الذكاء الاصطناعي يدخل المؤسسة. ما الذي يتغير فعلاً؟

Staggered workplace rollout; observational causal design. 5,172 agents · One business-software firm; mainly Philippines

Issues resolved per hour rose 15% on average; gains varied by experience.

Limits: One firm and tool; not randomized; not enterprise ROI.

What to measure: Measure accepted output, quality and worker heterogeneity.

Preregistered randomized online experiment. 453 participants · Incentivized professional writing tasks

Time fell 40%; assessed quality rose 18%.

Limits: Short tasks; no whole-firm cost or sustained-use result.

What to measure: Separate task speed, quality and total workflow outcomes.

Randomized field experiment. 758 participants · Consulting tasks with GPT-4

Within-frontier tasks improved speed and assessed performance; effects depend on task suitability.

Limits: Selected tasks and one employer; working-paper versions change.

What to measure: Evaluate task boundaries and human review rather than access alone.

Preregistered randomized field experiment. 776 participants · Product innovation; individuals and teams, with/without AI

Individuals with AI matched teams without AI on evaluated performance.

Limits: One company and bounded innovation exercise; not proof teams should disappear.

What to measure: Assess role boundaries and coordination, not only individual use.

Task-randomized controlled trial. 16 developers; 246 tasks · Mature repositories; February–June 2025 tools

AI-allowed tasks took 19% longer despite perceived speedup.

Limits: Small selected sample; period-specific tools and mature projects.

What to measure: Compare measured performance with participant estimates.

Follow-up randomized study; methodological disclosure. 57 developers; 800+ tasks; 143 repositories · Follow-up begun August 2025

Authors judge newer productivity estimates unreliable because of selection and time-measurement problems.

Limits: Nonparticipation, selective tasks, changed pay and concurrent agents.

What to measure: Record nonresponse, mode and task-mix changes; avoid timeless technology claims.

Six-month randomized field experiment. 7,137 workers · Cross-industry; integrated email, document and meeting tools

Email and document behaviors changed; meeting time did not significantly change.

Limits: Specific revision; user-subgroup and intention-to-treat estimates differ; vendor researchers.

What to measure: Measure shared workflows and adoption denominators, not only individual time saved.

Stratified firm surveys. Almost 6,000 respondents · US, UK, Germany, Australia

Widespread reported use coexisted with limited reported employment/productivity effects; expectations were stronger.

Limits: Self-reported impacts, four non-GCC economies; later revision differs.

What to measure: Keep adoption, realized effects and expectations separate.

Economic model and historical empirical analysis. Not a company AI trial · R&D, software and hardware investment

Complementary intangible investment can delay measured productivity gains.

Limits: Model/historical mechanism; not proof any AI investment pays back.

What to measure: Track complementary investment, timing and delayed outcomes.

Instrument development and cross-sectional survey. 143 responses · Multiple industries

AI capability combines technical, human and organizational resources associated with creativity/performance.

Limits: Self-selection and same-source reporting; association is not causal; pre-GenAI.

What to measure: Treat capability as multidimensional; do not equate software ownership with capability.

Structured management-practice interviews and linked performance analysis. 732 firms · US, France, Germany, UK

Practice scores were associated with several firm-performance measures.

Limits: Observational, historical manufacturing sample; not an AI index.

What to measure: Use trained interviewers and operational examples rather than executive self-labels.

Methodological handbook. Not applicable · Country indices; methodological principles transferable

Normalization, weighting, aggregation and sensitivity choices affect interpretation and rankings.

Limits: Does not validate Mayden dimensions, cutoffs or weights.

What to measure: Predeclare construct, anchors, weighting and uncertainty/sensitivity.

Methodological primer. Not applicable · Health, social and behavioral research

Separates item development, scale development and evaluation.

Limits: General guidance; reflective-scale procedures may not fit a formative index.

What to measure: Test comprehension, content validity, reliability and criterion validity.

Voluntary risk-management framework. Not applicable · Cross-sector governance

Organizes risk management around Govern, Map, Measure and Manage.

Limits: Normative guidance, not evidence that compliance causes productivity; version-specific.

What to measure: Assess ownership, measurement, incidents and lifecycle controls.

Preregistered systematic review/meta-analysis. 106 experiments; 370 effect sizes · Studies published January 2020–June 2023

Average combinations underperformed the better standalone agent; task type moderated results.

Limits: Historical systems; heterogeneous tasks; does not measure whole enterprises.

What to measure: Compare feasible human-only, AI-only and combined workflows.

Conceptual measurement-model analysis. · Structural equation measurement

Consequences of indicator selection and consistency depend on whether indicators cause or reflect a construct.

Limits: Conceptual argument, not validation of an AI index.

What to measure: Treat the public composite as a designed disclosure index; do not use internal consistency alone to validate it.

Methodological analysis with three case studies. · Composite indicators

Nominal weights can differ from statistical importance because indicators vary and correlate.

Limits: Case-study methods do not establish optimal weights for Mayden; small selected cohorts limit statistical estimation.

What to measure: Publish nominal weights and deterministic sensitivity; avoid claiming equal weights mean equal effective importance.

Methodological manual. · Health measurement

Content validity concerns relevance, comprehensiveness and comprehensibility, including intended respondents' understanding.

Limits: PROM guidance; clinical thresholds and grading do not automatically transfer to organisations.

What to measure: Test interpretation, omitted content and response processes before statistical scoring.

Reporting-guideline development. · Research synthesis

Sixteen reporting items improve transparency of searches; the checklist is for reporting, not a complete conduct standard.

Limits: Adapting a search log does not make a corporate review systematic or exhaustive.

What to measure: Log platform, exact query, date, limits, supplementary routes, deduplication and retrieved counts.

Reporting-guideline development. · Research synthesis reporting

Provides a 27-item checklist and flow diagrams for transparent systematic-review reporting.

Limits: Checklist adoption alone is not proof of complete searching or low bias.

What to measure: Call the current synthesis a scoped rapid evidence review; reserve systematic-review claims for a documented eligible design.

Normative data-model specification. · Domain-independent provenance

Represents entities, activities, agents and derivations so information can be traced across transformations.

Limits: A provenance graph records derivation; it does not establish factual truth.

What to measure: Link fetched artifacts, extracted spans, claims, reviews and releases with stable IDs and hashes.

Original statistical-agency survey analysis. · U.S. business AI use

The article describes a November 2025 question change from production-related AI to any business function and measures use over a recent two-week period.

Limits: Exact analytic sample not reported in inspected article; wording change limits historical comparisons.

What to measure: Keep a stable reference period and bridge instrument revisions before publishing adoption trends.

Observational linked enterprise-account and usage analysis. More than 1,500 organisations and 17 million messages at the six-month adoption horizon, as described in the abstract · Account records through March 2026

Adoption differs in speed, breadth and purpose; account use is linked to roles and firm characteristics.

Limits: Preprint; inspected abstract only. Selected customers of one product, observational data and incomplete enterprise-tool coverage.

What to measure: Separate available seats, active use, task purpose and role coverage; triangulate surveys with permitted telemetry.

Series of randomized field experiments. Millions of users and products; exact pooled total not stated in inspected abstract · Seven workflows, experiments in 2023–2024

Sales effects differ across workflow applications, with some showing no effect.

Limits: Preprint version 6; inspected abstract only. One platform and bounded sales outcomes do not establish enterprise profitability.

What to measure: Record baseline, task, outcome, comparison and period; keep sales effects separate from capability scoring.

Linked adoption surveys and administrative records; observational difference-in-differences. Approximately 25,000 complete responses per survey round; latest round spans about 7,000 workplaces · November–December 2023 and 2024 surveys; labour records through December 2024

Task reorganisation coexists with estimates near zero for earnings and recorded hours.

Limits: Working paper; nonrandom adoption, identification assumptions, short horizon; no post-adoption firm financial analysis.

What to measure: Separate task change, capacity, pay, employment and profit outcomes.

Provider telemetry with model-assisted classification. 998,481 sampled API tool calls; separate 500,000-session analysis; other denominators differ by exhibit · Late 2025–early 2026

Autonomy and oversight vary with product context and user experience.

Limits: Single provider; model classifications; API requests cannot be joined into complete sessions; evaluations may resemble production.

What to measure: Track action authority, interventions, reversibility and session boundaries separately.

Nonrandom post-implementation surveys, adjusted regressions and interviews. 1,716 licensed respondents of 3,549 invited; 2,535 comparison respondents of 9,300 invited; 19 interviews · Survey fieldwork February–March 2025

Respondents described task benefits and verification overhead.

Limits: No pre-trial baseline or objective outcome metrics; self-selection and nonresponse; regression does not remove all bias.

What to measure: Collect baselines and observed task metrics before rollout.

Stratified randomised, blinded-rating trial. 50 physicians, 25 per arm; 244 completed case responses across six vignettes · November–December 2023; one-hour sessions

Access to a chatbot did not significantly improve the primary diagnostic-reasoning score.

Limits: Small vignette study, one model, limited prompting instruction; not patient outcomes.

What to measure: Test the combined human-and-AI workflow against usual practice.

Separate employer telephone and worker online-panel surveys. 5,334 workers; 2,053 firms · Main fieldwork mid-January to mid-February 2022

Sector-specific surveys capture implementation and workers' experiences.

Limits: Perceptions and associations; online-panel coverage; predates mass generative-AI adoption; matched worker-employer pairs are not established in the inspected method.

What to measure: Ask operators as well as leadership; define the technology and reference period.

Professional-body online sentiment surveys. Over 2,200 Q1 respondents; separate skills survey over 2,000 · Six AI questions in Q1 2025 Global Construction Monitor; separate Q2 skills module

Respondents distinguish pilots, regular process use and organisation-wide adoption.

Limits: Sentiment sample; UK accounts for 48% of Q1 responses; MEA is not a GCC estimate; two survey bases must not be pooled.

What to measure: Keep project, asset, organisation and respondent scope explicit.

Synthetic interactive agent benchmark with repeated trials. 115 retail tasks and 50 airline tasks · Version 1; synthetic users, policies and databases

Repeated success differs from success in a single run; final-state tests can miss policy violations.

Limits: Simplified environments, simulated users and annotation assumptions; not production business outcomes.

What to measure: Measure consistency and policy compliance in addition to task completion.

Expert-authored task benchmark with pairwise deliverable evaluation. 1,320 tasks across 44 occupations; public gold subset 220 tasks · Version 1; professional deliverables

Capabilities can be assessed through work products and expert comparisons.

Limits: Precisely specified, mainly one-shot tasks; limited tacit context and interaction; capability and modelled savings do not establish realised enterprise ROI.

What to measure: Separate model evaluation from deployment and financial evidence.

Agent benchmark re-audit and safeguard experiments. 50 airline and 115 retail tasks; three runs for tau-bench experiments · Version 1; three model families; 30-turn cap

Authors identify annotation problems and test safeguards around state-changing actions.

Limits: Simulated confirmation; preprint; gains are not universal in table 2 and narrative overstates consistency.

What to measure: Review test labels, action consequences and regressions before deployment.

Retrospective observational evaluation; authors also describe a quasi-prospective approach.. 1.04 million test chest X-ray scans · January 2022–January 2024; low-prevalence screening

Reported AUROC 98.51%; overall false-negative rate 1.57% at the optimal threshold.

Limits: Potential 80% reporting reduction is calculated from test classifications, not demonstrated prospective savings. Zero TB false negatives follows annotation review/correction. Lower external-dataset AUROC; sputum confirmation restricted to radiologist-flagged cases.

What to measure: Separate diagnostic evaluation, projected workload and observed operating change.