AI use is broad but shallow. This paper turns adoption and change research into four NKK models: an Adoption Ladder, Four Locks diagnostic, Task Frontier Map and Shadow-to-Sanctioned loop, with metrics.
Written for: CIOs, CHROs, heads of AI, transformation and change leads
12 chapters, 59 cited sources, 26 models and charts. The first 2 chapters are open; the rest unlock with the form.
of employed adults used generative AI for work in the past week (Q2 2026; 28.2% in Q3 2024)
Most enterprises have an AI conversion problem, not an adoption problem: access has become use, but use has not become changed work, and changed work has not become captured value. In the St. Louis Fed's tracker, 39.2% of employed US adults aged 18 to 64 used generative AI for work in the past week in Q2 2026, yet its latest readings show only about 6.3% of work hours AI-assisted and about 2.2% saved, on self-reports the authors call "inherently approximate" [1]. Fewer than 3% of tasks have adoption rates above 50% [2].
The reasons are mostly organisational, and the evidence is mostly correlational. In AI-adopting organisations, 78% of employees are frequent users where they strongly agree their manager supports AI use, against 44% where they do not [3]. Only 13% of AI users say they are rewarded for reinventing work with AI even if results fall short [4]. Field experiments show task-level gains, but also a "jagged frontier" where AI users were 19 percentage points less likely to be correct [5].
This paper offers four NKK models, each a synthesis of cited prior art and none empirically tested: the Adoption Ladder, the Four Locks diagnostic (Fit, Skill, Safety, Signal), the Task Frontier Map (where to delegate, pair, pilot or protect) and the Shadow-to-Sanctioned loop, with a six-link metric chain. They let a leader find each team's binding constraint, spend there first, and measure conversion, not activity.
Ask how many people use AI at work and the answers range from 15% to 88%. Every figure can be correct, because each measures something different. The Stanford AI Index reports 88% organisational adoption in 2025, compiled from McKinsey's State of AI survey, where "adoption" means using AI in at least one business function; the AI Index itself says the survey data are self-reported and should be "viewed as directional rather than comprehensive" [6]. The US Census Bureau's firm survey, which asks whether a business used AI in any business function in the last two weeks, found 19.8% in early May 2026 and 23.8% in the most recent period in its data file [7]. Gallup's probability-based panel finds 15% of US employees using AI daily, 30% a few times a week or more, and 52% a few times a year or more, in cumulative bands [8].
| Criterion | What is measured | Latest figure | Caveat |
|---|---|---|---|
| Organisations (AI Index) | Use of AI in at least one business function, 2025 | 88% [6] | Self-reported; the AI Index calls it directional |
| US firms (Census BTOS) | Firm used AI in any business function in the last two weeks | 23.8% in the latest period in the file [7] | Question wording changed in November 2025 [9] |
| US employees (Gallup) | Use in own role, by frequency (data labelled May 2026) | 15% daily; 30% weekly or more; 52% a few times a year or more [8] | Bands are cumulative |
| Employed adults (St. Louis Fed) | Used generative AI for work in the past week, Q2 2026, ages 18 to 64 | 39.2% [1] | Different question from the 45% of workers using AI for their jobs in May 2026 [2] |
| Frontline staff (BCG) | Regular use among frontline employees, 2025 | 51%, stalled [10] | 11-country survey of more than 10,600 people; vendor-run |
| Tasks (St. Louis Fed) | Share of tasks with adoption above 50% | Fewer than 3% [2] | Task-level measure; NKK's reading: the most telling of the six |
Census changed its question in November 2025, from AI use "in producing goods or services" to use "in any business function", so firm-level figures from before and after are not comparable [9]. The consistent message across independent sources is breadth outrunning depth. The St. Louis Fed's own post carries the section heading "AI Adoption Is Widespread but Shallow" [2]. Adoption has also been fast: work adoption of generative AI has been as fast as the personal computer, measured from first mass-market launch [11].
of employed adults used generative AI for work in the past week (Q2 2026; 28.2% in Q3 2024)
Second, there is a supporting NKK concept, the capture gap: the distance between the work AI touches and the value an organisation can bank. The tracker's hours saved are roughly a third of hours assisted (2.2% against 6.3%, NKK arithmetic on approximate figures) [1]. In Denmark, a study linking surveys of about 25,000 workers per round to employer records found no significant effect on earnings or recorded hours in any exposed occupation, with adopters in the May 2025 version reporting time savings of about 3% of work hours [12]; the March 2026 revision rules out effects on earnings or recorded hours larger than 2% two years after ChatGPT's launch [13]. In Microsoft's field experiment across 66 firms and 7,137 knowledge workers, the 80% of treated workers who used the tool spent about two fewer hours on email each week, but the authors "do not detect shifts in the quantity or composition of workers' tasks" [14]. Saved minutes are not redeployed hours. Somebody has to decide what the time is for.
These are early numbers on early tools in specific settings, and the Danish and Microsoft studies cover different populations and periods. They do not show that AI cannot raise productivity. They show that the step from use to value is not automatic.
Every figure and regulatory statement in this piece traces to a numbered source below. Links open the original publisher in a new tab.
The EU lowered the AI-literacy standard on 27 July 2026. Adoption data points to managers, permission and peers as well as courses, so we set out an original champions model, with the evidence for and against it.
Self-reported benefits are common, yet P&L-visible value is a minority. Evidence on what AI delivers, why gains stall, and three NKK models for conversion, baselines and attribution.
Pilots are common and scaled agent use is rare. NKK's Oversight Map, Three-Meter Funding Model, Canvas and Maturity Ladder turn the operating-model research into owners, budgets and a scorecard.
