{ "version": "PE-2026.10-v5", "date": "2026-10-10", "researchCutoff": "2026-10-10", "publicationDate": "2026-10-11", "publicationPolicy": "Sector research without individual company scores. Attributed examples do not imply participation, endorsement or independent operating assurance. Confidential company assessment remains in development.", "counts": { "entityRecords": 80, "sectors": 8, "markets": 5, "themeCells": 720, "documentedThemeCells": 201 }, "dimensions": [ { "id": "ai-core", "name": "AI Core", "ar": "الذكاء الاصطناعي في صميم العمل", "question": "Where has AI changed a material operating workflow?" }, { "id": "workforce", "name": "Augmented Workforce", "ar": "القوى العاملة المعززة", "question": "How do people learn, use and supervise AI in their work?" }, { "id": "ecosystem", "name": "Ecosystem", "ar": "المنظومة المتكاملة", "question": "How do partners and shared capabilities support internal AI operations?" }, { "id": "interfaces", "name": "Interfaces & Automation", "ar": "الواجهات والأتمتة", "question": "How does AI change the way customers and employees interact with services?" }, { "id": "lineage", "name": "Decision Lineage", "ar": "تتبّع القرارات", "question": "Can consequential AI decisions be traced to their inputs, versions and authority?" }, { "id": "experimentation", "name": "Experimentation", "ar": "التجريب", "question": "How are AI use cases tested, compared, stopped and improved?" }, { "id": "autonomy", "name": "Autonomy & Operating Model", "ar": "الاستقلالية ونموذج التشغيل", "question": "What may AI execute, and how do people supervise, reverse or stop it?" }, { "id": "knowledge", "name": "Knowledge", "ar": "المعرفة", "question": "How does AI make organizational knowledge available and reusable?" }, { "id": "governance", "name": "Governance", "ar": "الحوكمة", "question": "Who owns AI decisions, risks and lifecycle controls?" } ], "sectors": [ { "name": "Aviation & Logistics", "entityCount": 10, "claimCount": 37, "includedClaimCount": 32, "heldClaimCount": 5, "themeCoverage": [ { "dimension": "ai-core", "count": 7 }, { "dimension": "workforce", "count": 3 }, { "dimension": "ecosystem", "count": 5 }, { "dimension": "interfaces", "count": 3 }, { "dimension": "lineage", "count": 1 }, { "dimension": "experimentation", "count": 2 }, { "dimension": "autonomy", "count": 0 }, { "dimension": "knowledge", "count": 2 }, { "dimension": "governance", "count": 2 } ], "sourceDates": { "2025onward": 14, "before2025": 9, "undated": 9 }, "reportSlug": "logistics-public-evidence" }, { "name": "Banking", "entityCount": 10, "claimCount": 63, "includedClaimCount": 52, "heldClaimCount": 11, "themeCoverage": [ { "dimension": "ai-core", "count": 7 }, { "dimension": "workforce", "count": 8 }, { "dimension": "ecosystem", "count": 4 }, { "dimension": "interfaces", "count": 7 }, { "dimension": "lineage", "count": 2 }, { "dimension": "experimentation", "count": 3 }, { "dimension": "autonomy", "count": 2 }, { "dimension": "knowledge", "count": 4 }, { "dimension": "governance", "count": 7 } ], "sourceDates": { "2025onward": 47, "undated": 5 }, "reportSlug": "banking-public-evidence" }, { "name": "Energy & Utilities", "entityCount": 10, "claimCount": 36, "includedClaimCount": 30, "heldClaimCount": 6, "themeCoverage": [ { "dimension": "ai-core", "count": 7 }, { "dimension": "workforce", "count": 6 }, { "dimension": "ecosystem", "count": 5 }, { "dimension": "interfaces", "count": 3 }, { "dimension": "lineage", "count": 0 }, { "dimension": "experimentation", "count": 3 }, { "dimension": "autonomy", "count": 1 }, { "dimension": "knowledge", "count": 3 }, { "dimension": "governance", "count": 1 } ], "sourceDates": { "undated": 8, "2025onward": 16, "before2025": 6 }, "reportSlug": "energy-public-evidence" }, { "name": "Healthcare", "entityCount": 10, "claimCount": 32, "includedClaimCount": 27, "heldClaimCount": 5, "themeCoverage": [ { "dimension": "ai-core", "count": 6 }, { "dimension": "workforce", "count": 3 }, { "dimension": "ecosystem", "count": 3 }, { "dimension": "interfaces", "count": 4 }, { "dimension": "lineage", "count": 0 }, { "dimension": "experimentation", "count": 3 }, { "dimension": "autonomy", "count": 0 }, { "dimension": "knowledge", "count": 1 }, { "dimension": "governance", "count": 4 } ], "sourceDates": { "undated": 22, "before2025": 1, "2025onward": 4 }, "reportSlug": "healthcare-public-evidence" }, { "name": "Manufacturing & Materials", "entityCount": 10, "claimCount": 17, "includedClaimCount": 17, "heldClaimCount": 0, "themeCoverage": [ { "dimension": "ai-core", "count": 5 }, { "dimension": "workforce", "count": 2 }, { "dimension": "ecosystem", "count": 2 }, { "dimension": "interfaces", "count": 1 }, { "dimension": "lineage", "count": 0 }, { "dimension": "experimentation", "count": 3 }, { "dimension": "autonomy", "count": 0 }, { "dimension": "knowledge", "count": 2 }, { "dimension": "governance", "count": 2 } ], "sourceDates": { "before2025": 4, "2025onward": 11, "undated": 2 }, "reportSlug": "manufacturing-public-evidence" }, { "name": "Real Estate & Construction", "entityCount": 10, "claimCount": 27, "includedClaimCount": 20, "heldClaimCount": 7, "themeCoverage": [ { "dimension": "ai-core", "count": 4 }, { "dimension": "workforce", "count": 1 }, { "dimension": "ecosystem", "count": 4 }, { "dimension": "interfaces", "count": 3 }, { "dimension": "lineage", "count": 0 }, { "dimension": "experimentation", "count": 3 }, { "dimension": "autonomy", "count": 1 }, { "dimension": "knowledge", "count": 1 }, { "dimension": "governance", "count": 0 } ], "sourceDates": { "undated": 11, "2025onward": 4, "before2025": 5 }, "reportSlug": "realestate-public-evidence" }, { "name": "Retail & Consumer", "entityCount": 10, "claimCount": 18, "includedClaimCount": 17, "heldClaimCount": 1, "themeCoverage": [ { "dimension": "ai-core", "count": 3 }, { "dimension": "workforce", "count": 1 }, { "dimension": "ecosystem", "count": 3 }, { "dimension": "interfaces", "count": 5 }, { "dimension": "lineage", "count": 0 }, { "dimension": "experimentation", "count": 0 }, { "dimension": "autonomy", "count": 1 }, { "dimension": "knowledge", "count": 1 }, { "dimension": "governance", "count": 2 } ], "sourceDates": { "2025onward": 14, "before2025": 3 }, "reportSlug": "retail-public-evidence" }, { "name": "Telecommunications", "entityCount": 10, "claimCount": 42, "includedClaimCount": 29, "heldClaimCount": 13, "themeCoverage": [ { "dimension": "ai-core", "count": 5 }, { "dimension": "workforce", "count": 5 }, { "dimension": "ecosystem", "count": 3 }, { "dimension": "interfaces", "count": 4 }, { "dimension": "lineage", "count": 1 }, { "dimension": "experimentation", "count": 2 }, { "dimension": "autonomy", "count": 3 }, { "dimension": "knowledge", "count": 3 }, { "dimension": "governance", "count": 3 } ], "sourceDates": { "undated": 22, "2025onward": 7 }, "reportSlug": "telecom-public-evidence" } ], "studies": [ { "id": "L01", "title": "Generative AI at Work", "url": "https://doi.org/10.1093/qje/qjae044", "publisher": "Quarterly Journal of Economics", "publicationDate": "2025-02-04", "design": "Staggered workplace rollout; observational causal design", "population": "Customer-support agents", "sampleSize": "5,172 agents", "setting": "One business-software firm; mainly Philippines", "keyFinding": "Issues resolved per hour rose 15% on average; gains varied by experience.", "limitations": "One firm and tool; not randomized; not enterprise ROI.", "transferabilityToGCC": "Test language, work mix and experience locally.", "measurementImplication": "Measure accepted output, quality and worker heterogeneity.", "evidenceType": "peer-reviewed empirical" }, { "id": "L02", "title": "Experimental evidence on the productivity effects of generative artificial intelligence", "url": "https://doi.org/10.1126/science.adh2586", "publisher": "Science", "publicationDate": "2023-07-13", "design": "Preregistered randomized online experiment", "population": "College-educated professionals", "sampleSize": "453 participants", "setting": "Incentivized professional writing tasks", "keyFinding": "Time fell 40%; assessed quality rose 18%.", "limitations": "Short tasks; no whole-firm cost or sustained-use result.", "transferabilityToGCC": "Test equivalent Arabic/English tasks; do not transfer effect size.", "measurementImplication": "Separate task speed, quality and total workflow outcomes.", "evidenceType": "peer-reviewed experiment" }, { "id": "L03", "title": "Navigating the Jagged Technological Frontier", "url": "https://www.hbs.edu/ris/download.aspx?name=24-013.pdf", "publisher": "Harvard Business School working paper 24-013", "publicationDate": "2023-09", "design": "Randomized field experiment", "population": "BCG consultants", "sampleSize": "758 participants", "setting": "Consulting tasks with GPT-4", "keyFinding": "Within-frontier tasks improved speed and assessed performance; effects depend on task suitability.", "limitations": "Selected tasks and one employer; working-paper versions change.", "transferabilityToGCC": "Local role/task evaluations needed.", "measurementImplication": "Evaluate task boundaries and human review rather than access alone.", "evidenceType": "working-paper experiment" }, { "id": "L04", "title": "The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise", "url": "https://www.nber.org/papers/w33641", "publisher": "NBER working paper 33641; published Organization Science June 2026", "publicationDate": "2025-04", "design": "Preregistered randomized field experiment", "population": "Procter & Gamble professionals", "sampleSize": "776 participants", "setting": "Product innovation; individuals and teams, with/without AI", "keyFinding": "Individuals with AI matched teams without AI on evaluated performance.", "limitations": "One company and bounded innovation exercise; not proof teams should disappear.", "transferabilityToGCC": "Test expertise-sharing locally.", "measurementImplication": "Assess role boundaries and coordination, not only individual use.", "evidenceType": "original field experiment" }, { "id": "L05", "title": "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", "url": "https://metr.org/Early_2025_AI_Experienced_OS_Devs_Study-paper.pdf", "publisher": "METR", "publicationDate": "2025-07-10", "design": "Task-randomized controlled trial", "population": "Experienced open-source developers", "sampleSize": "16 developers; 246 tasks", "setting": "Mature repositories; February–June 2025 tools", "keyFinding": "AI-allowed tasks took 19% longer despite perceived speedup.", "limitations": "Small selected sample; period-specific tools and mature projects.", "transferabilityToGCC": "Do not generalize to GCC developers or current tools.", "measurementImplication": "Compare measured performance with participant estimates.", "evidenceType": "original experiment" }, { "id": "L06", "title": "We are Changing our Developer Productivity Experiment Design", "url": "https://metr.org/blog/2026-02-24-uplift-update/", "publisher": "METR", "publicationDate": "2026-02-24", "design": "Follow-up randomized study; methodological disclosure", "population": "Experienced open-source developers", "sampleSize": "57 developers; 800+ tasks; 143 repositories", "setting": "Follow-up begun August 2025", "keyFinding": "Authors judge newer productivity estimates unreliable because of selection and time-measurement problems.", "limitations": "Nonparticipation, selective tasks, changed pay and concurrent agents.", "transferabilityToGCC": "Monitor participation and task-selection changes in repeated panels.", "measurementImplication": "Record nonresponse, mode and task-mix changes; avoid timeless technology claims.", "evidenceType": "original follow-up methods report" }, { "id": "L07", "title": "Shifting Work Patterns with Generative AI", "url": "https://www.nber.org/system/files/working_papers/w33795/revisions/w33795.rev1.pdf", "publisher": "NBER working paper 33795, revision 1", "publicationDate": "2025-05-06", "design": "Six-month randomized field experiment", "population": "Knowledge workers", "sampleSize": "7,137 workers", "setting": "Cross-industry; integrated email, document and meeting tools", "keyFinding": "Email and document behaviors changed; meeting time did not significantly change.", "limitations": "Specific revision; user-subgroup and intention-to-treat estimates differ; vendor researchers.", "transferabilityToGCC": "Coordination mechanisms need local testing.", "measurementImplication": "Measure shared workflows and adoption denominators, not only individual time saved.", "evidenceType": "working-paper experiment" }, { "id": "L08", "title": "Firm Data on AI", "url": "https://wfhresearch.com/wp-content/uploads/2026/03/Firm-Data-on-AI-16-February-2026.pdf", "publisher": "NBER working paper 34836; author-hosted February version", "publicationDate": "2026-02", "design": "Stratified firm surveys", "population": "Senior executives", "sampleSize": "Almost 6,000 respondents", "setting": "US, UK, Germany, Australia", "keyFinding": "Widespread reported use coexisted with limited reported employment/productivity effects; expectations were stronger.", "limitations": "Self-reported impacts, four non-GCC economies; later revision differs.", "transferabilityToGCC": "Context, not a GCC prevalence estimate.", "measurementImplication": "Keep adoption, realized effects and expectations separate.", "evidenceType": "original survey" }, { "id": "L09", "title": "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies", "url": "https://www.nber.org/papers/w25148", "publisher": "NBER; American Economic Journal: Macroeconomics 13(1), 2021", "publicationDate": "2021-01", "design": "Economic model and historical empirical analysis", "population": "Historical technology/intangible investment series", "sampleSize": "Not a company AI trial", "setting": "R&D, software and hardware investment", "keyFinding": "Complementary intangible investment can delay measured productivity gains.", "limitations": "Model/historical mechanism; not proof any AI investment pays back.", "transferabilityToGCC": "Treat reorganization as a hypothesis with explicit costs.", "measurementImplication": "Track complementary investment, timing and delayed outcomes.", "evidenceType": "peer-reviewed theory and empirical analysis" }, { "id": "L10", "title": "Artificial intelligence capability: Conceptualization, measurement calibration, and empirical study on its impact on organizational creativity and firm performance", "url": "https://doi.org/10.1016/j.im.2021.103434", "publisher": "Information & Management 58(3)", "publicationDate": "2021-04-01", "design": "Instrument development and cross-sectional survey", "population": "Senior technology managers recruited through professional network", "sampleSize": "143 responses", "setting": "Multiple industries", "keyFinding": "AI capability combines technical, human and organizational resources associated with creativity/performance.", "limitations": "Self-selection and same-source reporting; association is not causal; pre-GenAI.", "transferabilityToGCC": "Revalidate content and measurement in each target language/context.", "measurementImplication": "Treat capability as multidimensional; do not equate software ownership with capability.", "evidenceType": "peer-reviewed survey and measurement" }, { "id": "L11", "title": "Measuring and Explaining Management Practices Across Firms and Countries", "url": "https://www.nber.org/papers/w12216", "publisher": "NBER; Quarterly Journal of Economics 122(4), 2007", "publicationDate": "2007-11", "design": "Structured management-practice interviews and linked performance analysis", "population": "Medium-sized manufacturers", "sampleSize": "732 firms", "setting": "US, France, Germany, UK", "keyFinding": "Practice scores were associated with several firm-performance measures.", "limitations": "Observational, historical manufacturing sample; not an AI index.", "transferabilityToGCC": "Adapt behavior-anchored interviewing; test transfer.", "measurementImplication": "Use trained interviewers and operational examples rather than executive self-labels.", "evidenceType": "peer-reviewed organizational measurement" }, { "id": "L12", "title": "Handbook on Constructing Composite Indicators: Methodology and User Guide", "url": "https://doi.org/10.1787/9789264043466-en", "publisher": "OECD / European Union / EC-JRC", "publicationDate": "2008-08-22", "design": "Methodological handbook", "population": "Composite-indicator builders", "sampleSize": "Not applicable", "setting": "Country indices; methodological principles transferable", "keyFinding": "Normalization, weighting, aggregation and sensitivity choices affect interpretation and rankings.", "limitations": "Does not validate Mayden dimensions, cutoffs or weights.", "transferabilityToGCC": "Adapt principles; disclose cohort and missingness.", "measurementImplication": "Predeclare construct, anchors, weighting and uncertainty/sensitivity.", "evidenceType": "measurement guidance" }, { "id": "L13", "title": "Best Practices for Developing and Validating Scales for Health, Social, and Behavioral Research: A Primer", "url": "https://doi.org/10.3389/fpubh.2018.00149", "publisher": "Frontiers in Public Health", "publicationDate": "2018-06-11", "design": "Methodological primer", "population": "Scale developers", "sampleSize": "Not applicable", "setting": "Health, social and behavioral research", "keyFinding": "Separates item development, scale development and evaluation.", "limitations": "General guidance; reflective-scale procedures may not fit a formative index.", "transferabilityToGCC": "Cognitive interviews and invariance testing needed.", "measurementImplication": "Test comprehension, content validity, reliability and criterion validity.", "evidenceType": "peer-reviewed measurement guidance" }, { "id": "L14", "title": "Artificial Intelligence Risk Management Framework (AI RMF 1.0)", "url": "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf", "publisher": "NIST", "publicationDate": "2023-01", "design": "Voluntary risk-management framework", "population": "Organizations developing/deploying/using AI", "sampleSize": "Not applicable", "setting": "Cross-sector governance", "keyFinding": "Organizes risk management around Govern, Map, Measure and Manage.", "limitations": "Normative guidance, not evidence that compliance causes productivity; version-specific.", "transferabilityToGCC": "Contextualize controls; no automatic local compliance claim.", "measurementImplication": "Assess ownership, measurement, incidents and lifecycle controls.", "evidenceType": "official framework" }, { "id": "L15", "title": "When combinations of humans and AI are useful: A systematic review and meta-analysis", "url": "https://doi.org/10.1038/s41562-024-02024-1", "publisher": "Nature Human Behaviour", "publicationDate": "2024-10-28", "design": "Preregistered systematic review/meta-analysis", "population": "Human–AI task experiments", "sampleSize": "106 experiments; 370 effect sizes", "setting": "Studies published January 2020–June 2023", "keyFinding": "Average combinations underperformed the better standalone agent; task type moderated results.", "limitations": "Historical systems; heterogeneous tasks; does not measure whole enterprises.", "transferabilityToGCC": "Local baselines essential; no assumed human–AI synergy.", "measurementImplication": "Compare feasible human-only, AI-only and combined workflows.", "evidenceType": "peer-reviewed evidence synthesis" }, { "id": "L16", "title": "Conventional Wisdom on Measurement: A Structural Equation Perspective", "url": "https://doi.org/10.1037/0033-2909.110.2.305", "publisher": "Psychological Bulletin", "publicationDate": "1991-09", "design": "Conceptual measurement-model analysis", "population": null, "sampleSize": null, "setting": "Structural equation measurement", "keyFinding": "Consequences of indicator selection and consistency depend on whether indicators cause or reflect a construct.", "limitations": "Conceptual argument, not validation of an AI index.", "transferabilityToGCC": "Applicable design distinction; requires local construct work.", "measurementImplication": "Treat the public composite as a designed disclosure index; do not use internal consistency alone to validate it.", "evidenceType": "peer-reviewed-methodological" }, { "id": "L17", "title": "Weights and importance in composite indicators: Closing the gap", "url": "https://doi.org/10.1016/j.ecolind.2017.03.056", "publisher": "Ecological Indicators / European Commission JRC", "publicationDate": "2017-09", "design": "Methodological analysis with three case studies", "population": "Resource Governance, Good Country and Water Retention indices", "sampleSize": null, "setting": "Composite indicators", "keyFinding": "Nominal weights can differ from statistical importance because indicators vary and correlate.", "limitations": "Case-study methods do not establish optimal weights for Mayden; small selected cohorts limit statistical estimation.", "transferabilityToGCC": "Relevant to any proposed regional composite.", "measurementImplication": "Publish nominal weights and deterministic sensitivity; avoid claiming equal weights mean equal effective importance.", "evidenceType": "peer-reviewed-methodological" }, { "id": "L18", "title": "COSMIN methodology for assessing the content validity of PROMs: User manual", "url": "https://www.cosmin.nl/wp-content/uploads/COSMIN-methodology-for-content-validity-user-manual-v1.pdf", "publisher": "COSMIN", "publicationDate": "2018-02", "design": "Methodological manual", "population": "Patient-reported outcome measurement instruments", "sampleSize": null, "setting": "Health measurement", "keyFinding": "Content validity concerns relevance, comprehensiveness and comprehensibility, including intended respondents' understanding.", "limitations": "PROM guidance; clinical thresholds and grading do not automatically transfer to organisations.", "transferabilityToGCC": "Adapt interview principles with Arabic/English executives and operational users.", "measurementImplication": "Test interpretation, omitted content and response processes before statistical scoring.", "evidenceType": "methodological-guideline" }, { "id": "L19", "title": "PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews", "url": "https://doi.org/10.1186/s13643-020-01542-z", "publisher": "Systematic Reviews", "publicationDate": "2021-01-26", "design": "Reporting-guideline development", "population": "Systematic-review searches", "sampleSize": null, "setting": "Research synthesis", "keyFinding": "Sixteen reporting items improve transparency of searches; the checklist is for reporting, not a complete conduct standard.", "limitations": "Adapting a search log does not make a corporate review systematic or exhaustive.", "transferabilityToGCC": "Useful audit discipline for multilingual source discovery.", "measurementImplication": "Log platform, exact query, date, limits, supplementary routes, deduplication and retrieved counts.", "evidenceType": "reporting-guideline" }, { "id": "L20", "title": "The PRISMA 2020 statement: an updated guideline for reporting systematic reviews", "url": "https://doi.org/10.1136/bmj.n71", "publisher": "The BMJ", "publicationDate": "2021-03-29", "design": "Reporting-guideline development", "population": "Systematic reviews", "sampleSize": null, "setting": "Research synthesis reporting", "keyFinding": "Provides a 27-item checklist and flow diagrams for transparent systematic-review reporting.", "limitations": "Checklist adoption alone is not proof of complete searching or low bias.", "transferabilityToGCC": "Applicable reporting distinctions across research settings.", "measurementImplication": "Call the current synthesis a scoped rapid evidence review; reserve systematic-review claims for a documented eligible design.", "evidenceType": "reporting-guideline" }, { "id": "L21", "title": "PROV-DM: The PROV Data Model", "url": "https://www.w3.org/TR/prov-dm/", "publisher": "W3C", "publicationDate": "2013-04-30", "design": "Normative data-model specification", "population": null, "sampleSize": null, "setting": "Domain-independent provenance", "keyFinding": "Represents entities, activities, agents and derivations so information can be traced across transformations.", "limitations": "A provenance graph records derivation; it does not establish factual truth.", "transferabilityToGCC": "Useful architecture independent of geography.", "measurementImplication": "Link fetched artifacts, extracted spans, claims, reviews and releases with stable IDs and hashes.", "evidenceType": "technical-standard" }, { "id": "L22", "title": "Large Firms With at Least 20 Employees Biggest AI Users", "url": "https://www.census.gov/library/stories/2026/05/ai-use-businesses.html", "publisher": "U.S. Census Bureau", "publicationDate": "2026-05-26", "design": "Original statistical-agency survey analysis", "population": "U.S. businesses covered by Business Trends and Outlook Survey", "sampleSize": null, "setting": "U.S. business AI use", "keyFinding": "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.", "limitations": "Exact analytic sample not reported in inspected article; wording change limits historical comparisons.", "transferabilityToGCC": "Question design is transferable; U.S. prevalence is not a GCC estimate.", "measurementImplication": "Keep a stable reference period and bridge instrument revisions before publishing adoption trends.", "evidenceType": "original-survey-analysis" }, { "id": "L23", "title": "How Organizations Use AI: Evidence from ChatGPT", "url": "https://arxiv.org/abs/2608.12236v1", "publisher": "arXiv author preprint", "publicationDate": "2026-08-12", "design": "Observational linked enterprise-account and usage analysis", "population": "ChatGPT Enterprise organisations and workers", "sampleSize": "More than 1,500 organisations and 17 million messages at the six-month adoption horizon, as described in the abstract", "setting": "Account records through March 2026", "keyFinding": "Adoption differs in speed, breadth and purpose; account use is linked to roles and firm characteristics.", "limitations": "Preprint; inspected abstract only. Selected customers of one product, observational data and incomplete enterprise-tool coverage.", "transferabilityToGCC": "Suggests useful behavioural dimensions, not transferable adoption norms.", "measurementImplication": "Separate available seats, active use, task purpose and role coverage; triangulate surveys with permitted telemetry.", "evidenceType": "empirical-preprint-observational" }, { "id": "L24", "title": "Generative AI and Sales Productivity: Field Experiments in Online Retail", "url": "https://arxiv.org/abs/2510.12049v6", "publisher": "arXiv author preprint", "publicationDate": "2026-06-29", "design": "Series of randomized field experiments", "population": "Users and products of one cross-border retail platform", "sampleSize": "Millions of users and products; exact pooled total not stated in inspected abstract", "setting": "Seven workflows, experiments in 2023–2024", "keyFinding": "Sales effects differ across workflow applications, with some showing no effect.", "limitations": "Preprint version 6; inspected abstract only. One platform and bounded sales outcomes do not establish enterprise profitability.", "transferabilityToGCC": "Workflow-specific evaluation principle transfers; effect sizes cannot be assigned to GCC sectors.", "measurementImplication": "Record baseline, task, outcome, comparison and period; keep sales effects separate from capability scoring.", "evidenceType": "empirical-preprint-experimental" }, { "id": "L25", "title": "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI", "url": "https://www.nber.org/system/files/working_papers/w33777/w33777.pdf", "publisher": "NBER working paper 33777", "publicationDate": "2026-03", "design": "Linked adoption surveys and administrative records; observational difference-in-differences", "population": "Workers in 11 chatbot-exposed Danish occupations", "sampleSize": "Approximately 25,000 complete responses per survey round; latest round spans about 7,000 workplaces", "setting": "November–December 2023 and 2024 surveys; labour records through December 2024", "keyFinding": "Task reorganisation coexists with estimates near zero for earnings and recorded hours.", "limitations": "Working paper; nonrandom adoption, identification assumptions, short horizon; no post-adoption firm financial analysis.", "transferabilityToGCC": "Danish labour institutions and occupational selection prevent direct GCC effect transfer.", "measurementImplication": "Separate task change, capacity, pay, employment and profit outcomes.", "evidenceType": "empirical-working-paper-observational" }, { "id": "L26", "title": "Measuring AI agent autonomy in practice", "url": "https://www.anthropic.com/research/measuring-agent-autonomy", "publisher": "Anthropic original research", "publicationDate": "2026-02-18", "design": "Provider telemetry with model-assisted classification", "population": "Public API tool calls and interactive Claude Code activity", "sampleSize": "998,481 sampled API tool calls; separate 500,000-session analysis; other denominators differ by exhibit", "setting": "Late 2025–early 2026", "keyFinding": "Autonomy and oversight vary with product context and user experience.", "limitations": "Single provider; model classifications; API requests cannot be joined into complete sessions; evaluations may resemble production.", "transferabilityToGCC": "Measurement concepts transfer; traffic shares are not regional enterprise prevalence.", "measurementImplication": "Track action authority, interventions, reversibility and session boundaries separately.", "evidenceType": "provider-original-observational" }, { "id": "L27", "title": "An Evaluation of DWP’s Microsoft 365 Copilot Trial", "url": "https://www.gov.uk/government/publications/an-evaluation-of-dwps-microsoft-copilot-365-trial/an-evaluation-of-dwps-microsoft-365-copilot-trial", "publisher": "UK Department for Work and Pensions", "publicationDate": "2026-01-29", "design": "Nonrandom post-implementation surveys, adjusted regressions and interviews", "population": "DWP employees with and without licences", "sampleSize": "1,716 licensed respondents of 3,549 invited; 2,535 comparison respondents of 9,300 invited; 19 interviews", "setting": "Survey fieldwork February–March 2025", "keyFinding": "Respondents described task benefits and verification overhead.", "limitations": "No pre-trial baseline or objective outcome metrics; self-selection and nonresponse; regression does not remove all bias.", "transferabilityToGCC": "Office workflow questions transfer; reported benefits do not estimate GCC effects.", "measurementImplication": "Collect baselines and observed task metrics before rollout.", "evidenceType": "official-original-evaluation" }, { "id": "L28", "title": "Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial", "url": "https://doi.org/10.1001/jamanetworkopen.2024.40969", "publisher": "JAMA Network Open", "publicationDate": "2024-10-28", "design": "Stratified randomised, blinded-rating trial", "population": "US-licensed generalist physicians recruited through three academic institutions", "sampleSize": "50 physicians, 25 per arm; 244 completed case responses across six vignettes", "setting": "November–December 2023; one-hour sessions", "keyFinding": "Access to a chatbot did not significantly improve the primary diagnostic-reasoning score.", "limitations": "Small vignette study, one model, limited prompting instruction; not patient outcomes.", "transferabilityToGCC": "Interaction-design lesson is relevant; clinical effects require local evaluation.", "measurementImplication": "Test the combined human-and-AI workflow against usual practice.", "evidenceType": "peer-reviewed-randomized-trial" }, { "id": "L29", "title": "The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers", "url": "https://doi.org/10.1787/ea0a0fe1-en", "publisher": "OECD Social, Employment and Migration Working Papers", "publicationDate": "2023-03-22", "design": "Separate employer telephone and worker online-panel surveys", "population": "Manufacturing and finance in seven countries; employer firms with at least 20 staff", "sampleSize": "5,334 workers; 2,053 firms", "setting": "Main fieldwork mid-January to mid-February 2022", "keyFinding": "Sector-specific surveys capture implementation and workers' experiences.", "limitations": "Perceptions and associations; online-panel coverage; predates mass generative-AI adoption; matched worker-employer pairs are not established in the inspected method.", "transferabilityToGCC": "Sector wording is useful; no GCC country was sampled.", "measurementImplication": "Ask operators as well as leadership; define the technology and reference period.", "evidenceType": "official-original-survey" }, { "id": "L30", "title": "RICS artificial intelligence in construction report 2025", "url": "https://www.rics.org/news-insights/artificial-intelligence-in-construction-report", "publisher": "Royal Institution of Chartered Surveyors", "publicationDate": "2025", "design": "Professional-body online sentiment surveys", "population": "Construction and surveying professionals globally", "sampleSize": "Over 2,200 Q1 respondents; separate skills survey over 2,000", "setting": "Six AI questions in Q1 2025 Global Construction Monitor; separate Q2 skills module", "keyFinding": "Respondents distinguish pilots, regular process use and organisation-wide adoption.", "limitations": "Sentiment sample; UK accounts for 48% of Q1 responses; MEA is not a GCC estimate; two survey bases must not be pooled.", "transferabilityToGCC": "Project and process distinctions transfer; regional estimates do not.", "measurementImplication": "Keep project, asset, organisation and respondent scope explicit.", "evidenceType": "professional-body-original-survey" }, { "id": "L31", "title": "tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains", "url": "https://arxiv.org/html/2406.12045v1", "publisher": "Author preprint / Sierra research", "publicationDate": "2024-06-17", "design": "Synthetic interactive agent benchmark with repeated trials", "population": "Simulated retail and airline service tasks", "sampleSize": "115 retail tasks and 50 airline tasks", "setting": "Version 1; synthetic users, policies and databases", "keyFinding": "Repeated success differs from success in a single run; final-state tests can miss policy violations.", "limitations": "Simplified environments, simulated users and annotation assumptions; not production business outcomes.", "transferabilityToGCC": "Local policies, language and permissions require separate tests.", "measurementImplication": "Measure consistency and policy compliance in addition to task completion.", "evidenceType": "original-benchmark-preprint" }, { "id": "L32", "title": "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks", "url": "https://arxiv.org/html/2510.04374v1", "publisher": "OpenAI author preprint", "publicationDate": "2025-10-05", "design": "Expert-authored task benchmark with pairwise deliverable evaluation", "population": "Digital knowledge work selected from nine US economic sectors", "sampleSize": "1,320 tasks across 44 occupations; public gold subset 220 tasks", "setting": "Version 1; professional deliverables", "keyFinding": "Capabilities can be assessed through work products and expert comparisons.", "limitations": "Precisely specified, mainly one-shot tasks; limited tacit context and interaction; capability and modelled savings do not establish realised enterprise ROI.", "transferabilityToGCC": "Task-design method transfers; US occupational selection is not a GCC workforce sample.", "measurementImplication": "Separate model evaluation from deployment and financial evidence.", "evidenceType": "provider-original-benchmark-preprint" }, { "id": "L33", "title": "SABER: Small Actions, Big Errors — Safeguarding Mutating Steps in LLM Agents", "url": "https://arxiv.org/html/2512.07850v1", "publisher": "Amazon AGI Foundations author preprint", "publicationDate": "2025-11-26", "design": "Agent benchmark re-audit and safeguard experiments", "population": "tau-bench airline/retail variants and selected SWE-bench evaluation", "sampleSize": "50 airline and 115 retail tasks; three runs for tau-bench experiments", "setting": "Version 1; three model families; 30-turn cap", "keyFinding": "Authors identify annotation problems and test safeguards around state-changing actions.", "limitations": "Simulated confirmation; preprint; gains are not universal in table 2 and narrative overstates consistency.", "transferabilityToGCC": "Audit technique transfers; benchmark effects are not GCC operating effects.", "measurementImplication": "Review test labels, action consequences and regressions before deployment.", "evidenceType": "original-benchmark-methods-preprint" }, { "id": "L34", "title": "Population-scale cross-sectional observational study for AI-powered TB screening on one million CXRs", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12241593/", "publisher": "npj Digital Medicine; full text in PMC", "publicationDate": "2025-07-09", "design": "Retrospective observational evaluation; authors also describe a quasi-prospective approach.", "population": "Adult TB screening at Capital Health Screening Centre, Abu Dhabi", "sampleSize": "1.04 million test chest X-ray scans", "setting": "January 2022–January 2024; low-prevalence screening", "keyFinding": "Reported AUROC 98.51%; overall false-negative rate 1.57% at the optimal threshold.", "limitations": "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.", "transferabilityToGCC": "Local screening evidence; not enterprise capability or general clinical benefit.", "measurementImplication": "Separate diagnostic evaluation, projected workload and observed operating change.", "evidenceType": "original-observational-model-evaluation" } ], "reports": [ { "slug": "ai-native-enterprise", "title": "What makes an AI-native enterprise?", "dek": "A research synthesis on work, organizational capability and the evidence needed to measure them.", "kind": "Research synthesis", "arSummary": "تركيب بحثي يدرس كيف يغيّر الذكاء الاصطناعي العمل والقدرات المؤسسية، وما يلزم لقياس ذلك دون تعميم غير مبرر.", "sector": null, "body": "# What makes an AI-native enterprise?\n\n**Define the changed work before judging the enterprise.**\n\n## Define the changed work\n\nMayden proposes an AI-native enterprise as one that repeatedly converts AI capability into useful, accountable work across a defined organisational scope. This working definition concerns recurring decisions, workflows, responsibilities and learning. It is a research proposition, not an empirically validated classification.\n\nStart with a resolved service request, reviewed clinical note, accepted production schedule or authorised maintenance decision. State what AI contributes and where the work finishes. A demonstration shows that an output can be produced under particular conditions. A changed operating model needs a repeatable relationship between output, people and result.\n\nCounterexamples sharpen the definition. Buying licences establishes availability, not active use. A specialised model can be valuable without transforming an organisation. Restricting automation can reflect sound judgement; maximum autonomy is not the objective.\n\n## Test the mechanism and work boundary\n\nAI research does not deliver one transferable enterprise effect. A customer-support study found gains differing by experience; writing experiments and a bounded consulting exercise study other work. They identify mechanisms and questions but cannot be averaged into an expected return for a Gulf enterprise. [Generative AI at Work](https://doi.org/10.1093/qje/qjae044); [Experimental evidence on the productivity effects of generative artificial intelligence](https://doi.org/10.1126/science.adh2586); [Navigating the Jagged Technological Frontier](https://www.hbs.edu/ris/download.aspx?name=24-013.pdf).\n\nEvidence complicating the positive story belongs in the same review. An early-2025 developer study found slower completion in its small selected setting despite perceived speed-up. A subsequent methods update describes selection and time-measurement problems in newer estimates. This warns about transfer and measurement; it is not a universal conclusion about developers. [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/Early_2025_AI_Experienced_OS_Devs_Study-paper.pdf); [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/).\n\nDefine eligible work, observation period, comparison and acceptance criterion before evaluating results. Include rejected outputs, checking, interruptions and the people absorbing change. A faster component can shift work elsewhere.\n\n## Keep three evidence ledgers\n\nThe practice ledger records workflow, scope, accountable role, permissions, frequency and coverage. The outcome ledger records quality, time, throughput, cost, service and adverse effects, together with the comparison supporting interpretation. The evidence ledger records respondent role, document or observation, date, locator, provenance, contradictions and review.\n\nConnect the ledgers without collapsing them into one number. A documented disappointing result is scientifically valuable. A reported saving without a baseline may merit further study. Strong source access does not establish effectiveness; limited public disclosure does not establish a weak company.\n\nNine candidate themes organise the inquiry: AI in core work; workforce; ecosystem; interfaces and automation; decision lineage; experimentation; autonomy and operating model; organisational knowledge; governance. They are hypotheses about a useful composite, not interchangeable symptoms of a proven latent trait. Internal consistency alone would not validate them. [Conventional Wisdom on Measurement: A Structural Equation Perspective](https://doi.org/10.1037/0033-2909.110.2.305); [Handbook on Constructing Composite Indicators: Methodology and User Guide](https://doi.org/10.1787/9789264043466-en).\n\n## Build the benchmark around a useful decision\n\nA private report should help executives decide what to improve, test, stop or verify. It should state scope, show disagreements and propose a bounded action with an owner and measure. It must remain useful even if research does not support a single headline score.\n\nDevelopment requires content review, bilingual cognitive work, feasibility testing, independent coding and empirical evaluation for declared purposes. Sample requirements follow the construct, comparisons and precision, not a convenient target. A 40–60-company feasibility plan is a capacity assumption, not validation power. [Best Practices for Developing and Validating Scales for Health, Social, and Behavioral Research: A Primer](https://doi.org/10.3389/fpubh.2018.00149).\n\nPublic research describes source-backed examples and limits of a collection. A future participant study would examine permissioned evidence over a defined period. Their scales cannot be converted, and the public series cannot certify enterprise capability.\n\n### Leadership agenda\n\nExecutive sponsor: identify the recurring decision the assessment should improve.\n\nMethod lead: specify the construct and evidence needed for that use.\n\nOperating team: retain outcomes, interventions and uncertainty alongside successes.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "banking-public-evidence", "title": "AI in Gulf banking: where does the decision sit?", "dek": "A purposeful review of ten banking entities across five headquarters markets.", "kind": "Industry evidence review", "arSummary": "مراجعة استكشافية لأدلة عشرة بنوك في خمس أسواق. تصف النتائج الأدلة المنشورة، ولا تثبت ترتيب النضج المؤسسي أو أداء الدول.", "sector": "Banking", "body": "# AI in Gulf banking: where does the decision sit?\n\n**Information access, operational intelligence and decision authority require different evidence.**\n\nA banking assistant can retrieve policy, explain a transaction or help a colleague interpret customer conversations. Those activities carry different permissions. The first executive task is to identify what information the system receives and which decision can follow its answer. A bank-wide label hides precisely the boundary that a reviewer needs to inspect.\n\n## Start with the information boundary\n\nA source-linked answer has a different operating role from an instruction to move money, amend a customer record or approve credit. Distinguish read access, recommendation, preparation of an action and authority to execute it. Record the accountable business owner, intended user and permitted data. A general statement about responsible AI cannot answer all of those questions for a particular system.\n\nFollow an actual completed item: recover the source version, generated response, colleague’s intervention and accepted decision. This is a proposed evidence request, not a claim that the banks in this report lack those records.\n\n## Separate an answer from an accepted decision\n\nPolicy retrieval needs tests of accuracy, currency and ambiguity. Operational insight needs a trace to the decision it supports: which customer issue was accepted for action, who owned it, and what happened afterward? Access restrictions need a test where prohibited data or an impermissible action would enter the workflow. These are different evaluation plans.\n\nThe cases below illustrate different operating questions. They do not constitute a bank comparison. Their inclusion does not imply participation in Mayden research or endorsement of Mayden.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### Customer conversation\n\nEmirates NBD says PCA turns calls into recurring themes, intent and operational issues shared across teams. Named platform; decision-use quality not demonstrated. The next research question is: Which issue was acted on, and did the service outcome change?\n\nSource: [Turning millions of customer conversations into better banking experiences with AI](https://www.emiratesnbd.com/en/innovation/insights/turning-millions-of-customer-conversations-into-better-banking-experiences-with-ai).\n\n### Policy answer\n\nKFH says ShariaBot references specific supervisory-board fatwas when answering staff questions. KFH Kuwait; source links do not reconstruct the full decision. The next research question is: Can a reviewer recover the source version and check a disputed answer?\n\nSource: [KFH Reinforces Banking Leadership with Forward-Looking AI Initiatives](https://www.kfh.com/en/home/Personal/news/2025/KFH-Reinforces-Banking-Leadership-with-Forward-Looking-AI-Initiatives.html).\n\n### Permitted data\n\nNBK says Copilot Web excludes confidential data and disables document upload and download. Specified access controls; broader framework remains on the roadmap. The next research question is: Can an attempted prohibited action demonstrate that the boundary is enforced?\n\nSource: [NBK Sustainability Report 2025](https://argaamplus.s3.amazonaws.com/4b4e6fe7-09dd-4311-be33-c97735b392e7.pdf).\n\n## Measure work completed, including the checking\n\nAvailability, active use and useful completion need separate denominators. Start with the employees or decisions eligible for a workflow; then observe actual use, accepted outputs, corrections, escalations and repeated work. A tool that saves drafting time can still increase verification effort. Include that effort and later rework.\n\nThe DWP Copilot evaluation records reported benefits and checking overhead, but had no pre-trial baseline or objective outcome metrics. Its survey helps frame questions; it cannot estimate a Gulf bank’s productivity gain. [An Evaluation of DWP’s Microsoft 365 Copilot Trial](https://www.gov.uk/government/publications/an-evaluation-of-dwps-microsoft-copilot-365-trial/an-evaluation-of-dwps-microsoft-365-copilot-trial).\n\n## Give the board a decision-specific request\n\nThe business owner should explain the objective and decision boundary. The control owner should show a disputed output and the response. Finance should establish the full cost and outcome denominator. The sponsor should decide whether to expand, constrain or stop the workflow using that record.\n\nThe next research wave should examine the same defined work over a stated period, seek an operator’s account alongside leadership’s, and record changes in data access and responsibilities. That can make a private assessment useful without publishing a league table of banks.\n\n### Leadership agenda\n\nBusiness owner: define one recurring decision and its permitted inputs and actions.\n\nControl owner: trace a disputed output through intervention and correction.\n\nFinance and operations: measure accepted work, verification, rework and cost over the same period.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "telecom-public-evidence", "title": "AI in Gulf telecoms: how far does the loop close?", "dek": "Separate internal AI practice, customer offerings and future infrastructure across ten telecom entities.", "kind": "Industry evidence review", "arSummary": "مراجعة لأدلة عشر جهات اتصالات، تفصل بين الاستخدام الداخلي ومنتجات العملاء والخطط المستقبلية، مع إظهار اختلاف نطاقات المجموعات والشركات التابعة.", "sector": "Telecommunications", "body": "# AI in Gulf telecoms: how far does the loop close?\n\n**Network assurance, customer service and cyber investigation have different action and rollback boundaries.**\n\nTelecom AI sits in different loops: detecting a network problem, completing a service request or investigating a security alert. Each loop can stop at observation, recommendation or action. An executive should know where the system stops and an accountable operator takes over.\n\n## Name the loop before discussing autonomy\n\nFor a network workflow, identify the service, assets and operating unit. For customer service, distinguish answering a question from changing an account. For cyber investigation, separate evidence gathering from containment or remediation. None establishes the authority of the others.\n\nA group disclosure may describe a policy, shared platform or one subsidiary. Record that boundary. Group ambition and a subsidiary implementation cannot be treated as interchangeable observations.\n\n## Follow observation through to action\n\nTrace the signal, recommendation, authorised action and return signal. Ask what happens when information is missing, contradictory or outside the permitted action range. Rollback and escalation are part of the work.\n\nThe selected accounts illustrate assurance, service interaction and security investigation. They are attributed source descriptions. This report does not infer an architecture, permission or intervention rate where the source does not provide one.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### Network assurance\n\nstc Bahrain reports AI assurance deployed particularly for fixed wireless access. Bahrain operation; award account is not an operating audit. The next research question is: Which faults can it correct, and which require an engineer?\n\nSource: [stc Bahrain wins awards for AI-powered CX assurance and convergent billing](https://www.stc.com.bh/media-center/template?id=4346).\n\n### Customer interaction\n\nBatelco says Basma connects to backend systems for billing, device searches and fraud-reporting support. Batelco Bahrain; completion and escalation rates undisclosed. The next research question is: Does the interaction finish the request or transfer unfinished work?\n\nSource: [Beyon Sustainability Report 2025](https://beyon.com/wp-content/uploads/2026/03/Beyon_ESG2025_English.pdf).\n\n### Security investigation\n\nBeyon Cyber describes Oryx AI investigating alerts through live queries with human oversight. Managed-security service; not proof of Batelco-wide adoption. The next research question is: What action can follow an investigation, and who authorises it?\n\nSource: [Beyon Annual Report 2025](https://beyon.com/wp-content/uploads/2026/03/Beyon-AR2025-ENGLISH.pdf).\n\n## Test routine and exceptional work separately\n\nRoutine completion can conceal a fragile exception process. Test representative requests, ambiguous cases and prohibited actions, recording both completion and policy compliance. Repeat the test where variability matters. Successful demonstrations do not establish reliable operation under changing load.\n\nThe tau-bench research distinguishes repeated success from one-run success in simulated retail and airline tasks and discusses limits of final-state checks. These concepts inform test design; synthetic tasks do not measure telecom production reliability. [tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains](https://arxiv.org/html/2406.12045v1).\n\n## Extend authority with an operating case\n\nBefore widening an action boundary, record what changed, supporting evidence, exception owner and rollback conditions. Compare interruption and escalation burden with the service result. More automated action is not automatically desirable: reversible low-impact tasks and consequential network changes require different decisions.\n\nA future assessment should ask operational roles about the same loop. Disagreement between engineering, service and governance accounts is an issue to investigate, not a number to average away.\n\n### Leadership agenda\n\nEngineering: trace a deployed loop and actual action permissions.\n\nService and security: review completed cases alongside escalations and reversals.\n\nGovernance: record the evidence required before increasing authority.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "energy-public-evidence", "title": "AI in Gulf energy: from prediction to control", "dek": "Production, maintenance and control: evidence across ten selected operating entities.", "kind": "Industry evidence review", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": "Energy & Utilities", "body": "# AI in Gulf energy: from prediction to control\n\n**Predicting maintenance, adjusting a process and exploring scenarios confer different operational authority.**\n\nA prediction, a process adjustment and an engineer’s scenario analysis are different forms of work. In energy and utilities, the point of action may affect a physical asset. The useful question is what the system can change, within which operating envelope, and who remains accountable.\n\n## Locate the asset and decision\n\nBegin at the named plant, equipment class or engineering workflow. One facility does not establish group-wide coverage. A maintenance prediction can inform planning without executing maintenance. A scenario tool can help engineers choose setpoints while leaving authority with them.\n\nRecord the source’s relationship to the operation. An issuer account and supplier account can be informative; neither is automatically independent verification of continuing use or achieved benefit.\n\n## Inspect the control boundary\n\nIdentify the observations entering the model, how the output is used, who authorises change and what constrains it. Keep source-described steps separate from proposed questions. A conceptual diagram must not imply an undisclosed connection.\n\nReview intervention and fallback evidence within the operating envelope. This report offers a research structure; it does not assess process safety, certify a control system or replace engineering assurance.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### Predict a maintenance need\n\nADNOC describes Neuron5 using sensor readings at NEB and Taweelah to predict maintenance needs. Prediction is not autonomous maintenance execution. The next research question is: Who accepts a prediction and schedules the intervention?\n\nSource: [ADNOC Deploys Pioneering AI-Enabled Process Optimization Technology](https://www.adnoc.ae/en/news-and-media/press-releases/2024/adnoc-deploys-pioneering-ai-enabled-process-optimization-technology).\n\n### Adjust a process\n\nAramco describes AI/ML automatically adjusting the oil-stabilisation process. Named site; group-wide control authority is not established. The next research question is: What operating envelope, intervention and fallback have been tested?\n\nSource: [AI and Big Data](https://www.aramco.com/en/what-we-do/energy-innovation/digitalization/ai-and-big-data).\n\n### Compare operating scenarios\n\nThe supplier account describes a model through which engineers test scenarios and choose setpoints. Supplier account of an initial field test; continuing routine use unresolved. The next research question is: Which decisions remained with engineers after the field test?\n\nSource: [Geminus AI, Bapco Energies and SLB to Deploy Physics-Native AI to Optimize Bahrain’s Energy Infrastructure](https://www.automation.com/article/geminus-ai-bapco-energies-slb-physics-native-ai-bahrain-energy-infrastructure).\n\n## Build a complete benefit account\n\nMaintenance prediction might be evaluated against intervention decisions and downtime; process adjustment against output, quality, energy and constraints; engineering support against accepted decisions and verification time. Record baseline, operating conditions and period before combining results.\n\nA technology label or projected target does not establish benefit. The productivity J-curve literature describes how complementary investment can delay measured gains; it does not prove that an AI investment will repay its costs. [The Productivity J-Curve: How Intangibles Complement General Purpose Technologies](https://www.nber.org/papers/w25148).\n\n## Decide what justifies the next boundary\n\nThe asset owner should state the decision and constraints. Engineering should show intervention history and responses to unusual conditions. Operations and finance should reconcile benefit claims with checking work, maintenance, costs and observed output. Expansion should follow that record.\n\nFor the eventual private study, source publication date and confirmation of continuing use must be separate fields. A historic implementation remains useful, while a current-operation claim needs dated evidence of its own.\n\n### Leadership agenda\n\nAsset owner: specify the site, operating envelope and authority.\n\nEngineering: inspect an actual intervention and tested fallback.\n\nOperations and finance: reconcile output, quality, energy, downtime and full delivery cost.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "logistics-public-evidence", "title": "AI in Gulf logistics: the allocation decision", "dek": "A review of selected airlines, logistics groups and airport operators.", "kind": "Industry evidence review", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": "Aviation & Logistics", "body": "# AI in Gulf logistics: the allocation decision\n\n**A better recommendation matters only when it changes allocation within real operating constraints.**\n\nAI creates a practical question in aviation and logistics: which allocation changes because of the recommendation? Cargo space, instructor time and operational information do not share a single performance denominator. Follow the handoff from an output to an accepted operating decision.\n\n## Identify the scarce resource\n\nA cargo recommendation may inform demand, inventory or overbooking. Training scheduling allocates instructors and facilities. Executive information can assemble a view without commanding operations. Define the resource, decision maker and constraints that make the decision consequential.\n\nKeep airline, cargo, academy, airport and group IT boundaries visible. An implementation within one setting does not establish end-to-end autonomy across a transport group.\n\n## Trace the handoff, including rejection\n\nIdentify who accepts the recommendation, what is checked, how an override is recorded and what follows rejection. Coverage should describe relevant routes, modules or decisions, rather than parent-company size.\n\nThe cases below illustrate separate allocation and information workflows. They are not a common performance test. Vendor descriptions can identify a deployment and intended workflow; achieved network-wide value requires a different observation.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### Cargo allocation\n\nWiremind describes progressive rollout of optimisation tools within Qatar Airways Cargo. Cargo scope; complete network and module coverage unestablished. The next research question is: Which decisions use the recommendation, and which exceptions revert to people?\n\nSource: [Qatar Airways Cargo goes live with CARGOSTACK Optimiser](https://www.wiremind.io/blog/qatar-airways-cargo-cargostack-optimiser).\n\n### Training allocation\n\nMicrosoft describes Saudia Academy implementing AI-enabled scheduling in training management. Academy workflow; not the airline’s wider operating system. The next research question is: How many eligible scheduling decisions use it, and what does a planner override?\n\nSource: [Saudia Academy transforms aviation training with Dynamics 365](https://www.microsoft.com/en/customers/story/1782079017420911107-saudia-academy-dynamics-365-customer-service-travel-and-transportation-en-saudi-arabia).\n\n### Executive information\n\nGulf Air’s FalconEye account combines flight, passenger and crew information for executive questions. Airline information interface; not autonomous operational command. The next research question is: Can the answer be traced to current data before an executive acts?\n\nSource: [Public preview: Voice-native agents in Microsoft Foundry](https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/public-preview-voice-native-agents-in-microsoft-foundry/4502756).\n\n## Evaluate service, constraints and rework together\n\nA faster schedule is useful only if it remains workable. Define the comparison period, workload, constraints, service outcomes and downstream effort. Count manual repair, repeated planning and exceptions. Preserve model and policy changes so apparent improvement does not conceal a different task.\n\nAgent benchmark research warns that an expected final state need not demonstrate that intermediate actions followed policy. Synthetic airline tasks are not carrier-performance evidence, but motivate inspecting the route as well as the outcome. [tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains](https://arxiv.org/html/2406.12045v1).\n\n## Create a repeatable operating record\n\nThe planning owner should select one allocation decision with a clear comparison boundary. Operations should identify where a plan becomes an accepted commitment. Evaluation should retain rejected recommendations and disruption cases alongside normal work.\n\nA future diagnostic can return a bounded priority: improve evidence around a handoff, clarify responsibility or test an expansion. It should not extrapolate one scheduling application into a company-wide label or promise a financial result without compatible data.\n\n### Leadership agenda\n\nPlanning: identify one allocation decision and its binding constraints.\n\nOperations: trace recommendation, acceptance, override and completed service.\n\nEvaluation: retain disruption cases, rework and the full planning cost.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "manufacturing-public-evidence", "title": "AI in Gulf manufacturing: close the operating loop", "dek": "Production, maintenance and quality across ten selected industrial entities.", "kind": "Industry evidence review", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": "Manufacturing & Materials", "body": "# AI in Gulf manufacturing: close the operating loop\n\n**Inspection, safety attention and order planning demand different proof at the point of action.**\n\nFactory AI becomes an organisational question when somebody acts on a signal. Inspecting an anode, prioritising safety attention and planning an order are different decisions. A model result alone does not show how the plant changed its work.\n\n## Start at the production event\n\nDefine the event, product or schedule, the site and the employee receiving the result. The denominator might be inspected units, eligible operating areas or manufacturing orders. Group revenue, model count and announced investment cannot substitute for those operating denominators.\n\nThe selected accounts cover different tasks. Read them as examples of what to investigate, not evidence that one manufacturer has a better overall operating model.\n\n## Join detection quality to intervention\n\nTrace a quality signal to review, rejection, rework or acceptance; safety prioritisation to an action and the risks it may miss; an order recommendation to the schedule a planner accepts. Each pathway needs its own evidence.\n\nA conceptual evaluation can examine missed events and false alerts alongside delay, workload and adverse consequences. These are proposed inquiry dimensions, not a populated chart of company performance. This review establishes no common numerical threshold.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### Inspect quality\n\nEGA reports real-time AI vision checking carbon-anode quality at Al Taweelah. Site-specific inspection; group-wide autonomy not established. The next research question is: What is rejected, reviewed or reworked after the signal?\n\nSource: [EGA designated Industry 4.0 Global Lighthouse](https://media.ega.ae/ega-designated-industry-40-global-lighthouse-by-world-economic-forum-in-uae-and-world-aluminium-industry-first/).\n\n### Prioritise safety attention\n\nAlba describes AI using incident patterns and risk drivers to identify higher-risk areas. Identification is reported; causal safety improvement is not verified. The next research question is: What action changes, and how are false alerts and missed risks reviewed?\n\nSource: [Al Bayan Q1 2026](https://www.albasmelter.com/uploads/albayan2026_q1.pdf).\n\n### Plan production\n\nDucab states that AI software is used in order planning within Blade. Facility coverage undisclosed; targets are not achieved outcomes. The next research question is: Which schedule changes are accepted, overridden and measured?\n\nSource: [Ducab launches Blade smart factory project](https://ducab.com/media-center/press-releases/ducab-group-expedites-operations-with-the-launch-of-blade-the-smart-factory-project-of-the-future-based-on-industry-4-0-technologies/).\n\n## Include the people absorbing the change\n\nOperators can reveal checking, workarounds and new tasks that leadership does not see. OECD employer and worker surveys illustrate sector-specific questions and multiple perspectives. Their manufacturing and finance samples came from seven non-GCC countries; the reported perceptions are not a local causal effect estimate. [The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers](https://doi.org/10.1787/ea0a0fe1-en).\n\nRecord who is trained, who handles uncertainty and what happens during unavailability. An intervention needs an owner even when the model performs well. Keep planned pilots, completed models and sustained production use distinct.\n\n## Make the intervention record the management unit\n\nConnect event, data/model version, signal, reviewer, decision and subsequent observation. Preserve accepted and rejected recommendations. The production owner can then decide whether the tool improved the process and whether its boundary should change.\n\nA future benchmark must retain site and product scope, workload shifts and process changes. Otherwise a repeated company questionnaire may appear to show progress while describing another factory or another task.\n\n### Leadership agenda\n\nPlant owner: define the process, event and output denominator.\n\nOperators: identify checking, false alerts, missed events and workarounds.\n\nEvaluation: connect outputs to interventions and observed consequences.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "retail-public-evidence", "title": "AI in Gulf retail: who handles the exception?", "dek": "Demand, inventory and service across ten selected retail and consumer entities.", "kind": "Industry evidence review", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": "Retail & Consumer", "body": "# AI in Gulf retail: who handles the exception?\n\n**Forecasting, service and supplier workflows reveal how AI is connected to the organization.**\n\nRetail AI is visible in an interaction and often consequential in the work behind it. Replenishment, service and supplier questionnaires reach different decision boundaries. Ask who handles the exception and whether the workflow actually finishes.\n\n## Name the system and decision\n\nA forecast can inform an order without placing it. An assistant can connect to customer systems without resolving every request. A drafted questionnaire can begin a supplier-evidence process without demonstrating responses or compliance. Make these distinctions before discussing benefit.\n\nA rollout described for the UAE should retain that scope. A historical store example should retain its date. Neither establishes current group-wide practice across markets.\n\n## Follow the exception\n\nIdentify the rules bounding an action, who reviews an exception and the evidence that survives review. For replenishment this may be an accepted order; for service, a completed or escalated transaction; for suppliers, a reviewed response informing a decision.\n\nThe examples below expose different completion tests. They do not establish retail prevalence, and their public description does not imply Mayden access to internal systems.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### Forecast to replenishment\n\nLuLu says UAE planning teams review AI outputs and exceptions under defined rules and thresholds. Reported approval workflow; no inference to every market. The next research question is: Does the accepted order improve availability without shifting waste elsewhere?\n\nSource: [LuLu Retail Integrated Annual Report 2025](https://www.luluretail.com/media/pqobufrm/lulu-retail_iar2025_eng.pdf).\n\n### Conversation to service\n\nLuLu describes Salem on WhatsApp integrated with customer and order systems. Integration alone does not establish successful resolution. The next research question is: Is the transaction completed, escalated or repeated by the customer?\n\nSource: [AI-powered customer experiences on WhatsApp with Yellow.ai](https://www.luluretail.com/media/news/lulu-hypermarket-launches-new-ai-powered-customer-experiences-on-whatsapp-in-collaboration-with-yellowai).\n\n### Draft to supplier evidence\n\nMezzan reports using GenAI to develop an ESG questionnaire for targeted high-spend suppliers. Target population is not confirmed responses or compliance. The next research question is: Which responses change a supplier decision after human review?\n\nSource: [Mezzan Sustainability Report 2025](https://www.mezzan.com/wp-content/uploads/2026/08/Mezzan-Sustainability-Report-2025-English-.pdf).\n\n## Choose a denominator at the work boundary\n\nReplenishment evaluation should consider availability, waste and planner effort over comparable products and periods. Service should include repeat contacts and escalation. Supplier analysis should distinguish intended recipients, responses, usable evidence and decisions informed. Do not pool these ledgers into an AI return percentage.\n\nA retail field-experiment preprint reports effects differing across seven applications, including some with no effect. The retained review inspected its abstract only. It supports a workflow-specific question, not a transferable outcome magnitude or proof of enterprise profit. [Generative AI and Sales Productivity: Field Experiments in Online Retail](https://arxiv.org/abs/2510.12049v6).\n\n## Design the next operating test\n\nChoose a workflow and the decision it can change. Review normal and exception cases. Include checking, rework, compute and implementation effort in the cost account. Interpret customer or supplier feedback with its actual response denominator.\n\nA future participant report can identify a bounded experiment and owner. It should not turn one public account into a company-wide maturity judgement or treat an undisclosed result as failure.\n\n### Leadership agenda\n\nBusiness owner: define the accepted order, completed service or reviewed response.\n\nControl owner: inspect thresholds, exceptions and escalation.\n\nFinance: include checking, repeated work and implementation costs.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "healthcare-public-evidence", "title": "AI in Gulf healthcare: evidence at the point of care", "dek": "Clinical and administrative workflows, responsibility and the boundaries of deployment.", "kind": "Industry evidence review", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": "Healthcare", "body": "# AI in Gulf healthcare: evidence at the point of care\n\n**Operating throughput, retrospective evaluation and clinical authority are separate lines of evidence.**\n\nHealthcare AI needs distinct evidence layers: a system may process examinations, perform well retrospectively and still need separate evidence of benefit in a live pathway. Mayden follows the workflow and accountable clinical decision; this is not a clinical-quality rating.\n\n## Identify the clinical and organisational question\n\nDocumentation, image interpretation and screening coordination change different work. Specify provider, service, patient population, professional role and period. A company-wide statement cannot establish suitability for every clinical use.\n\nThroughput describes activity. Retrospective performance describes evaluation under a study design. Prospective effects on care, workload and safety need further observations. A technical result should not carry a broader claim than its evidence supports.\n\n## Read the study in its actual setting\n\nThe AIRIS-TB paper evaluates 1.04 million chest X-rays from January 2022 to January 2024 in an observational retrospective analysis in Abu Dhabi. Projected reporting reduction is calculated from classifications, not demonstrated prospective staffing savings. External-dataset performance and restricted sputum-confirmation pathways limit generalisation. M42/CHSC affiliations and model-developer involvement remain relevant. [Population-scale cross-sectional observational study for AI-powered TB screening on one million CXRs](https://pmc.ncbi.nlm.nih.gov/articles/PMC12241593/).\n\nThis is locally relevant research, not an enterprise-capability measure or proof of a new deployment’s effect in another population. The accompanying cases concern distinct evidence layers and must not be pooled into a provider comparison.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### Operating activity\n\nM42 reports AIRIS-TB conducting 2,000 AI chest-X-ray examinations per day. Issuer-reported throughput; staffing/comparator details incomplete. The next research question is: How many eligible examinations reach a reviewed clinical decision?\n\nSource: [M42 and GE HealthCare join forces to advance AI-Enabled Patient Care](https://m42.ae/media-resources/news/m42-and-ge-healthcare-join-forces-to-advance-ai-enabled-patient-care/).\n\n### Retrospective evaluation\n\nThe original AIRIS-TB study evaluates 1.04 million chest X-rays from January 2022 to January 2024 in an observational retrospective analysis. M42/CHSC-affiliated research; model evaluation is not prospective patient benefit. The next research question is: Does performance hold in the intended workflow, population and prevalence?\n\nSource: [M42 announces breakthrough results for its AI-powered tuberculosis screening](https://m42.ae/media-resources/news/m42-announces-breakthrough-results-for-its-ai-powered-tuberculosis-screening/).\n\n### Clinical authority\n\nHMG says its Responsible AI Framework retains clinician final authority and includes fairness monitoring. A different provider and evidence layer; no assurance result supplied. The next research question is: Can an actual disagreement show how review, intervention and follow-up operate?\n\nSource: [Annual Report 2025](https://hmg.com/ir/en/Lists/Reports_Publications/Attachments/111/HMG_Annual_Report_2025__English.pdf).\n\n## Evaluate the combined workflow\n\nFinal authority should be observable in an accepted or rejected output, correction and follow-up. For documentation, ask who verifies the note and how errors are corrected. For screening, trace the output to a reviewed decision and identify eligible examinations outside the workflow. These are research questions, not treatment recommendations.\n\nA randomised study of 50 physicians found no significant improvement in its primary diagnostic-reasoning score from chatbot access. It used vignettes, not patient outcomes. This illustrates why the combined workflow needs evaluation; it does not settle the value of other systems. [Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial](https://doi.org/10.1001/jamanetworkopen.2024.40969).\n\n## Keep the next study accountable\n\nClinical, operational and technical reviewers should agree intended use, comparator, period and outcome definitions before collection. Include verification burden and subgroup performance where appropriate. Preserve disputed outputs and corrections.\n\nThe future company diagnostic requires qualified human review and permissioned evidence. It must not award clinical assurance, infer patient benefit from disclosure volume or interpret limited public detail as unsafe care.\n\n### Leadership agenda\n\nClinical owner: define intended use, population and final authority.\n\nOperations: measure eligible work, reviewed outputs and correction effort.\n\nResearch reviewer: separate retrospective performance, projected effects and observed workflow outcomes.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "realestate-public-evidence", "title": "AI in Gulf real estate: name the asset and the action", "dek": "Property, construction and asset operations across ten selected entities.", "kind": "Industry evidence review", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": "Real Estate & Construction", "body": "# AI in Gulf real estate: name the asset and the action\n\n**A district, an environmental survey and a hotel kitchen cannot stand in for an entire property enterprise.**\n\nA developer’s name can span design, construction, asset operation, hospitality and environmental management. Useful AI evidence identifies which unit changed a recurring decision. A smart-city description cannot answer that question on its own.\n\n## Locate the operating boundary\n\nBegin with asset, project, operating entity and responsibility. Design capacity differs from a live installation; a subsidiary differs from the full portfolio. A kitchen and a district control system can sit within related corporate narratives while needing different evaluation.\n\nThe cases below make those distinctions concrete. They do not rank developers, measure investment quality or assess building safety.\n\n## Trace observation into management action\n\nImage analysis can support environmental monitoring without establishing ecological improvement. District or parking applications can inform decisions without showing which assets are live or how authority is allocated. Ask which result was acted on, who accepted it and what followed.\n\nPreserve the actual source account. An unreported action link is an open question. Do not draw a complete architecture from a statement naming only a partnership or capability.\n\n### Attributed operating examples\n\nThese separate cases were selected by the operating question, not company standing. The diagram does not imply one connected system.\n\n### District and parking\n\nAldar describes district AI with Siemens and parking applications through Aldar Estates/Parkonic. Specific operating entities; design capacity is not active deployment. The next research question is: Which assets are live, and who owns the consequential decision?\n\nSource: [Aldar Innovation Report 2025](https://app-aprp01saasa21mtp001.cms.optimizely.com/globalassets/aldar-innovation-report-2025---eng---final-09-april.pdf).\n\n### Environmental monitoring\n\nRSG describes CoralNet processing survey imagery after training on human-labelled images. Monitoring capability is not evidence of ecological improvement. The next research question is: Which management action follows a survey result, and what is observed afterwards?\n\nSource: [Innovation meets preservation: technology’s impact on Red Sea biodiversity conservation](https://www.redseaglobal.com/en/media-center/blog/innovation-meets-preservation-technology-s-impact-on-red-sea-biodiversity-conservation/).\n\n### Hospitality kitchen\n\nWinnow’s 2019 account describes AI-enabled food-waste recognition after a proof of concept. Historical supplier account; current use and coverage not confirmed. The next research question is: Which kitchens still use it, and how does the signal change purchasing or production?\n\nSource: [Emaar Hospitality Group first to bring AI technology into UAE kitchens](https://www.winnowsolutions.com/resources/news/emaar-hospitality-group-first-to-bring-cutting-edge-technology-into-uae-kitchen).\n\n## Keep time visible\n\nA 2019 supplier account illustrates a workflow described then; it cannot establish current use, portfolio or result. Publication date, operating period, retrieval date and confirmation of continuing use need separate fields.\n\nRICS research distinguishes pilots, regular process use and organisational adoption within a geographically uneven sentiment sample. Separate survey bases must not be pooled. It supports careful units and periods, not a prevalence estimate for Gulf developers. [RICS artificial intelligence in construction report 2025](https://www.rics.org/news-insights/artificial-intelligence-in-construction-report).\n\n## Design an asset-level evidence request\n\nFor district operations, identify live assets, responsible operator and decision boundary. For monitoring, connect observation, management action and later result. For hospitality, trace a signal into purchasing or production while recording waste and workload comparably.\n\nA future participant study should track asset transfers, project stages, contracts and portfolio changes. Apparent company change between editions may otherwise reflect a different portfolio rather than a different way of working.\n\n### Leadership agenda\n\nAsset owner: identify the live asset, operator and responsibilities.\n\nOperations: show a completed decision following the AI output.\n\nResearch: separate publication, observation and continuing-use dates.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship." }, { "slug": "state-of-ai-native-enterprise", "title": "The AI-native enterprise: an operating evidence atlas", "dek": "Eight sectors. Five markets. A closer examination of how AI enters recurring work.", "kind": "Measurement research", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": null, "body": "# The AI-native enterprise: an operating evidence atlas\n\n**Follow the decision, then test the operating change.**\n\n## Read the atlas as a research collection\n\nThe atlas connects eight inquiries: banking; telecoms; energy and utilities; aviation and logistics; manufacturing and materials; retail and consumer; healthcare; real estate and construction. Each contains ten selected entity records. Relevance and identifiable public disclosures influenced purposeful selection. This is not a census or probability sample.\n\nEighty records create 720 possible entity-theme cells across nine themes. Retained coding contains qualifying evidence in 201 cells; 519 remain unknown in this review. These are documentation counts, not 519 absent capabilities. A source can support several claims or themes; related corporate records can overlap.\n\nCoverage charts count records with qualifying disclosures out of ten selected records per sector. They do not estimate adoption, national performance or which sector is most AI-native. Scope, publication practice and retrieval success can change what is visible.\n\n## Separate deployment, use and continuing operation\n\nOf 224 eligible claim records, 79 have no recorded source-publication date, 28 have a date before 2025 and 117 have a date in 2025 or later. These sum to the claim denominator. They are not company, project or independent-confirmation counts. An undated source can discuss a dated reporting period; a recent publication can describe an older implementation.\n\nDeployment does not establish how much eligible work uses a system today. Continuing-use evidence needs a dated observation of the operation. A source-date filter can expose dependence on older documentation, but cannot show that an earlier practice stopped.\n\nLater editions must distinguish discovery from operating change. Access recovery, coding correction, changed corporate scope and revised questions need separate change records.\n\n## Ask a different first question in each sector\n\nIn banking, trace information and authority. In telecoms, locate where the observation-to-action loop closes. In energy, distinguish prediction from physical-process control. In logistics, identify allocation and constraints.\n\nIn manufacturing, connect detection to intervention. In retail, follow exceptions until work finishes. In healthcare, separate activity, model evaluation and clinical consequences. In real estate, locate asset, operating entity and period. These distinctions draw on attributed cases and Mayden’s synthesis; they are not estimated maturity stages.\n\nA common research record can still capture work unit, accountable role, actual use, authority, result and period. Sector modules ask what those fields mean in context. Comparability must be tested, not assumed because headings match.\n\n## Make the next edition answer a stronger question\n\nA broader public crawl could improve discovery and repeatability. It cannot reveal confidential operations or remove disclosure bias. The existing collector is a bounded public-source checker; retrieval and text matches create leads, not accepted operating claims.\n\nNext work includes reconciling source families and cases, resolving review gaps, testing the candidate construct with knowledgeable humans and designing permissioned fieldwork. Multi-informant records should preserve disagreement. Outcomes should include checking and rework; later Mayden advisory work must be recorded as a possible influence on change.\n\nPublish bounded findings, including nulls and unresolved questions. Return specific priorities privately to participants. Benchmark status must be earned through usefulness, repeat participation, stable methods and validation; collection size and visual finish do not establish it.\n\n### Leadership agenda\n\nBoard: ask for the operating boundary behind each headline claim.\n\nResearch: improve source dependencies and continuity evidence before expanding numerical claims.\n\nProgramme owner: fund the next wave and human review before promising an annual benchmark.\n\n### Scope and reading boundaries\n\nThe source corpus contains 80 purposefully selected entity records across eight sectors, ten per sector. Headquarters markets are Saudi Arabia, the UAE, Qatar, Kuwait and Bahrain; Oman is outside this edition. Groups, subsidiaries and specific operations have different scopes. This is not a representative sample. Source cutoff: 10 October 2026; publication edition: 11 October 2026.\n\nSource extraction and challenge were AI-assisted, not independent human assurance or verification of current operations. Source visibility is not adoption prevalence. Unreported details are unresolved questions, not absent capabilities. Named examples imply neither participation, endorsement, confidential access nor a Mayden client relationship.\n\nCorpus provenance: [Mayden aggregate research inputs](https://www.mayden.ai/research/mayden-index-2026-10.json) and [public methodology](https://www.mayden.ai/ai-native-index/methodology). Operating synthesis is developed in the [banking report](https://www.mayden.ai/ai-native-index/reports/banking-public-evidence), [healthcare report](https://www.mayden.ai/ai-native-index/reports/healthcare-public-evidence) and [full research collection](https://www.mayden.ai/ai-native-index). These are Mayden analytical outputs from the same selected corpus, not independent corroboration." }, { "slug": "research-depth", "title": "How much should you trust an AI index?", "dek": "Evidence quality, disclosure bias and what changes when scoring assumptions change.", "kind": "Measurement research", "arSummary": "دراسة للأدلة المنشورة وحدود التقييم. التقرير التفصيلي متاح بالإنجليزية.", "sector": null, "body": "# How much should you trust an AI index?\n\n**An index earns trust through its construct, evidence and reproducibility, not its visual precision.**\n\n## Begin with the claim the number would make\n\nNormalisation makes observations calculable on a chosen scale. It does not establish that components represent the intended construct or deserve their weights. The OECD/JRC handbook exposes such modelling choices; it does not validate Mayden’s dimensions. [Handbook on Constructing Composite Indicators: Methodology and User Guide](https://doi.org/10.1787/9789264043466-en).\n\nHistorical public-evidence editions used the strongest qualifying disclosure in each of nine themes, a fixed denominator and equal nominal weights. Arithmetic could be replayed, but one narrow implementation could fill a theme, unknown public evidence contributed no documented points, and access varied. These properties limit enterprise-capability interpretation. The current edition removes individual company scores from active public research.\n\nThe future instrument should distinguish a formative composite, reflective subscales and direct descriptive measures. A formative design defines a concept through selected components; high correlation is not automatically desirable. Reflective components need their own model and tests. Neither is validated by a convenient distribution. [Conventional Wisdom on Measurement: A Structural Equation Perspective](https://doi.org/10.1037/0033-2909.110.2.305); [Best Practices for Developing and Validating Scales for Health, Social, and Behavioral Research: A Primer](https://doi.org/10.3389/fpubh.2018.00149).\n\n## Keep source, claim and operating case separate\n\nA document can support several claims; URLs can repeat one assertion; claims can describe one case; a group can appear through subsidiaries. URL volume is not independent confirmation. Record original assertion, document version, extracted passage, claim, review and release. A provenance graph exposes derivation, not truth. [PROV-DM: The PROV Data Model](https://www.w3.org/TR/prov-dm/).\n\nLegacy review records are not all bound to exact claim hashes, and some sources lack recorded dates. New controls should surface those gaps without rewriting history as if reviews had occurred. A queue cannot confer eligibility, confirm current operation or manufacture a second reviewer.\n\nMissingness needs a reason: not searched, no qualifying evidence within a declared search, inaccessible source, outside scope or an observed negative finding. A participant instrument also needs unknown and not-applicable responses. Treating every state as zero creates precision about a different question.\n\n## Challenge modelling choices without inventing probabilities\n\nThe 11 October challenge tested 512 joint weight-corner specifications and 36 ordinal-spacing specifications against frozen v4, alongside a separately labelled baseline reference. Independent replay checked 43,840 specification-record results across 548 nonbaseline choices and 80 records. It verifies tested arithmetic, not source truth or construct validity.\n\nWithin the 512 weight choices, 53 of 59 historically ranked records changed position in at least one tested specification. Analysts selected the set; it is not a probability distribution, an exhaustive plausible model space or an error rate. This cannot become a claim that a corresponding percentage of companies was misclassified. Scope, disclosure inequality and intended use remain primary reasons to reconsider public comparison.\n\nSensitivity also needs source families, case dependence, coding disagreement, missing evidence and operating breadth. Nominal weight can differ from statistical influence when components vary and correlate. Estimating influence in a small selected corpus would not supply correct future weights. [Weights and importance in composite indicators: Closing the gap](https://doi.org/10.1016/j.ecolind.2017.03.056).\n\n## Make institutional trust part of the method\n\nThe publication model is deliberate. Inspected BCG Applied AI Index and McKinsey State of AI articles use aggregate respondent comparisons. BCG’s Most Innovative Companies 2023 publicly names and ranks firms using a different reputation/financial-performance composite. These observations establish neither client relationships nor universal consultancy policy. [BCG AI research](https://www.bcg.com/publications/2026/the-formula-for-agentic-ai-value); [McKinsey AI research](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2025); [BCG innovation methodology](https://www.bcg.com/assets/2023/most-innovative-companies-reaching-new-heights-in-uncertain-times.pdf).\n\nMayden’s accepted direction is public sector research with attributed facts and a future confidential opt-in diagnostic. Previously published data cannot become anonymous. Preserve history and explain the change. Public facts do not imply participation or endorsement.\n\nAppoint a research lead, a consequential-decision reviewer, a correction route and authority to halt unsupported release. The World Bank discontinued Doing Business after data irregularities: institutional standing does not substitute for controls. That case is not causal evidence about Mayden. [World Bank statement](https://www.worldbank.org/en/news/statement/2021/09/16/world-bank-group-to-discontinue-doing-business-report).\n\nPredeclare intended decisions, units, sampling frame, validation criteria and analysis. Test content, comprehension, coding and role agreement, sector/language comparability and outcomes against simple alternatives. Report attrition and distinguish within-company change from composition or method change. Human appointments, fieldwork and empirical validation remain future gates.\n\n### Leadership agenda\n\nMethod lead: specify what each item measures and its required evidence.\n\nRelease editor: place denominator, period, access limits and assumptions beside each finding.\n\nProgramme owner: appoint accountable humans and protect their right to stop publication." } ], "sourceCount": 206, "evidenceCount": 272, "eligibleEvidenceCount": 224, "methodology": "# Public Evidence Edition: method and publication policy\n\nVersion PE-2026.10-v5 · Source-corpus cutoff: 10 October 2026 · Publication and methodological challenge: 11 October 2026\n\n## What the current public edition does\n\nMayden studies how AI changes recurring work, decisions, responsibilities and organisational learning. Its working definition of an AI-native enterprise is one that repeatedly converts AI capability into useful, accountable work across a defined organisational scope. That definition and the nine-theme framework are research hypotheses, not a validated enterprise classification.\n\nThe current edition publishes **sector research, corpus-level documentation statistics and attributed operating examples, without individual company scores, ranks or coded profiles**. Its purpose is to improve executive questions and make research boundaries inspectable. It is not an enterprise audit, certification, clinical or investment rating, representative adoption survey, or a confidential company assessment. The confidential opt-in diagnostic is in development; it is not yet a validated or operating service.\n\nNamed facts can make an operating example useful. Each must retain its publisher, period or unknown date, locator, operating scope and relevant boundary. Selection is by an explicit research question, not score or standing. Public mention implies neither participation, endorsement, client relationship nor permission for confidential access. No company is portrayed as weak because qualifying public evidence was not located. These publication rules apply to the website, PDF, chart, downloadable data and client payload, not just the visible interface.\n\n## The selected research corpus\n\nThe retained corpus contains 80 entity records: ten in each of Banking; Telecommunications; Energy & Utilities; Aviation & Logistics; Manufacturing & Materials; Retail & Consumer; Healthcare; Real Estate & Construction. It contains 272 recorded claims, 224 retained as eligible under the historical evidence rules. These are collection quantities, not independent experimental observations.\n\nEach sector has two headquarters records in Saudi Arabia, UAE, Qatar, Kuwait and Bahrain. Oman is outside this edition; coverage is not the full GCC. Selection was purposeful, influenced by sector relevance and identifiable corporate disclosures. There was no probability sample, market-share threshold or complete industry census. No country league table, regional prevalence estimate or population percentile follows from this design.\n\nUnits mix international groups, local operators and subsidiary practices. A headquarters country does not identify where all reported work takes place. Corporate scopes can overlap. Sector cohorts also contain different business models. Public aggregates describe the selected records; they do not remove those limitations. Country-by-sector cells contain only two records and are not published as a new comparative product. Suppressing names alone would not ensure anonymity, especially because earlier named releases and source links remain public.\n\n## Search, access and source truth\n\nThe desktop review used bounded corporate-domain searches and annual, sustainability, governance and implementation reports, mainly in English with limited Arabic discovery. Search was not exhaustive or a preregistered systematic review. Logs describe performed work; they do not establish equal search intensity or recall across firms, themes or languages.\n\nThe retained archive includes claims, URLs, source dates where known, access dates, locators, implementation descriptions, scope limits, search notes and review decisions. A readable corporate report establishes what its author reports. It is not independent assurance of the operation. Vendor descriptions and mirrored reports retain origin and access limits. Republishing an assertion does not create independent confirmation. A search snippet is a discovery lead, not proof that a full source was read.\n\nThe edition does not announce a new 80-company crawl or source re-verification. The source corpus remains frozen at 10 October. The 11 October work adds a methodological challenge, institutional comparison and publication redesign. Changed presentation is not changed company performance.\n\n## How aggregate figures are constructed\n\nThe historical research codes evidence across nine candidate themes: AI Core; Augmented Workforce; Ecosystem; Interfaces & Automation; Decision Lineage; Experimentation; Autonomy & Operating Model; Knowledge; Governance. For public coverage, a record contributes one to a theme only when at least one retained eligible claim supports that theme. Each sector-theme denominator is its ten selected records. A record can contribute to several themes, so theme counts must not be added as unique companies.\n\nThere are 720 possible entity-theme cells (80 × 9); 201 contain qualifying evidence and 519 remain unknown in this review. Unknown means the review has no qualifying evidence for that cell, not that a capability is absent. Chart ordering is editorial/alphabetical, not a sector performance rank. Coverage is documentation coverage, distinct from search completeness and operational coverage of employees, workflows or decisions.\n\nSource-date figures use the 224 eligible claim records as denominator: 79 have no recorded publication date; 28 are dated before 2025; 117 are dated in 2025 or later. A claim is not a company or independent project. Publication date, reporting period, retrieval date and confirmation of continuing use answer different questions. An undated source may describe a dated period; a recent publication may describe an older deployment. No source-date category proves continuing or discontinued use.\n\nEvery exhibit specifies basis, unit, denominator where applicable, cutoff and interpretation boundary. Conceptual diagrams are proposed analytical tools, not measured distributions or maturity stages. Attributed accounts remain separate cases; connecting them visually must not imply an undisclosed system architecture. Reported benefits stay outside practice indicators and require their own baseline, period, comparison and cost boundary.\n\n## Historical arithmetic and why publication changed\n\nFrozen v1–v4 editions retain the original public-disclosure method and artifacts. Each theme used the maximum qualifying anchor: a specific commitment/pilot, an implemented practice, or an implemented practice with operating detail. Unknown contributed no documented points while remaining unknown in the profile. Historical normalisation was 100 × sum of nine theme anchors / 27, with equal nominal weights and within-sector tied positions. It did not estimate percentage AI nativeness.\n\nThat arithmetic can be replayed, but maximum anchors can reward one narrow example, ordinal distances were uncalibrated, source availability differs and organisational scopes are not homogeneous. A fixed denominator does not solve these problems. The public-policy change reflects construct, intended-use and engagement considerations; it does not claim that historical companies changed or that every earlier calculation was arithmetically wrong.\n\nThe 11 October challenge tested 512 joint weight-corner choices and 36 ordinal-spacing choices on frozen v4, plus a separately labelled baseline reference. Independent replay checked 43,840 specification-record results (548 × 80). In the 512 weight choices, 53 of 59 historically ranked records changed position in at least one tested specification. The specifications are analyst-selected deterministic tests. They are not a probability distribution, error rate, confidence interval, exhaustive plausible model space or validation study. Full individual diagnostics remain internal; this aggregate account retains its historical basis and limits.\n\nEarlier files must remain explicitly historical, with their exact data, method/code provenance and correction context. They are not current enterprise ratings. Previously public information cannot become confidential or anonymous retroactively. A material correction or lawful removal, if ever required, needs its own recorded decision rather than an unannounced overwrite.\n\n## Review, reproducibility and its limits\n\nSource extraction, coding challenge and publication review were AI-assisted. Separate AI passes can share errors. No completed independent human assurance, blinded human reliability study or empirical validation is claimed. Legacy decisions that lack exact-claim binding remain identified as legacy; a new queue must not relabel them as completed review.\n\nCurrent public exports contain aggregate figure inputs, their definitions, selected attributed facts and method information. They exclude individual company scoring records and sensitivity outputs. Readers can reproduce published chart arithmetic from its aggregate inputs and inspect cited sources. They cannot independently reconstruct every microdata decision from the current aggregate export alone. Mayden retains the detailed research records and private replay. Historical public bundles retain their own reproducibility contracts and limits. Hashes bind bytes or structured records, not factual truth.\n\nA source, claim and operating case are different units. A two-claim provenance pilot in earlier research does not establish a reconciled source-family graph across all 80 records. Required dependencies, mirrors, common assertion origins and case overlap need further review. Public-source checking can flag a changed page; it cannot automatically decide what claim changed or whether a practice continues.\n\n## Research synthesis and institutional comparison\n\nThe eleven reports draw on 34 selected research and methodological references in the retained corpus. Designs range from experiments and observational studies to benchmarks, normative frameworks and guidance. They are not 34 comparable experiments, a complete literature census or a new statistical meta-analysis. A writing-task estimate, contact-centre result and repository experiment cannot establish one enterprise-wide productivity effect. Each study retains design, setting, population, access depth and transfer limits.\n\nThe additional 11 October institutional review inspected 21 primary sources for eight benchmark analogues and six consultancy-publication examples. This was a purposive comparative review, not a causal study of commercial success. Provider descriptions show operating mechanisms and public offers; they do not prove audience size, revenue, retention or incremental consulting sales. Some evidence was abstract-only, with that limit retained.\n\nThe inspected BCG and McKinsey AI survey publications use aggregate comparisons. BCG's Most Innovative Companies 2023 also demonstrates that a consultancy can publish named rankings under a different construct. No client relationship or consent arrangement is inferred. Mayden's publication choice is contextual, not a universal claim about consultancy practice.\n\n## The future participant instrument\n\nKeep practice, outcome and evidence-confidence records separate. Specify the assessment entity, eligible work, reference period, respondent role, evidence unit and intended decision before choosing a scale. A candidate enterprise composite may be formative; reflective subscales require their own justification. High internal consistency does not validate a formative index, and correlations generated partly by the scoring formula cannot establish construct validity.\n\nBefore intake: appoint accountable humans; obtain content review; specify permissions, secure storage and retention; conduct English/Arabic cognitive interviews and mode checks; predeclare pilot questions and acceptance rules. Before benchmark claims: conduct independent coding and role checks, an appropriate test–retest design, sector/language comparability work, construct/outcome evaluation and sensitivity analysis. Analytical sample requirements follow intended comparisons and precision. A 40–60-company capacity scenario is not a statistically justified validation sample or regional norm.\n\nSeek several knowledgeable perspectives where feasible; preserve disagreement rather than averaging it away. Distinguish no, unknown, not applicable and unavailable evidence. Record workflow breadth and sustained use. No percentiles, certification or “validated score” until the relevant empirical and governance gates are passed. Paid advice and prior Mayden involvement must be recorded as potential conflicts and confounders.\n\n## Collection, governance and longevity\n\nThe collector is a manually invoked, bounded public-source checker with an approved URL register, access/robots checks, limits, pacing, hashes and retrieval receipts. HTML matches are leads; PDF collection is fingerprint-only. It has no private-system credentials, paid-data connector, continuous schedule or authority to accept, score or publish claims. It is not an official government crawler. DNS resolution remains subject to the host resolver outside its HTTP deadline. A production service would need an owner, validated extraction, source-family management, secure storage, monitoring and review capacity.\n\nThe accepted programme design separates research from commercial influence: no fee buys inclusion, favourable assessment or suppressed findings. Accountable human roles and external assurance have not yet been appointed. The proposed charter provides stop-release authority, conflict/recusal records, equal correction handling and versioned change decisions; implementation and actual sign-offs must be reported rather than assumed.\n\nFuture change reports must distinguish operating change from discovery, access recovery, coding correction, scope, sample composition and method change. Track attrition, mergers, respondent replacement, interventions and translation/mode revisions. Bridge studies are required before interpreting materially changed instruments on one trend line. Fund and accept the next wave before promising a cadence.\n\nMayden owns its original analysis and software subject to applicable rights, not third-party publications, public facts or a participant's entrusted material. The AI Native Index remains a working name; no trademark clearance or registration is claimed. The ambition is a durable organisational benchmark. This publication does not claim that ambition has already been achieved.\n\n## Selected methodological foundations\n\n[OECD/JRC composite-indicator handbook](https://doi.org/10.1787/9789264043466-en); [scale-development primer](https://doi.org/10.3389/fpubh.2018.00149); [formative/reflective measurement](https://doi.org/10.1037/0033-2909.110.2.305); [NIST AI RMF](https://www.nist.gov/itl/ai-risk-management-framework); [W3C provenance model](https://www.w3.org/TR/prov-dm/). These inform choices; none validates Mayden's instrument.\n", "collection": { "uniqueSources": 40, "retrievedSources": 19, "runs": 4, "lastRun": "2026-10-10T01:55:29.370889+00:00" }, "limits": [ "Purposefully selected corpus; no population prevalence, country ranking or representative GCC inference. Oman is outside this edition.", "Theme coverage is located documentation, not enterprise operating coverage. Mixed groups and subsidiaries can overlap.", "Source publications span different periods. Publication date does not confirm current use.", "Aggregate figures can be reproduced from sector chart inputs; company-level coding is retained privately. Historical editions remain explicit archives." ], "archive": "/research/editions/PE-2026.10-v4/index.html" }