Appendix: Evidence for use in training

Why task-specific scenarios?

Licensing an AI tool does not automatically embed its use in everyday work. In BCG’s 2025 survey of more than 10,600 employees across 11 countries and regions, more than three-quarters of leaders and managers use generative AI several times a week, while regular use among frontline employees remains at 51%. Only about a third of employees say they have received sufficient training. Regular use is significantly higher among those who have received at least five hours of training and have access to in-person training and coaching. [1]

There is also a scaling gap between usage and enterprise value. In Microsoft’s 2025 study of 31,000 knowledge workers across 31 countries, 82% of leaders see the year as critical for rethinking strategy and operations; 51% of managers say AI training or upskilling will become a team responsibility within five years. [192] [193] In McKinsey’s survey of 1,993 respondents, regular AI use in at least one function has reached 88%, yet only about a third are scaling their programs across the organization, and 39% report an impact on enterprise EBIT. Organizations reporting high value from AI are 2.8 times more likely than others to report fundamentally redesigning workflows. [2]

Training should therefore be more than a general explanation of “What is AI?” Employees need to see how to prepare inputs for their own roles, at which step to use AI, which criteria to use to check the output, and when to hand over to a human. MIT NANDA’s 2025 report, which examines more than 300 publicly disclosed initiatives, 52 organizational interviews, and responses from 153 senior leaders together, states that 95% of organizations report no measurable return, while only 5% of integrated AI pilots deliver value. The report identifies fragile workflows, a lack of contextual learning, and misalignment with daily operations as the main problems with most custom enterprise tools; it also notes that the findings are directional because category samples and definitions of success vary. [194]

This approach also makes sense from a legal perspective. Article 4 of the EU AI Act requires providers and deployer organizations to take measures to support the AI literacy of their employees and people operating the system on their behalf. The European Commission explains that the obligation began to apply on 2 February 2025; that training should be tailored to roles, risks, technical knowledge, experience, and the context of use; and that simply having people read the instructions for use will be insufficient in most cases. [195] [4] A task-specific scenario is therefore a practical training unit that supports both adoption and measurable, responsible use.

Evidence of productivity gains

In Noy and Zhang’s preregistered online experiment published in Science, 453 college-educated professionals were assigned occupation-specific, incentivized, mid-level writing tasks. Half of the participants were randomly given access to ChatGPT. Access reduced average task time by 40% and increased output quality by 18%. Workers with weaker skills benefited more. Those given access in the experiment were twice as likely to report using AI in their actual work two weeks later, and 1.6 times as likely two months later. This is a strong effect for this type of writing task and under short-term experimental conditions; it cannot be generalized to all office work. [196]

Brynjolfsson, Li and Raymond’s 2025 QJE study examined the phased rollout of a GPT-3-based assistant that suggested responses in real time to 5,172 customer support agents at a Fortune 500 company. Agents could reject or edit the suggestions. AI assistance increased the number of chats resolved per hour by an average of 15%. The increase reached approximately 30% among lower-skilled and less experienced agents, and 36% in the lowest skill quintile; no significant productivity effect was observed among the most highly skilled agents, and there was an indication of a small decline in conversation quality. The result shows that the same tool can deliver different value to different employees. [117]

In the field study conducted by Dell’Acqua and colleagues with BCG, 758 knowledge workers were randomly assigned to conditions with no AI, GPT-4, or GPT-4 plus an explanation of prompt engineering. Across 18 realistic consulting tasks within AI’s capability boundary, users completed 12.2% more tasks, finished tasks 25.1% faster and produced higher-quality solutions. However, on a complex management task outside that boundary, AI users were 19% less likely than the control group to produce the correct solution. Training should frame the question not as “AI helps with every task,” but as “Which tasks is it reliable for?” [3]

In Peng and colleagues’ controlled Copilot experiment with 95 professional programmers, participants wrote a JavaScript HTTP server; 45 people had access to Copilot, and 50 were in the control condition. The control group could use the internet and Stack Overflow. Among those who completed the task, the average time was 71.17 minutes in the Copilot group and 160.89 minutes in the control group; the difference represented a 55.8% reduction in time, with a 95% confidence interval of 21%-89%. Although the success rate was 7 percentage points higher in the treatment group, the difference was not statistically significant. A single standardized task does not represent the full scope of real-world team software development. [197]

In METR’s 2025 randomized experiment, 16 experienced open-source developers worked on 246 real issues in large repositories to which they regularly contributed. Tasks took an average of two hours, and AI use was randomly allowed or prohibited. In this context, where developers mostly used Cursor Pro with Claude 3.5/3.7 Sonnet, allowing AI increased completion time by 19%. Developers expected a 24% speedup beforehand and a 20% speedup after the experiment. METR specifically notes that this result applies to experienced developers, familiar repositories and early 2025 tools, and cannot be generalized to other areas. [198]

METR’s 2026 update reports that measurement with newer tools was distorted by selection effects. In the subsequent study involving 57 developers, 143 repositories and more than 800 tasks, some developers did not want to work on tasks without AI; 30%-50% said they had not submitted some tasks for this reason. The estimated speedup was -18% in the subset of returning participants and -4% among new participants; both confidence intervals included zero. This result is a reminder that tools change over time and that task selection affects measurement. [199]

Risks and limitations

NIST’s generative AI risk profile defines “confabulation” as confident but false or erroneous content that misleads the user. The same profile identifies data privacy risks as the leakage, unauthorized use or disclosure of personal or sensitive data, or the breakdown of de-identification. It recommends testing claims using empirical methods before deployment, verifying sources in outputs and regularly evaluating the system under conditions similar to those in which it will be used. Daily scenarios should therefore include source checks, a second review and human approval for high-impact decisions. [12]

According to KVKK’s November 2025 guidance, Law No. 6698 applies if personal data is processed at any stage of the generative AI lifecycle. Even if user input contains no personal data, information from model training may generate personal data in the output. KVKK also emphasizes separate processing conditions and risks for special categories of personal data. Its March 2026 workplace notice states that using publicly available third-party tools without a clear organizational policy makes monitoring and management more difficult. [200] [201]

HBR defines “workslop” as AI-generated work that appears polished but requires little effort, shifting cognitive work to the recipient and costing them time and effort. Fluent text does not necessarily advance the task or mean that it is correct. Training scenarios should teach not only how to produce outputs quickly, but also how to reduce the work the recipient will have to redo and meet quality criteria. [202]

Corporate data protection varies from tool to tool. OpenAI states that it does not use business data for model training by default and provides access, retention and encryption controls; this is not a blanket guarantee for all providers. An organization should not put employee data into a publicly available tool without verifying the contract, data transfers, retention period, access roles and administrator settings. [203]

Evidence points

• In the BCG 2025 survey, regular GenAI use among frontline employees is 51%. [1]
• BCG links at least five hours of training and access to coaching with regular use. [1]
• The Microsoft 2025 Work Trend Index covers 31,000 knowledge workers in 31 countries. [193]
• In the Microsoft study, 82% of leaders see 2025 as critical for rethinking strategy and operations. [192]
• In the Microsoft report, 51% of managers see AI training as a team responsibility within five years. [193]
• In McKinsey 2025, regular AI use in at least one function is 88%. [2]
• In McKinsey 2025, the proportion reporting an organization-wide EBIT impact is 39%. [2]
• According to the Stanford AI Index, organizational AI use rose from 55% in 2023 to 78% in 2024. [204]
• The MIT NANDA report states that 95% of organizations report no measurable returns, but also notes that the findings are indicative. [194]
• In Noy and Zhang, writing task completion time decreased by 40%, while quality increased by 18%. [196]
• In Brynjolfsson, Li and Raymond, customer support productivity increased by an average of 15%. [117]
• In the Dell’Acqua field experiment, tasks within the frontier saw 12.2% more tasks completed and 25.1% less time spent. [3]
• In the Dell’Acqua experiment, the likelihood of a correct solution on a management task outside the frontier decreased by 19%. [3]
• The Copilot experiment by Peng and colleagues involved 95 programmers, with completion times of 71.17 minutes versus a baseline of 160.89 minutes. [197]
• In METR 2025, completion time increased by 19% across 16 experienced developers and 246 real-world tasks. [198]
• The METR 2026 update reported that the newer measurement was unreliable due to selection effects. [199]
• Article 4 of the EU AI Act has applied AI literacy measures since 2 February 2025. [195] [4]
• KVKK states that Law No. 6698 applies at all stages of the GenAI lifecycle if personal data is processed. [201]
• NIST classifies incorrect but confident AI output as confabulation. [12]