Individual AI productivity is not an enterprise outcome
McKinsey’s survey separates self-reported gains for workers from reported EBIT impact. Leaders should test handoffs, queues and rework before treating tool use as organisational productivity.
What happened
McKinsey’s 2026 survey of 1,719 respondents in 97 countries reports that 80% saw individual productivity improvement, while 37% attributed some positive EBIT impact to AI and 6% qualified as high performers.
Why it matters
The figures are different measures with different denominators and cannot prove that workflow redesign causes profit. They still warn against using individual speed or token volume as a proxy for system value.
McKinsey’s 2026 global AI survey covered 1,719 respondents in 97 countries and was weighted by national GDP. Eighty percent said AI improved their individual productivity; 37% attributed some positive EBIT impact to AI, essentially unchanged from 2025; and 6% met McKinsey’s definition of an AI high performer. About one in five said AI operating costs constrained use.
These are self-reported survey measures, not audited productivity or a causal comparison. Individual productivity and EBIT are not inverse sides of one metric. Revenue conditions, investment timing, selection into the survey and respondent knowledge can affect the results. Business Insider’s report quotes McKinsey’s Eric Kutcher arguing that companies must inspect work across the organisation rather than only distribute tools.
The comparison also contains no matched before-and-after enterprise denominator. A company can report many faster workers and still see no margin movement because demand, quality, coordination or complementary investment changed at the same time. Conversely, EBIT can improve for reasons unrelated to AI. Treat the figures as a diagnostic gap that motivates measurement, not as proof that one operating model outperforms another.
Instrument the transfer path
Choose one end-to-end workflow. Baseline cycle time, queue age, handoff failures, rework, quality, cost and outcome volume. Then identify where individual AI-assisted speed changes the constraint. If one analyst finishes faster but work waits for approval, data access or downstream review, the local gain has not yet increased system throughput.
Run a bounded redesign with the same outcome definition. Include model cost, verification time, exceptions and displaced work. Compare with the baseline over a declared window; do not use token consumption or active users as success measures.
The immediate decision is to fund scale only when a local productivity gain survives the complete workflow and appears in a business outcome without a hidden quality or labour transfer.