Expand: New Source Leads, Grouped by the Gap They Close
Project Plan section 3.4 named three evidence gaps: no consistent definition or measure of AI value, little evidence directly linking trust and change readiness to financial results, and an evidence base dominated by Western consultancy research. The groups below attack those gaps, plus two more needs that surfaced in scoping: governance benchmarks for the maturity model, and fresher failure data. Each card states why it earns a place now.
Gap 1 · Causal, measured evidence of AI value (not self reported surveys)
The strongest weakness in the current base: almost every value number is a self reported executive survey. These are randomized and natural field experiments with measured outputs, the highest quality evidence available on what AI actually does to productivity and quality.
Navigating the Jagged Technological Frontier
Peer Reviewed JournalDell'Acqua, McFowland, Mollick, Lifshitz, Kellogg and colleagues ran a field experiment with 758 BCG consultants. Inside AI's capability frontier, consultants with AI finished more tasks, faster, at measurably higher quality; outside the frontier, AI use made results worse. Originally the famous 2023 HBS working paper, now published in a top peer reviewed journal.
Generative AI at Work
Peer Reviewed JournalBrynjolfsson, Li, and Raymond studied 5,172 customer support agents in a staggered real world rollout. AI assistance raised productivity 15% on average, with the largest gains for less experienced workers, improved customer sentiment, and improved employee retention.
Experimental Evidence on the Productivity Effects of Generative AI
Peer Reviewed JournalNoy and Zhang's randomized experiment on professional writing tasks: ChatGPT cut time taken by about 40% while raising output quality, and it compressed the gap between weaker and stronger performers rather than widening it.
The Cybernetic Teammate
Academic Working PaperDell'Acqua, Sadun, Lakhani and colleagues' follow up field experiment at Procter and Gamble: individuals working with AI matched the performance of two person teams without AI, and AI use broke down the usual silo between commercial and technical staff.
Artificial Intelligence, Firm Growth, and Product Innovation
Peer Reviewed JournalBabina, Fedyk, He, and Hodson measure AI investment across US firms using resume based data and find that AI investing firms grow faster in sales, employment, and market value, with the growth driven by product innovation. Caution: the design is observational, and the gains show up in growth more than in measured productivity.
Gap 2 · Linking trust, readiness, and change to outcomes
The plan concedes that trust evidence measures perception, not results. These sources get closer to the behavior and leadership side of the link.
Superagency in the Workplace
Consultancy ResearchSurvey finding that employees are far more ready for AI than leaders believe: employees are already using AI for a large share of their work while leaders underestimate usage by roughly three times, and the biggest barrier to scaling is leadership, not employee resistance.
Work Trend Index
Vendor ResearchMicrosoft and LinkedIn's annual study of tens of thousands of workers across 31 countries, tracking real usage telemetry alongside survey data: employee adoption running ahead of official programs, the rise of human agent teams, and the emergence of frontier firms.
Gap 3 · Widening the lens beyond Western consultancy data
The plan names the Western, consultancy heavy skew as a limitation. These intergovernmental datasets cover most of the world with published methods and no product to sell.
AI Preparedness Index
Intergovernmental DatasetIMF index covering 174 economies across digital infrastructure, human capital and labor policies, innovation, and regulation. Shows advanced economies better positioned but more exposed to disruption, while emerging markets face a widening readiness gap.
OECD.AI Policy Observatory
Intergovernmental ObservatoryLive database of national AI strategies, policies, and incidents across 70 plus jurisdictions, plus the OECD AI Principles that most national frameworks trace back to. Includes trend data on investment, skills, and regulation by country.
Gap 4 · Governance benchmarks the maturity model can anchor to
Dr. Kokkonen asked for a maturity model with stage gates. These three are the recognized external benchmarks a governance dimension can be graded against, which keeps the framework from being arbitrary.
AI Risk Management Framework (AI RMF)
Government FrameworkThe US national framework for trustworthy AI: Govern, Map, Measure, Manage. Voluntary, widely adopted by enterprises as the de facto governance baseline, with a Generative AI profile added for GenAI specific risks.
ISO/IEC 42001:2023 AI Management Systems
International StandardThe first certifiable international standard for AI management systems, covering governance, risk, lifecycle management, and continual improvement. Organizations can now be audited and certified against it.
The EU AI Act: Regulatory Framework on AI
RegulationThe world's first comprehensive AI law, phasing in through 2026 and 2027 with risk tiered obligations and significant penalties. The reference point for Europe's regulation first posture and for compliance as a value protection activity.
Gap 5 · Fresher failure and abandonment data
The initial base leans on the contested MIT 95% figure and Gartner's forward prediction. These add measured abandonment rates and systematically studied root causes.
The Root Causes of Failure for AI Projects
Policy Research InstituteRAND interviewed 65 experienced data scientists and engineers and synthesized five root causes of AI project failure, led by leaders misunderstanding or miscommunicating the problem to be solved, with inadequate data and infrastructure close behind. Technology shortcomings rank last.
AI Abandonment Jumps from 17% to 42%
Analyst ResearchS&P's survey of 1,000 plus enterprises across North America and Europe: the share of companies abandoning most of their AI initiatives jumped from 17% in 2024 to 42% in 2025, with an average 46% of proofs of concept scrapped before production. Cost, data privacy, and security cited as top causes.
30% of GenAI Projects Abandoned After Proof of Concept
Analyst ResearchThe predecessor to the agentic prediction already carded: Gartner forecast at least 30% of generative AI projects would be abandoned after proof of concept by end of 2025, citing poor data quality, inadequate risk controls, escalating costs, and unclear business value.
Sources Cited in the Project Plan, Added to the Hub
Five sources the Project Plan cites that were not yet carded, closing priority gap 8. Two anchor the maturity model build for O5 and two are the method sources behind the plan's integrative review and thematic synthesis.
Building Enterprise AI Maturity
AcademicWeill, Woerner, and Sebastian define four stages of enterprise AI maturity from research on 721 companies, with financial performance rising by stage. Caution: the evidence is correlational, not causal.
Developing Maturity Models for IT Management
Peer Reviewed JournalBecker, Knackstedt, and Pöppelbuß set out a tested procedure model for building maturity models, including the design steps and evaluation criteria a defensible model must meet.
Using Thematic Analysis in Psychology
Peer Reviewed JournalBraun and Clarke's standard reference for thematic analysis: the six phase process the team's thematic synthesis of qualitative evidence follows.
Literature Review as a Research Methodology
Peer Reviewed JournalSnyder's overview and guidelines for literature reviews, including the integrative review design the plan's research approach is built on.
Global AI Adoption in 2025: A Widening Digital Divide
Vendor ResearchCountry level AI diffusion data: 64.0% of the working age population in the UAE and 60.9% in Singapore use AI, against 28.3% in the US, with the divide between high and low income economies widening.
Public Cases for the Four Stage Walk Through (O4 and O5)
Five shortlisted public cases that trace capability, adoption, outcome, and sustained value inside one organization, across industries, regions, success, failure, and reversal. Added 9 October to move O4 to Strong; the team picks three at the 14 October review, and Harvard entries for the chosen cases' articles get added then. Figures are company or media reported, so verify each against the linked source before citing.
Klarna: The Full Chain, Reversal Included
Public CaseThe OpenAI built assistant launched in early 2024 handling two thirds of service chats, doing the work of about 700 agents with a projected $40 million profit improvement. By May 2025 the CEO was hiring human agents back on quality grounds while revenue per employee climbed toward $1 million. Capability, adoption, outcome, and a contested sustained value stage, all on the public record.
McDonald's Drive Thru AI: Capability Without Outcome
Public CaseA roughly two year automated order taking test across more than 100 US drive thrus, ended in mid 2024 after widely shared accuracy failures. A clean case of task capability that never converted into an operating outcome, with the partnership wound down rather than scaled.
JPMorgan: Scaled, Sustained Enterprise Value
Public CaseLLM Suite rolled out to roughly 200,000 employees with publicly described use cases across advisors, operations, and engineering, backed by one of the largest technology budgets in banking. The success at scale candidate. Caution: value figures are company self reported at investor events, so verify current numbers before citing.
IKEA: Redeployment Instead of Replacement
Public CaseThe Billie bot absorbed a large share of routine customer inquiries while 8,500 call center workers were retrained as interior design advisors, turning a cost center into a revenue stream. Adoption with workforce transition rather than headcount removal.
Commonwealth Bank: The Retracted Outcome Claim
Public CaseAustralia's biggest bank announced 45 role cuts tied to its AI chatbot in 2025, then reversed the decision within weeks and called it a mistake after the union showed call volumes had not fallen as claimed. An outcome claim that failed public scrutiny.
Primary evidence the team can create
Expanding is not just finding documents. The team's own position gives it primary evidence no consultancy has, and Ducere explicitly invites surveys and personas as deliverables. Anything involving people goes through the plan's section 5 ethics and approval gate first.
Practitioner Interview Program
Primary EvidenceSix to ten short semi structured interviews with leaders and practitioners across the team's networks, using the LinkedIn channel Dr. Kokkonen opened (she accepts all connection invites and encouraged cross group networking). Target a spread of industries, regions, and seniority, asking how they define AI value, where it leaked, and what they would measure differently.
Stakeholder Perception Pulse Survey
Primary EvidenceA short anonymous survey built on the Melbourne and KPMG question structure, distributed through the team's combined networks across the six RFS stakeholder groups. Even 50 to 100 responses gives the perception heat map an original data layer to compare against the global benchmarks.
DXC Public Materials Sweep
Primary EvidenceA systematic pass through DXC's own public record: the full AdvisoryX study download, annual report and investor materials (how DXC describes AI demand to shareholders), newsroom releases, and Dr. Kokkonen's publication and speaking trail including her Trust in Government's use of AI talk.
Cross Industry Mini Cases from the Team Itself
Primary EvidenceFive members spanning software sales (UK and global), digital marketing operations (London), government contracting operations (Hawaii), and construction and refrigeration services (Dallas). Each member writes a one page structured vignette of AI value creation or value loss observed in their own sector, using a common template.