. That is the point: the time the friction takes is the time the organization actually thinks.
A CIO who delivers time saved is rewarded even when the saving hollows out interpretive capacity. An L&D function that delivers AI literacy as tool knowledge is rewarded even when it misses what Article 4 requires. A governance team focused on data leakage is rewarded even while it allows cognitive leakage at the same time. Each of these incentive structures stands on its own logic, and together they form a system where the interpretation gap has no defender.
The question is not only which tools the organization gives its employees. The question is who owns the time the employee spends thinking, and what happens to the organization's judgment when no one does.
The interpretation gap in other contexts
- The Confirmation Machine — the concept in interplay with sycophantic AI systems.
- Cognitive Impairment in the AI Era — what happens to cognitive capacity when the gap is systematically shrunk.
Sources
1: OpenAI (2025). The State of Enterprise AI 2025. Time saving of 40 to 60 minutes a day. https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf
2: National Bureau of Economic Research (February 2026). Survey of nearly 6,000 managers in the US, UK, Germany, and Australia. Over 80 percent report no measurable effect on employment or productivity.
3: PwC (2026). 29th Annual Global CEO Survey. 4,454 leaders. 56 percent saw neither revenue gain nor cost reduction from AI.
4: Deloitte (2026). The ROI of AI. 1,854 managers in 14 countries. 85 percent increased AI investment, 91 percent plan more, 15 percent report significant measurable ROI.
5: MIT (February 2026). Mathematical proof that sycophantic models can drive rational users toward incorrect conclusions through confirmation alone.
6: Gerlich, M. (2025). Study with 666 participants in the UK. Correlation r = 0.72 between AI use and cognitive offloading, r = -0.75 between offloading and critical thinking.
