Designing AI Operating Models Around Humans

Narrative

Practitioner evidence from 2024 to 2026 points to a split picture. Google's enterprise RCT found about a 21% reduction in task time for a complex internal task, and IBM's internal study also reports net productivity gains for many developers. But DORA's 2024 report found that rising AI adoption was associated with lower delivery stability, lower throughput, and less time spent on valuable work, which RedMonk argues is what happens when firms optimise code generation rather than the real bottleneck in the delivery system.

The strongest cautionary evidence comes from settings where people must supervise AI inside mature systems rather than greenfield tasks. METR's 2025 field experiment found experienced open-source developers were 19% slower with frontier tools, despite expecting a 24% speed-up, and another 2025 study found that AI-assisted output can shift maintenance burden onto core developers, increasing review load while reducing their own original-code productivity. A separate open-source study found project productivity gains, but also a 41.6% rise in integration time, which is a direct measure of coordination cost rather than coding speed.

Workplace reports now give names to the hidden human tax. Harvard Business Review and BetterUp describe "workslop", polished AI output that transfers effort to colleagues, while Glean's 2026 findings describe "botsitting", with workers spending 6.4 hours a week supplying context, checking outputs, and fixing errors. Developer surveys line up with that pattern: Stack Overflow's 2025 results show AI usage has become mainstream, yet distrust has risen, and JetBrains' ecosystem data shows many teams still want humans to keep hold of code review and testing.

The operating-model implication is consistent across these sources. DORA's newer framing around burnout, friction, and perceived value, the SPACE study's emphasis on team support, and the Stack Overflow and JetBrains survey results all point away from top-down usage mandates and towards selective deployment, explicit review boundaries, and workflow redesign. The evidence is strongest where organisations treat AI as a system component that must fit human attention, judgement, and collaboration, not as a token-maximising substitute for them.