peopleanalyst

AI Human Interaction Guide · Appendix A

Case Studies Catalog

Appendix A. A reference catalog of the cases the seven parts draw on — aggregate sector studies, documented enterprise failures, successful people-analytics work, diffusion-mechanism field experiments, and workforce-side adoption signals.

How to use this catalog

The guide cites these cases inline where they matter, and aggregates them here so the reader can see the empirical record in one place. Entries are organized by the kind of evidence each case provides — sector-aggregate, single-deployment, successful-program, diffusion-mechanism, workforce-signal — rather than by topic or chronology. Each entry closes with a Where this appears in the guide pointer.

The catalog is not exhaustive. It covers the cases the guide's argument depends on plus the cases a reader would expect a guide to this material to address. For the broader corpus, the substrate manifests are at content/research/ai-encyclopedia/manifests/ in the source repository.

A.1 Aggregate sector studies

These are the multi-organization studies that produce the 95% organizational-failure rate stylized fact the guide treats as load-bearing for Parts II and VII.

Stanford 51-deployment analysis (Pereira, Graylin, Brynjolfsson — April 2026)

The Stanford Digital Economy Lab's Enterprise AI Playbook: Lessons from 51 Successful Deployments studied 51 production AI deployments across 41 organizations, 7 countries, 5 regions, and more than 1 million collective employees.1 The methodology paired structured interviews with deployment teams against quantitative measures of business impact, organizational stability, and operational outcomes.

The headline finding is that roughly 95% of enterprise AI failures across the broader population (i.e., outside the curated successful-deployment set) trace to organizational variables rather than technical ones. The four organizational predictors the study names for success:

  1. Workflow mapping before technology selection — successful deployments invested in mapping the workflow the AI would integrate into before choosing the tooling; unsuccessful ones picked the tool first.
  2. Day-one governance architecture — the governance framework (oversight, deployment-gate criteria, monitoring) was in place when the deployment started, not bolted on after problems appeared.
  3. Pre-production observability — the instrumentation to detect AI failures (drift, hallucination, calibration loss) was deployed before the AI was in front of real users.
  4. 18-month leadership continuity — the executive sponsoring the deployment remained in role for at least 18 months. Sponsor turnover was the single most predictive failure precursor.

Stanford's framing is descriptive — this is what worked — rather than mechanistic. The guide reads the four predictors as operationally adjacent to a network-topology mechanism the study stops short of naming explicitly; the structural-network-variable claim is the guide's contribution. Where this appears: Part II §2.6; Part VII §7.1, §7.3.

MIT NANDA GenAI Divide (August 2025)

MIT's NANDA project (Networked AI Agents in Decentralized Architecture) published The GenAI Divide: State of AI in Business 2025 in August 2025.2 Methodology combined 150 senior-leader interviews, a 350-employee survey, and a 300-deployment public-record analysis. The headline finding — that 95% of enterprise GenAI pilots fail to deliver measurable financial impact — converged independently with Stanford's 95% figure and made the stylized fact citable as more than a single-source claim.

NANDA's diagnosis differs from Stanford's: the failure mechanism MIT names is the learning gap — too much investment at the pilot stage, not enough investment at the scaling stage; pilots that demonstrate AI's capability never translate to the methodology, governance, and operational discipline scaling requires. The shape of the failure pattern is consistent with Stanford's organizational-predictor account but framed at the investment-stage level rather than the predictor level.

The report's policy implication is that enterprises should treat pilot-success as a signal that the methodology stack needs further investment, not as a signal that the deployment is ready to scale. The guide's network-mediated-adoption argument in Part VII extends this — what the conventional rollout playbook treats as scaling is structurally the wrong move; the topology-aware alternative is saturation inside the seed cohort's tie clusters before broadcast expansion. Where this appears: Part II §2.6; Part VII §7.1, §7.2.

McKinsey State of AI 2025 (November 2025)

McKinsey QuantumBlack's annual State of AI report (Singla, Sukharevsky, Yee, et al., November 2025) surveyed 1,993 respondents across 105 countries.3 Headline numbers: 88% of organizations report using AI in at least one function; only 39% attribute any EBIT impact to AI; only 6% — the "high performers" cohort — report EBIT impact greater than 5%.

The report distinguishes between AI use (operational presence) and AI value (measurable EBIT contribution). The gap between the two — 88% use versus 6% material-value capture — is the operationally precise version of the broader 95% failure-rate finding. McKinsey's analytical layer adds organizational-practice variables that distinguish the high-performer cohort: redesigned workflows, AI-specific governance, dedicated talent-development pipelines, and risk-management practices integrated into deployment rather than appended after-the-fact.

The high-performer profile aligns with Stanford's success predictors and MIT's investment-stage diagnosis. The convergence across three independent methodologies — interview-based deployment study, mixed-methods aggregate, large-N global survey — is what makes the underlying pattern citable as a stylized fact rather than as any single firm's reading. Where this appears: Part II §2.6.

BCG Widening AI Value Gap (September 2025)

BCG's Build for the Future 2025 report quantifies the bimodal distribution at the center of the topology argument: 60% of companies are achieving little or no material AI value; 5% are "future-built" AI leaders capturing substantial value at scale.4 The middle ~35% sits between the failure cohort and the leader cohort, with mixed value capture.

BCG's diagnostic adds a temporal dimension McKinsey's snapshot doesn't capture: the gap is widening. The future-built leaders' value capture is accelerating relative to the trailing cohort. The mechanism BCG names is infrastructural — the leaders invested in foundational capabilities (data quality, governance, talent, change management) before the GenAI wave, and the wave amplified the lead.

The bimodality matters for the guide's argument because it isolates the organizational variance component of AI success from the capability variance component. All four reports — Stanford, MIT NANDA, McKinsey, BCG — operate against approximately the same capability frontier (foundation models from the major labs, with broadly similar performance characteristics). The 60/35/5 distribution is therefore not explained by which capability each firm bought; it's explained by which organization each firm is. Where this appears: Part II §2.6; Part VII §7.1.

HBR Beware the AI Experimentation Trap (August 2025)

The Harvard Business Review piece by an HBR editorial-board author (August 2025) frames the pilot-stuck pattern documented in the sector aggregates as the AI experimentation trap.5 The argument: enterprises invest in pilot-stage experiments that demonstrate AI capability but never produce the methodology, governance, or operational discipline that translates capability into business value. The trap is structural rather than technical — the experimentation pattern is the failure mode, not a precursor to it.

The framing language has become a rhetorical handle for the broader 95% pattern. The guide engages with the framing in Part II's adoption-vs-readiness section and in Part VII's network-mediated-adoption argument: the experimentation trap is the operational consequence of the topology mismatch between pilot environments (clustered topology, high local density) and broader-rollout environments (small-world topology, low local density). Where this appears: Part II §2.6; Part VII §7.2.

A.2 Documented enterprise failures (deployment-level)

These are the specific named-organization deployments that anchor the failure-mode arguments in Parts III, IV, and V.

IBM Watson Health × MD Anderson (2013-2017)

The IBM Watson × MD Anderson Cancer Center partnership is the canonical AI-deployment-failure case. Announced in 2013 as a $62M project to deploy IBM Watson for oncology decision support, the partnership was terminated in 2017 after a University of Texas internal audit raised concerns about cost overruns (final cost exceeded $62M), missed deliverables, and clinical-deployment readiness.6

The technical issues, per subsequent reporting in STAT News (Casey Ross) and Forbes (Matthew Herper) and clinical-medicine commentary in JAMA: Watson's training methodology relied on hypothetical patient cases generated by MD Anderson clinicians rather than real patient records, which left the system poorly calibrated to actual clinical decision contexts; Watson's recommendations included unsafe treatment options in evaluation testing; the integration with MD Anderson's existing electronic-health-record systems never materialized at clinical scale.7

The organizational issues: the project was sponsored at the institution's leadership level but never integrated into the clinical-decision workflow at the unit-of-care level. The methodology gap from Part I §1.3 — AI rollouts are not software rollouts — is the structural account: the project was scoped, governed, and executed using software-project methodology in a context that required clinical-deployment methodology (which is itself a specialized discipline with its own rigor requirements).

The case became canonical because of the visibility of the institutions involved, the size of the investment, and the clinical-decision-stakes of the failure. It is cited across the consulting literature and academic commentary as the cautionary example AI-deployment teams are warned about — though, on the evidence of the convergent 95% failure rate across the post-2024 GenAI wave, the cautionary value has not generalized. Where this appears: Part IV §4.5; Part V §5.5.

Google Glass Enterprise (2013-2023)

Google Glass shipped as the Explorer Edition in 2013, was withdrawn from consumer markets in 2015 after sustained privacy and form-factor controversy, then re-launched as Glass Enterprise Edition in 2017 for industrial-use cases (warehouse picking, surgical assistance, field-service technician support). Glass Enterprise Edition 2 followed in 2019. In March 2023, Google announced it was discontinuing Glass Enterprise Edition entirely.8

The Enterprise pivot worked in narrow vertical deployments — pre-pandemic Boeing, GE, AGCO, and a handful of healthcare deployments produced documented productivity gains. The cases where Glass Enterprise produced value shared structural features: high-task-repetition contexts where heads-up information delivery saved measurable seconds per task, dedicated training programs, and integrated workflows around the device. Where Enterprise deployments failed, the failure mode was usually environmental (devices broke or were lost) or workflow-misfit (the heads-up benefit didn't compound across the actual work pattern).

The Enterprise discontinuation in 2023 wasn't a single failure event so much as a slow-decline outcome: hardware iteration stopped, vertical-specific software partners stopped investing, and the total addressable market never reached the scale required for sustained product support.

The case is in this catalog as the successful-but-not-scalable counterexample to the Watson Health failed-and-canceled shape. Both are enterprise-AI failures by the time-horizon outcome metric, but the structural account differs: Watson was a methodology-gap failure; Glass Enterprise was a market-structure failure. Where this appears: Part IV §4.1.

Air Canada chatbot ruling (February 2024)

The British Columbia Civil Resolution Tribunal ruled in February 2024 that Air Canada was liable for misinformation its customer-service chatbot provided to a passenger about bereavement-fare eligibility.9 The chatbot fabricated a policy stating bereavement-fare claims could be retroactively applied to already-purchased tickets; the airline argued the chatbot's statement was not binding because it was "a separate legal entity." The Tribunal rejected the argument and awarded the passenger $812 CAD.

The case is precedent-setting for AI-deployment liability: courts will hold companies responsible for fabricated outputs from their AI customer-service systems. The legal framing — chatbot outputs are agent statements binding on the principal — extends the conventional law of agency to AI systems without requiring new statute.

For the guide's argument, the case is the empirical anchor for Part III §3.3 (conversational AI in customer service): the fabrication risk is not theoretical; it is operationalized through legal liability the deploying organization bears. The structural correction — anchoring reasoning to verifiable substrate, calibration instrumentation, deployment gates — is what the Part III chapter advocates. Where this appears: Part III §3.3; Part V §5.2.6.

Cursor "Sam" support-agent fabrication (April 2025)

In April 2025, Cursor's AI-powered customer-support agent (named "Sam") fabricated a policy in response to a user inquiry about device-limit handling. The user had asked whether the recent device-limit warning was real; Sam responded that the company had a (non-existent) policy preventing simultaneous use across multiple devices on a single subscription. The exchange went viral on Hacker News and tech-press, and Cursor publicly apologized and clarified that no such policy existed.10

The Cursor case is in this catalog for a specific reason: Cursor is an AI-native company whose product is itself an AI tool used by software engineers. The expectation — externally and presumably internally — would be that an AI-native firm has the discipline to deploy AI support systems with adequate guardrails against fabrication. The fact that Cursor produced exactly the Watson-style fabrication failure six months after entering the support agent market is empirical evidence that the methodology gap from Part I §1.3 applies even to AI-native deployers, and is not solved by the deploying organization's AI sophistication.

The structural correction is identical to Part III §3.3's recommendation: verifiable-substrate anchoring; deployment gates that require ground-truth retrieval before generative response; calibration instrumentation that flags low-confidence outputs to the user before they hit them. Where this appears: Part III §3.3; Part V §5.2.6.

A.3 Successful enterprise people-analytics cases

These are positive cases that demonstrate what people-analytics work looks like when it works. The guide draws on them as evidence that the underlying methodology is mature — the problem is not that people-analytics doesn't work; it's that the typical deployment doesn't reach the discipline these cases demonstrate.

Google Project Oxygen (2008-present)

Project Oxygen was Google's internal people-analytics study of what makes a good manager, conducted by Prasad Setty's People Analytics team beginning 2008-2009.11 The methodology: regression analysis of manager-quality scores against engineer satisfaction, retention, and team performance, against initially-skeptical engineers who believed (per Google culture) that technical excellence was the only manager quality that mattered.

The findings, published in HBR (Garvin 2013) and the Google re:Work hub, identified eight (later ten) manager behaviors that predicted team outcomes — being a good coach; empowering rather than micromanaging; expressing interest in team-member success; being productive and results-oriented; being a good communicator; helping with career development; having a clear vision and strategy; having technical expertise to advise the team.12 The technical-expertise behavior ranked eighth in predictive importance, against the engineering-culture prior that it would rank first.

Project Oxygen continues to inform Google's manager-training programs and has become the canonical citation for people-analytics work that surfaces non-obvious behavioral findings in practitioner literature. The methodological move — letting the regression speak rather than confirming the cultural prior — is what makes it citable as evidence that disciplined people analytics produces findings the conventional consulting framework does not. Where this appears: Part II §2.5.

Google Project Aristotle (2012-2015)

Project Aristotle was Google's follow-on people-analytics study of what makes effective teams, published in NYT Magazine (Charles Duhigg, February 2016).13 The study analyzed 180+ Google teams across two years on group-effectiveness measures, looking for the team-composition variables (skill mix; tenure mix; introvert/extrovert balance; reporting structure) that predicted outcomes.

The finding — that psychological safety (the team norm that members can take interpersonal risks without retaliation) was the single strongest predictor of effectiveness — surprised both the research team and the broader management literature. Psychological safety, defined and instrumented in Amy Edmondson's earlier academic work, became the practitioner-recognizable construct people analytics teams could instrument and intervene on.14

The methodological move — disciplined empirical testing of management folk-wisdom against outcome data — is the canonical example of people-analytics work that produces practitioner-changing findings. The guide cites it as evidence that the field has the methodology to surface non-obvious answers; the gap is deployment of that methodology, not existence of it. Where this appears: Part II §2.5.

Microsoft Workplace Analytics → Viva Insights (2015-present)

Microsoft Workplace Analytics launched in 2015 as a productivity-mining tool built on Office 365 collaboration metadata (meeting patterns, email flows, focus-time fragments). The MyAnalytics consumer product followed, presenting individual-level dashboards. In 2020, the productivity-score feature triggered a privacy controversy when European regulators and labor-organization advocates raised concerns about surveillance-grade individual measurement.15

Microsoft responded by re-architecting the product around privacy-by-design principles: aggregation thresholds (minimum N before any cohort can be surfaced); anonymization at the substrate layer; manager-side dashboards that show team-level patterns without individual identification; opt-in for personal MyAnalytics features. The renamed Viva Insights product (2021) reflects the privacy architecture.

The case is in this catalog because it illustrates two simultaneous things: people-analytics products operate against real privacy constraints that determine product viability; and those constraints can be addressed at the substrate level (min-N gates, aggregation, anonymization) rather than at the policy layer (which depends on enforcement). The structural-correction lesson — substrate primitives outperform configurable policies — is the same lesson the guide names in Part VI §6.2 and Part VII §7.3. Where this appears: Part II §2.5; Part VI §6.2.

A.4 Diffusion-mechanism field evidence

These are the empirical studies of the contagion mechanism the guide's network-mediated-adoption argument in Part VII rests on.

Sociological Science Vol. 12 — country-scale complex-contagion RCT (October 2025)

The October 2025 issue of Sociological Science (Vol. 12) published a country-scale randomized controlled trial of the complex-contagion mechanism, using a peer-encouragement design that manipulated whether participants were exposed by one friend or by two friends to a behavioral signal.16 The trial was conducted at population scale (effectively a national field experiment), with causal identification via random assignment to exposure conditions.

The headline result: signals from multiple sources interact rather than acting as independent cascades. Two-friend exposure produced significantly larger adoption-rate increases than the linear combination of two single-friend exposures would predict. The result is consistent with Centola's lab-experiment finding from Science 2010 (clustered topologies support complex contagion; small-world topologies do not), now replicated at country-scale in a non-laboratory setting.

For the guide's argument, this is the strongest available empirical anchor for treating AI adoption as a complex contagion: AI-adoption decisions in organizations have the structural shape of multi-exposure-required behaviors, and the topology through which adoption propagates is therefore load-bearing for whether it propagates. The 60-year diffusion canon (Rogers, Granovetter, Burt) supplied the mechanism hypothesis; Centola 2010 supplied the lab evidence; the Sociological Science 2025 trial supplied the country-scale field replication. Where this appears: Part VII §7.2, §7.4.

Centola 2010 + 2018 — complex contagion canon

Damon Centola's 2010 Science paper ("The Spread of Behavior in an Online Social Network Experiment") was the foundational laboratory demonstration that behaviors require clustered network topology to propagate, while information flows readily through small-world topology.17 Centola's 2018 book How Behavior Spreads synthesized the experimental record across multiple behavioral domains and methodological replications.

The mechanism the work isolates: complex contagions (behaviors requiring multiple reinforcing exposures from trusted ties) propagate through tie clusters where each adopter is surrounded by other adopters in their immediate network. Small-world graphs — which characterize most large organizations by structure — are excellent at moving information but bad at moving practice. The 2025 Sociological Science trial replicated the finding at country-scale; PMC/PNAS Nexus 2024 and Phys. Rev. Research 2025 extended the model to belief-system dynamics and temporal networks respectively.1819

The body of work is the structural foundation for the network-mediated-adoption argument the guide advances in Part VII: AI adoption is a complex contagion; pilots succeed because their teams sit in clustered topologies by selection accident; rollouts fail because they push adoption into the broader organization's small-world topology. Where this appears: Part VII §7.2, §7.4.

A.5 Workforce-side adoption signals

These are signals from the workforce side of the adoption equation that ground the workforce dimension of the guide's argument in Part II.

Gallup State of the Global Workplace 2025

Gallup's 2025 State of the Global Workplace report documented global employee engagement declining from 23% to 21% year-over-year — the sharpest single-year decline since the COVID-19 pandemic.20 The report names manager engagement as the binding constraint on team engagement: 70% of the variance in team engagement traces to the manager.

The methodological framing — that manager-level variables are the single largest controllable input into engagement outcomes — is structurally adjacent to the People Analytics Toolbox's CAMS (Continuous Action Measurement System) instrument's binding-constraint move from Part II §2.5. Gallup operates at aggregate-survey scale; CAMS operates at within-organization instrumentation scale. Both name the manager as load-bearing.

For the guide's argument, the Gallup finding matters because it identifies the organizational layer at which AI-adoption topology operates: managers are the trust-graph bridge actors whose adoption decisions propagate to their teams, and whose disengagement (per the 2025 trend) is itself a signal that the trust-graph substrate AI adoption depends on is under stress. Where this appears: Part II §2.5; Part VII §7.2.

Writer 2026 employee-sabotage survey

Writer's Q1 2026 enterprise-AI-adoption survey of 1,200 employees and 1,200 C-suite executives reported that 29% of employees admitted actively sabotaging their company's AI strategy; the figure rose to 44% among Gen Z respondents. 54% of C-suite respondents said AI adoption was "tearing the company apart."21

The survey is a vendor-published source (Writer is an enterprise-AI vendor with commercial stake in the framing), so it falls under the credibility caveats the guide applies to non-peer-reviewed sources. The methodology — survey-instrument framing, sample selection, question wording — is not independently validated.

The signal is included here because it is internally consistent with the convergent organizational-failure rate and because the sabotage framing operationalizes a mechanism the consulting literature does not name: workforce-side resistance is not passive (slow uptake) but active (deliberate counter-action). If the survey's directionality is roughly correct — and the cross-source plausibility check makes the directional claim plausible even if the precise percentages should be treated with vendor-source caveats — the active-resistance mechanism is a workforce-topology variable the guide's network-mediated-adoption argument has not fully integrated. The Field-Reporting initiative (PA-400+ in the work queue) is the closer empirical anchor the guide should converge to as field evidence accumulates. Where this appears: Part II §2.5; Part VII §7.3.

Cross-reference index

CasePrimary cross-referenceSecondary cross-references
Stanford 51-deployment (2026)Part II §2.6Part VII §7.1, §7.3
MIT NANDA GenAI Divide (2025)Part II §2.6Part VII §7.1, §7.2
McKinsey State of AI 2025Part II §2.6—
BCG Widening AI Value Gap (2025)Part II §2.6Part VII §7.1
HBR Experimentation Trap (2025)Part II §2.6Part VII §7.2
IBM Watson Health × MD AndersonPart IV §4.5Part V §5.5
Google Glass EnterprisePart IV §4.1—
Air Canada chatbot ruling (2024)Part III §3.3Part V §5.2.6
Cursor "Sam" fabrication (2025)Part III §3.3Part V §5.2.6
Google Project OxygenPart II §2.5—
Google Project AristotlePart II §2.5—
Microsoft Workplace Analytics → Viva InsightsPart II §2.5Part VI §6.2
Sociological Science Vol. 12 trial (2025)Part VII §7.2Part VII §7.4
Centola 2010 + 2018 canonPart VII §7.2Part VII §7.4
Gallup State of the Global Workplace 2025Part II §2.5Part VII §7.2
Writer 2026 sabotage surveyPart II §2.5Part VII §7.3

Footnotes

  1. Pereira, Elisa, Alvin Wang Graylin, and Erik Brynjolfsson. The Enterprise AI Playbook: Lessons from 51 Successful Deployments. Stanford Digital Economy Lab, April 2026. https://digitaleconomy.stanford.edu/publication/enterprise-ai-playbook/. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 1 (quality rating 6/7, action-recommendation: read-now). ↩

  2. MIT NANDA (Networked AI Agents in Decentralized Architecture). The GenAI Divide: State of AI in Business 2025. August 2025. Primary report behind the Fortune secondary coverage at https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 9. ↩

  3. Singla, Alex, Alexander Sukharevsky, Lareina Yee, et al. The State of AI in 2025: Agents, Innovation, and Transformation (Global Survey). McKinsey QuantumBlack, November 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 2. ↩

  4. BCG Build for the Future team. The Widening AI Value Gap: Build for the Future 2025. BCG, September 2025. https://media-publications.bcg.com/The-Widening-AI-Value-Gap-Sept-2025.pdf. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 3. ↩

  5. Beware the AI Experimentation Trap. Harvard Business Review, August 2025. https://hbr.org/2025/08/beware-the-ai-experimentation-trap. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 10. ↩

  6. University of Texas System Administration. Audit of the Oncology Expert Advisor Project at the University of Texas MD Anderson Cancer Center. Audit Report, 2016. The audit identified procurement-process and project-management issues; subsequent reporting documented the clinical-deployment difficulties. ↩

  7. Ross, Casey. IBM's Watson supercomputer recommended "unsafe and incorrect" cancer treatments, internal documents show. STAT News, July 2018. Herper, Matthew. MD Anderson Benches IBM Watson In Setback For Artificial Intelligence In Medicine. Forbes, February 2017. Multiple JAMA editorials in 2017-2018 commented on the clinical-AI-readiness implications of the case. ↩

  8. Google. Glass Enterprise Edition Support Resources (discontinuation notice). March 2023. Press coverage at https://www.theverge.com/2023/3/15/23640110/google-glass-discontinued-enterprise-edition-2. ↩

  9. Moffatt v. Air Canada. British Columbia Civil Resolution Tribunal, Case 2024 BCCRT 149. February 14, 2024. Coverage at the Globe and Mail, BBC, and Reuters. The Tribunal awarded the plaintiff $812 CAD plus tribunal fees and ruled that the airline's chatbot statements were binding on the airline as agent statements. ↩

  10. Hacker News thread and tech-press coverage, April 2025. Cursor publicly responded acknowledging the fabrication and clarifying no such policy existed. The incident is widely cited in the AI-deployment-failure literature post-April 2025 as the canonical "AI-native firm produces Watson-style fabrication failure" case. ↩

  11. Setty, Prasad, and the Google People Analytics team. Project Oxygen, internal study, 2008-2009. Methodology and findings documented at the Google re:Work hub (https://rework.withgoogle.com/). ↩

  12. Garvin, David A. How Google Sold Its Engineers on Management. Harvard Business Review, December 2013. The HBR piece is the canonical public summary of the Project Oxygen methodology and findings, including the eight (later ten) manager behaviors. ↩

  13. Duhigg, Charles. What Google Learned From Its Quest to Build the Perfect Team. The New York Times Magazine, February 25, 2016. https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-the-perfect-team.html. The piece is the canonical public summary of Project Aristotle's methodology and the psychological-safety finding. ↩

  14. Edmondson, Amy C. Psychological Safety and Learning Behavior in Work Teams. Administrative Science Quarterly 44, no. 2 (1999): 350-383. The academic anchor underneath the Project Aristotle finding; Edmondson's subsequent work (The Fearless Organization, 2018) is the practitioner-facing synthesis. ↩

  15. Microsoft. Workplace Analytics → MyAnalytics → Viva Insights documentation, 2015-2021. The privacy controversy is documented in coverage at the Guardian, Reuters, and AlgorithmWatch, December 2020. Microsoft's response, including the substrate-layer privacy architecture, is at https://learn.microsoft.com/en-us/viva/insights/privacy/privacy-architecture. ↩

  16. Authors per Sociological Science Vol. 12, October issue (Network Science Institute affiliation). Complex Contagion in Social Networks: Causal Evidence from a Country-Scale Field Experiment. Sociological Science 12 (October 2025): article 28-685. https://sociologicalscience.com/articles-v12-28-685/. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 4 (quality rating 7/7). ↩

  17. Centola, Damon. The Spread of Behavior in an Online Social Network Experiment. Science 329, no. 5996 (September 2010): 1194-1197. The foundational laboratory demonstration that clustered topology supports complex contagion; small-world topology does not. ↩

  18. Emergence of simple and complex contagion dynamics from weighted belief networks. PMC / PNAS Nexus (PMC11014438), 2024. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 5. Extends the contagion model to belief-system dynamics in clustered networks. ↩

  19. Competition between simple and complex contagion on temporal networks. Physical Review Research, 2025. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 6. Models temporal-network setting where contagion-mechanism switching occurs. ↩

  20. Gallup workplace research team. State of the Global Workplace 2025. Gallup, 2025. https://www.gallup.com/workplace/708071/global-employee-engagement-continues-decline.aspx. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 12. ↩

  21. Writer research team. Enterprise AI Adoption 2026. Writer (vendor-published survey), Q1 2026. https://writer.com/blog/enterprise-ai-adoption-2026/. Cataloged in manifests/new-candidates-2026-05-14.json as Finding 11 (quality rating 3/7 — vendor source caveat applied). ↩

← Previous

VII.Network-Mediated AI Adoption

Next →

B.Practitioner's Toolkit + Further Reading

← All guide parts