Our selected research collection examines AI at work, implementation conditions, economic value and risk. Each note explains what the source found, how we interpret its relevance to a business decision and where the evidence has limits.
A controlled experiment, a workplace study and a management survey answer different questions. We look at the tasks, participants, tools and measured outcomes before drawing a practical implication. Conflicting findings are useful when they reveal where the circumstances differ.
This is a selected reading collection, not a systematic review or evidence of Versta customer results. Research helps us assess what to investigate in a client case; it does not supply a ready-made savings percentage.
AI assistance increased resolved issues per hour by 15% on average across 5,172 customer-support agents. Gains varied by experience, with limited benefits and some quality declines among the strongest workers.
Quarterly Journal of EconomicsRead the analysis ↗An experiment with 758 consultants found benefits on tasks within the tested AI capabilities. On a task beyond that boundary, AI users were less likely to reach the correct answer.
Harvard Business SchoolRead the analysis ↗In a study of 453 professionals, half received access to ChatGPT. Those with access completed short writing tasks faster and produced better-rated outputs. Average task time fell 40%; rated quality rose 18%.
ScienceRead the analysis ↗A six-month experiment across 66 firms found roughly two fewer hours of email work per week among treated workers using an integrated AI tool. Broader changes in task quantity or composition were not detected.
NBERRead the analysis ↗In product-innovation exercises with 776 P&G professionals, individuals using AI matched the performance of teams without AI. The experiment also explored expertise sharing across professional boundaries.
NBERRead the analysis ↗Representative US surveys found rapid adoption, while regular use and the share of work assisted were smaller than the headline number of people who had tried AI.
NBERRead the analysis ↗Danish work reorganised and AI-related tasks emerged, while average earnings and recorded hours showed no detectable change large enough to exceed 2% in the studied period.
Humlum & VestergaardRead the analysis ↗Combined evidence from 4,867 developers estimated more completed tasks with AI code-completion tools. Individual experiments were noisy and effects varied.
Three workplace experimentsRead the analysis ↗Sixteen experienced developers working on 246 issues took 19% longer with the studied tools, despite believing they had worked faster.
METRRead the analysis ↗Newer developer data suggested improvement, but participant selection, task selection and concurrent agent use made the effect estimate unreliable.
METRRead the analysis ↗New technologies require complementary investment in processes, skills and organisational assets. These investments affect when productivity benefits appear and how they are measured.
American Economic JournalRead the analysis ↗Aggregate benefits depend on improvement within affected tasks and the share of activity those tasks represent. Estimates are sensitive to assumptions about capabilities and exposure.
NBER · Daron AcemogluRead the analysis ↗Workers reported benefits to performance and job enjoyment, alongside concerns about work intensity, data use and inequality.
OECDRead the analysis ↗A 2024 survey of over 5,000 SMEs in seven countries found reported performance benefits more common than increased revenue. Skills, legal concerns and preparation remained relevant.
OECDRead the analysis ↗Human–AI combinations did not consistently outperform the better of the human or AI alone. Results differed between decision tasks and content creation.
Nature Human BehaviourRead the analysis ↗Interviews with 65 AI/ML practitioners highlighted unclear goals, unsuitable data, technology-led decisions, missing infrastructure and tasks beyond available capabilities.
RANDRead the analysis ↗The AI Risk Management Framework organises governance, context, measurement and risk management across the lifecycle of an AI system.
NISTRead the analysis ↗NIST’s generative-AI profile addresses unreliable content, privacy, security and problematic configurations of people and AI systems.
NISTRead the analysis ↗Task duration, measured against human completion time, offers a lens on agent capability. Performance improved rapidly in the tested software-task suite.
METRRead the analysis ↗The annual index documents advancing capabilities, uneven performance and gaps in responsible-AI measurement and transparency.
Stanford HAIRead the analysis ↗