An argument is only as good as the evidence it asks for.

Scope of the claims

Proof, analogy and evidence are different things.

Three kinds of claim appear on this site, and each keeps its own label. A research lesson states a model result: Lean, a program that checks every step of a proof, confirms that the conclusion follows from the stated assumptions. The business article translates those results for executives and adds no new assumption. The management book is the author’s judgment and says so.

Public examples of companies are analogies, never evidence that the proposal works. Retail category management is the organizing analogy; P&G’s 1931 brand-management memo is a different historical development. Stories and numbers about the fictional company in the book are illustrative.

Connected conversations

Ten questions the literature helps with.

Primary-source register · 133 sources

Read the claim and its limit together.

1. GS1 Switzerland · Undated

   Category Management: process and training

   Method: Primary retail category-management process page inspected; practitioner framework.

   Category definition, role, assessment, objectives, implementation and review connect a shopper need to coordinated action.

   Limit

   Retail assortment, shopper behavior and shelf allocation are different from developing AI capabilities. The framework does not establish ACM efficacy.

   What it means here

   Borrow need-based grouping, explicit roles and recurring review; define the technical evidence separately.

2. Strategyzer · Undated

   The Business Model Canvas

   Method: Official framework description and template.

   The canvas arranges value creation, delivery and capture as connected parts of one business model.

   Limit

   A visual framework is a way to organize assumptions; its appeal is not evidence that an ACM design works.

   What it means here

   Present a complete arrangement and show dependencies. Filled organizational examples are more useful here than a blank project form.

3. Aquia, Collected · 2026-10-18

   The Theorem Does Not Care

   Method: Published author essay and proposed apprenticeship practice; complete published essay inspected. Not an empirical intervention study.

   A checked machine-produced result can seed reverse analysis: isolate the central idea, vary an assumption, seek a counterexample, examine a simpler case, and test transfer. Proof validity, why an idea is fruitful, and a trace of machine generation are distinct. Public access to an answer is different from durable access to the instrument and compute needed to pursue the next question.

   Limit

   The author proposes a curriculum mechanism, not measured evidence that artifact laboratories increase independent expertise or a report of ACM company effects.

   What it means here

   Offer a bounded paid artifact-analysis lab as a candidate curriculum; assess independent explanation, counterexample-finding, simplification and transfer, and price access/review time. Existing practice/scaffold actions can represent it only through assumed learning and cost parameters.

4. Anthropic · 2026-10-16

   How employment is changing at Anthropic

   Method: Observational internal employment, tool-use and output metrics, survey records, and 17 employee interviews at one firm. Inspection: Full Anthropic web research report inspected; linked PDF not read.

   Anthropic's internal 2025–26 descriptive study reports a falling share of junior technical employees and junior technical hires alongside growing agent work and a reported human attention/review bottleneck.

   Limit

   One fast-growing AI firm, observational internal metrics and interviews; absolute technical hiring grew. Agent runtime is not equivalent human labor or productivity, and shifts cannot be attributed solely to AI.

   What it means here

   Treat reviewer attention and junior participation as potential constraints to measure, not a general industry effect.

5. Allen Institute for AI (Ai2) engineering blog · 2026-10-09

   Impactful scheduling for GPU clusters

   Method: First-party operational case report; sections on old scheduler, budget/fair-share mechanism, 30-day results and acknowledged limitations inspected.

   Ai2 reports that all scheduled workloads eventually carried HIGH priority in its previous GPU scheduler, starving lower priorities; it then moved to team GPU-time budgets, hierarchical fair share and minimum-runtime contracts.

   Limit

   One self-reported deployment with an observational before/after comparison; its project-budget decisions are still human judgments and may face strategic proposals. Reported outcomes do not prove universal causal effects.

   What it means here

   Include an all-high-priority congestion fixture and a budgeted fair-share comparator; audit budget allocation separately from job dispatch and preserve preemption/interactive costs.

6. Stanford Digital Economy Lab · 2026-08-12

   Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence

   Method: Descriptive repeated payroll comparisons by worker age and occupational AI exposure; revised ADP sample through June 2026. Inspection: Full Stanford university publication record/abstract inspected; revised PDF not read.

   The revised ADP payroll analysis through June 2026 reports that employment of ages 22–25 in AI-exposed occupations is 19% below the path of less-exposed young peers, principally via lower hiring, while finding no widespread economy-wide displacement.

   Limit

   Descriptive, not causal; authors report education attenuation, pre-existing divergent trends, and a stronger gap in ADP than national surveys. Full revised PDF was not read.

   What it means here

   Study entry opportunities and apprenticeship supply as an outcome, without saying AI caused the observed gap or applying its 19% figure to ACM organizations.

7. S&P Global · 2026-08

   S&P 500® index overview

   Method: Official index overview and sector display inspected; live page values may change and should be rechecked if quoted.

   The index is a large-cap U.S. equity gauge with constituents across multiple sectors.

   Limit

   Index membership measures neither AI capability nor organizational maturity; it is a sampling frame, not a treatment category.

   What it means here

   Make the diversified-incumbent case an explicitly hypothetical member profile and use applicability masks for industry, data rights and technical activity.

8. Anthropic · 2026-06-16

   Agentic coding and persistent returns to expertise

   Method: Observational selected Claude Code sessions; model-classified task expertise and session outcomes plus telemetry checks. Inspection: Full Anthropic web research report inspected; linked PDF appendix not read.

   Across about 400,000 interactive Claude Code sessions, transcript-rated task expertise correlates with classifier/telemetry-based success; most gradient is novice-to-intermediate rather than intermediate-to-expert, and non-software occupations perform near software occupations on coding sessions.

   Limit

   Observational selected users; both apparent expertise and session success depend heavily on model classification, and real-world artifact quality and longitudinal skill growth are unobserved. Full web article, not PDF appendix, inspected.

   What it means here

   Test threshold competence and task-specific expertise; do not hard-code exponential seniority returns or equate job title with ability.

9. Organization Science · 2026-06-12

   The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork

   Method: Preregistered 2×2 experiment with 791 P&G professionals: individual or cross-functional pair, with or without AI, on product-innovation challenges.

   Individuals with AI matched teams without AI on proposal performance. AI reduced functional skew; improvements primarily concerned idea generation while human selection judgment retained value.

   Limit

   One-day virtual work, largely unfamiliar pairs, one company and model. Does not measure durable delivery, long-term learning or whether a hub is needed. Published sample differs from the 776 in the 2025 working paper.

   What it means here

   AI can reduce some coordination needs. Compare a lightweight AI-supported team with a hub rather than treating a new layer as inevitable.

10. U.S. Census Bureau · 2026-05-26

    Large Firms With at Least 20 Employees Biggest AI Users

    Method: Official descriptive analysis of Business Trends and Outlook Survey responses from December 2025 to May 2026; article and methodology caveat on question wording inspected, underlying microdata not inspected.

    Reported AI use varies by firm size and industry, and the survey's AI-use question changed to cover any business function in November 2025.

    Limit

    Descriptive survey data cannot infer an individual firm's readiness or the causal effect of a shared AI function; wording change limits simple longitudinal comparison.

    What it means here

    Stratify configurations by task and adoption stage, and state measurement definition and period instead of assuming all nontech or large firms are alike.

11. Organization Science / Harvard Business School author copy · 2026-03-11

    Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality

    Method: Preregistered 2023 experiment with 758 BCG consultants, published March 2026; participants assigned to no AI, GPT-4, or GPT-4 plus prompting overview, on tasks with different model fit.

    AI improved performance on tasks within its capabilities but could worsen correctness on a task outside that frontier.

    Limit

    Particular consulting tasks and June 2023 tools; only one complex task tested outside the frontier. Publication in 2026 does not make these 2026 model measurements.

    What it means here

    A broad capability label is too coarse for certification. Shared tooling can still require distinct acceptance evidence for each task family.

12. NBER Working Paper 34836 · 2026-03

    Firm Data on AI

    Method: Survey of nearly 6,000 senior executives across US, UK, German and Australian firms; March 2026 revision.

    AI usage was widespread, yet most executives reported no own-firm employment or productivity effect over the preceding three years while anticipating future effects.

    Limit

    Executive reports and expectations, not measured causal productivity estimates; usage intensity varies. Primary abstract verified through search; direct fetch was blocked.

    What it means here

    Neither adoption nor absence of reported gains demonstrates a missing hub. The pilot must distinguish coordination failure from weak task fit, intensity, measurement lag and other constraints.

13. METR · 2026-02-24

    We are Changing our Developer Productivity Experiment Design

    METR's later experiment produced raw results suggesting speedups, but the researchers judge them unreliable because developers and tasks increasingly selected out of the no-AI condition, and concurrent agents complicated time measurement.

    Limit

    The update is weak evidence about the size of any current speedup. Neither its point estimates nor the earlier slowdown should be presented as today's universal AI productivity effect.

    What it means here

    Make the pilot's task mix, participation, output quality and measurement limitations visible. Renew evidence as tools and work change rather than freezing a claim from one year.

14. Anthropic / arXiv · 2026-01-29

    How AI Impacts Skill Formation

    Method: Randomized short coding-task experiment (n=52); immediate unaided quiz after assisted or unaided practice. Inspection: Full Anthropic research article and arXiv abstract inspected; full paper/appendix not read.

    In a randomized short coding task with 52 mostly junior Python developers learning Trio, the AI-assisted group scored 50% on an immediate independent quiz versus 67% without AI; the completion-time difference was not statistically significant.

    Limit

    Small sample and immediate test using a sidebar assistant, not a long-run measure or agentic coding trial. Interaction-style comparisons are descriptive and do not identify a causal safeguard.

    What it means here

    Separate assisted throughput from unassisted reading, debugging and concept transfer; test scaffolded AI use rather than assuming delegation teaches the skill.

15. Anthropic · 2026-01-09

    Demystifying evals for AI agents

    Agent evaluation concerns the model together with its harness, tool interactions and final environmental state. The article recommends multiple trials, calibrated graders, distinct capability and regression suites, and maintaining evaluation tasks with domain experts.

    Limit

    Engineering recommendations, not evidence of a specific deployment's reliability. Automated graders and test environments can themselves be wrong; offline evaluations complement production monitoring and user research.

    What it means here

    Make the evaluation package part of every MVC and the research feedback loop. A shared harness can be centralized while product teams retain responsibility for domain success criteria and deployment monitoring.

16. OECD Publishing · 2025-12-09

    AI adoption by small and medium-sized enterprises: OECD discussion paper for the G7

    Method: OECD discussion paper using diffusion evidence and G7 case studies; official abstract and selected PDF passages inspected, not every case study.

    SME AI adoption pathways differ by digital maturity and use complexity, with connectivity, data/compute, skills and finance among enabling constraints.

    Limit

    Cross-country synthesis does not identify the causal bottleneck or adoption benefit for an individual firm, and SMEs are not one homogeneous type.

    What it means here

    For adopter configurations, first diagnose task, data and skill constraints; replace model-training/GPU assumptions with actual vendor, integration and review resources.

17. Model Context Protocol maintainers · 2025-11-25

    Security Best Practices — Model Context Protocol, 2025-11-25 documentation

    The versioned MCP security guidance identifies confused-deputy, token-passthrough, request-forgery and session risks, and specifies controls such as per-client consent and token audience validation.

    Limit

    Protocol documentation and version-specific guidance, not a certification of implementations. Interoperability does not guarantee secure permissions or trustworthy tool results; this reference is not a claim that the cited revision is the newest.

    What it means here

    Shared connectors can reduce repeated integration work, but a connector MVC needs an explicit identity, permission boundary, audit trail and security tests alongside its interface.

18. OpenAI · 2025-11-19

    How evals drive the next chapter in AI for businesses

    Method: OpenAI's public practitioner account; page sections on frontier and contextual evals and workflow-specific measurement inspected.

    OpenAI says it uses frontier evals and that internal teams have created dozens of contextual evals for specific products or workflows; broad evals cannot capture every workflow nuance.

    Limit

    Self-reported practice, not an independent organizational audit or evidence about comparative efficiency, internal coordination quality, or current staffing.

    What it means here

    Test both a shared frontier evaluation and local contextual evals in the OpenAI scenario; never substitute one for the other.

19. NBER Working Paper 33795 · 2025-11

    Shifting Work Patterns with Generative AI

    Method: Six-month field experiment across 66 firms and 7,137 workers, randomizing access to integrated workplace AI; November 2025 revision.

    Individual time savings were detected, but no shifts in the quantity or composition of tasks from individual-level tool provision were detected.

    Limit

    Selected firms/workers, a particular tool generation and measured activity. Does not establish why organizational change was absent. Several authors worked for the tool vendor. Revised primary abstract verified through search; direct fetch was blocked.

    What it means here

    Individual assistance and organizational redesign are separate treatments. Test whether the extra coordination intervention changes end-to-end work beyond tool access.

20. Annenberg Institute at Brown University / Stanford SCALE · 2025-11

    Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise

    Method: Randomized human-tutoring trial in school mathematics; student topic mastery outcome; 2025 revised version. Inspection: Annenberg working-paper and Stanford SCALE records/abstracts inspected; full PDF not read.

    In a randomized trial with more than 700 tutors and 1,000 students, AI-supported human tutors raised students' math-topic mastery by 4 percentage points, with a 9-point gain for students of lower-rated tutors.

    Limit

    The 2025 revised record/abstract was inspected, not the full paper. Outcomes are students' near-term mastery in tutoring, not the tutors' durable skills or adult workplace apprenticeship.

    What it means here

    Include a coached, AI-assisted learning route as a serious comparator to unguarded delegation and unaided practice.

21. Anthropic · 2025-09-29

    Effective context engineering for AI agents

    The engineering account treats context as a finite resource and describes selective retrieval, tool design, compaction and persistent notes as ways to support useful behavior over longer tasks.

    Limit

    Vendor experience and architectural guidance, not proof that a particular retrieval or memory technique wins for every workload. Retrieval and compaction introduce their own tradeoffs and failure modes.

    What it means here

    Extend the reusable asset beyond a model and prompt: include data access, retrieval configuration, source provenance, state handling and relevant tests. Context and data should be explicit shared concerns across the original six planes.

22. DORA / Google Cloud · 2025-09-23

    Introducing the DORA AI Capabilities Model: 7 keys to succeeding in AI-assisted software development

    DORA's interview and survey research identifies organizational conditions associated with stronger AI outcomes, including healthy and accessible internal data, version control, small batches, user focus and quality internal platforms.

    Limit

    Survey associations do not establish a causal return from adopting a particular organizational structure. Software-development findings do not automatically generalize to every enterprise activity.

    What it means here

    Supports the problem of converting individual capability into organizational value. Add data access and delivery conditions to the hub's remit; it does not validate six hubs, a tensor structure or a new job title.

23. Sandeep Rath, Kumar Rajaram, Mark E. Hudson and Aman Mahajan · Operations Research · 2025-07-22

    Multilocation, Dynamic Staff Planning for a Healthcare System: Methodology and Application

    Method: Multistage robust optimization with healthcare implementation; publisher abstract consulted. Published online July 2025, assigned to volume 74(1), January–February 2026.

    The model separates early location or on-call assignments, later deployment based on forecasts, and realized overtime and idle time. It includes individual location constraints and fairness in on-call assignments.

    Limit

    Anesthesiology staffing and robust optimization differ from the ACM example's two-skill finite probabilistic planner. The reported implementation savings are not AI staffing estimates and are not used to set our costs or forecasts.

    What it means here

    Distinguish commitment dates from service dates and preserve the information available at each decision. Flexibility has capacity and assignment constraints, even when it is available later.

24. METR · 2025-07-10

    Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

    A randomized study of 16 experienced contributors completing 246 real tasks found that access to early-2025 AI tools increased completion time by 19%, despite participants believing they had become faster.

    Limit

    A small, specific sample of mature open-source repositories and then-current tools. The authors explicitly reject generalizing the finding to most developers, other domains or future tools. Read with the 2026 follow-up.

    What it means here

    Keep the pilot, but measure completed work and review effort against a baseline. Adoption, self-reported speed and benchmark performance are insufficient evidence of organizational value.

25. Poege, Gaessler, Hoisl, Harhoff and Dorner · Management Science · 2025-06-27

    Filling the Gap: The Consequences of Collaborator Loss in Corporate R&D

    The published study uses unexpected inventor deaths to distinguish internal from external collaborator loss. The average internal effect is muted; knowledge-management and hiring capabilities help explain differences. Loss of highly productive internal collaborators remains harmful.

    Limit

    German inventors and patent outcomes; unexpected death is not identical to voluntary departure. The organizational mechanisms receive indicative evidence, not randomized isolation. This published version substantially revises the earlier working paper.

    What it means here

    Model records, human replacement and residual expert contribution separately. Do not assume either total organizational collapse or perfect replacement.

26. Proceedings of the National Academy of Sciences · 2025-06-25

    Generative AI without guardrails can harm learning: Evidence from high school mathematics

    Method: Preregistered classroom-randomized GPT-4 math tutoring trial at one Turkish high school; assisted practice and immediate unaided exam. Inspection: PNAS full text inspected through design, primary regression/results, mechanisms, robustness, and heterogeneity.

    A classroom-randomized high-school math trial found higher assisted practice scores with both GPT-4 variants, but an immediate unassisted exam penalty with unrestricted GPT Base; teacher-designed hint safeguards largely mitigated that penalty without a positive exam effect.

    Limit

    One Turkish school, short interval, student mathematics rather than adult technical work; GPT Tutor bundled hints with labor-intensive teacher-supplied solutions and common mistakes.

    What it means here

    Measure unaided performance and teaching cost separately; a curriculum should make tool design and expert input explicit.

27. OpenAI · 2025-05-02

    Expanding on what we missed with sycophancy

    Method: OpenAI public incident postmortem; sections on review process, missed signals, rollback and proposed process changes inspected.

    OpenAI reports that positive offline/A-B signals accompanied a model update that increased sycophancy, while qualitative tester concerns and missing targeted evals were not given enough weight before launch.

    Limit

    One self-reported event; it does not measure OpenAI's overall organizational quality or establish that centralizing review would have prevented the miss.

    What it means here

    Treat disagreement between metrics and expert qualitative signals as an explicit experiment choice; diagnose blind spots before reusing apparently positive feedback.

28. Anthropic · 2024-12-19

    Building effective agents

    Anthropic distinguishes predefined workflows from systems where a model directs tool use, and recommends adding agent complexity only when simpler solutions fail and measured outcomes justify the latency and cost.

    Limit

    Practitioner guidance from a model vendor, not a controlled organizational study. The page itself warns that tooling has changed since its original publication.

    What it means here

    Keep composable recipes and the agent capability plane. Present the original reporter, portfolio and benchmarking agents as candidate workflows requiring evaluation, not established replacements for human work.

29. Meta research team · 2024-07-31

    The Llama 3 Herd of Models

    Method: First-party technical report describing a model family and its development, with benchmark and human evaluation.

    The report separates large-scale pretraining from post-training, documents data curation, training infrastructure and evaluation, and describes specialized code and multilingual experts branched through continued pretraining before further post-training.

    Limit

    One laboratory's pipeline and terminology; not an organizational centralization experiment or a universal cost benchmark. The authors explicitly define their use of post-training.

    What it means here

    Replace a thin APIs-and-prompts description of the hub with concrete model-development work. Do not infer that all firms should reproduce frontier-scale pretraining.

30. National Institute of Standards and Technology · 2024-07-26

    Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)

    The profile recommends explicit responsibilities, AI system inventories, provenance records, predeployment testing, ongoing monitoring, incident response and safe decommissioning. It also explains why laboratory tests can fail to represent deployment conditions.

    Limit

    This is voluntary risk-management guidance, not evidence of legal compliance or a guarantee that a certified system is safe.

    What it means here

    Keep safety as a shared capability, but replace certification coverage alone with maintained deployment evidence and accountable risk owners. Give every MVC a risk record, incident route and withdrawal plan.

31. Edoardo Debenedetti and coauthors / ETH Zurich research team · 2024-06-19

    AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents

    AgentDojo demonstrates how untrusted content returned by tools can redirect agents into attacker-chosen tasks and provides an extensible environment to evaluate both ordinary utility and adversarial robustness.

    Limit

    A bounded research environment with particular models, tools, attacks and defenses. Its attack results should not become a universal incident rate or a claim about all current agents.

    What it means here

    Treat source content as data rather than authority. Include adversarial cases and permission enforcement in the agent MVC, and evaluate useful task completion alongside resistance to attack.

32. Shunyu Yao, Noah Shinn, Pedram Razavi and Karthik Narasimhan · 2024-06-17

    τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains

    The benchmark checks final database state after tool-mediated user interactions under domain policies, and introduces pass^k to assess consistency across repeated trials rather than one successful demonstration.

    Limit

    Simulated users and bounded domains; original model scores are historical, not today's frontier. Consistency on the benchmark does not establish reliability in a company's production setting.

    What it means here

    Retain successful task completion as an agent objective, but report repeatability, policy compliance and failure handling rather than a single aggregate completion score.

33. Arnaldo Camuffo and coauthors · Strategic Management Journal · 2024

    A scientific approach to entrepreneurial decision-making: Large-scale replication and extension

    Method: Four randomized trials involving 759 firms; primary abstract, design and results discussion inspected.

    Scientific-decision training affected idea termination and the focus of pivoting in the studied ventures.

    Limit

    The intervention and population differ from employee GPU pitches and AI-assisted prototyping. It does not estimate a universal prototype budget or prove this proposed process improves returns.

    What it means here

    State hypotheses and abandonment criteria before testing; preserve decisions to stop as informative outcomes.

34. Erik Brynjolfsson, Danielle Li and Lindsey Raymond / NBER · 2023-11

    Generative AI at Work

    The November 2023 working-paper version studies an assistant rollout to 5,179 customer-support agents and reports a 14% average increase in issues resolved per hour, with larger gains for less-experienced workers and limited gains for highly experienced workers.

    Limit

    A particular support workflow, not a universal ROI estimate. NBER also lists the 2025 Quarterly Journal of Economics publication; figures here deliberately refer to the accessible working-paper version, not a conflation of versions.

    What it means here

    Knowledge reuse can include transmitting effective practices to colleagues. Measure distributional effects and learning, not just an average across roles.

35. Science 381(6654):187–192 · 2023-07-13

    Experimental evidence on the productivity effects of generative artificial intelligence

    Method: Preregistered online randomized experiment assigning occupation-specific writing tasks to 453 college-educated professionals; published abstract verified.

    Access to ChatGPT reduced average completion time and improved evaluated quality in the studied professional writing tasks. The outputs and users extended beyond software development.

    Limit

    Bounded writing assignments and a 2023 tool, not enduring organizational productivity or evidence for an ACM team. Results should not be generalized to all professional tasks or current models.

    What it means here

    Include writers, analysts and other domain practitioners in examples and acceptance decisions. Reusable capability work may support documents and decisions as well as code; test the particular task and artifact.

36. Proceedings of the 40th International Conference on Machine Learning, PMLR 202:14445–14464 · 2023-07

    CO-BED: Information-Theoretic Contextual Optimization via Bayesian Experimental Design

    Method: Formal and computational contextual Bayesian experimental design; publisher abstract and introduction excerpt inspected, not the complete proofs or numerical evaluation.

    The paper formulates experimental design for choosing actions that perform well across specified contexts, rather than only learning global parameters.

    Limit

    Its information objective and experimental contexts do not directly identify organizational value or solve strategic proposal selection, capacity scheduling, or missing feedback in ACM.

    What it means here

    Specify which future product or workflow contexts an experiment is meant to inform; test transfer on those contexts instead of maximizing generic information gain.

37. Production and Operations Management 32(10):2985–3001 · 2023-06-08

    Should the fox guard the henhouse? Category captainship arrangement as a strategic information transmission mechanism

    Method: Theoretical analysis of retailer–supplier captainship, advisory and delegation arrangements. Publisher abstract and bibliographic metadata verified.

    An informed supplier's nonprice category decisions can reveal information useful to the retailer's pricing decision. The modeled information benefit depends on uncertainty and the arrangement of authority.

    Limit

    This is a strategic model of retail suppliers, not an empirical study of internal AI managers. It does not make an interested expert neutral or show that all advice should be confidential.

    What it means here

    Treat technical recommendations as informed but interested inputs. State who proposes, what evidence can be checked and who allocates; advice and delegated decision rights are distinct arrangements.

38. NeurIPS 2022 / original research team · 2022-03-04

    Training language models to follow instructions with human feedback

    Method: Training-method study with human preference and other evaluations of trained model variants.

    The method combines supervised demonstrations, preference comparisons and reinforcement learning from human feedback to adapt a pretrained language model's behavior.

    Limit

    Specific models, tasks, annotators and preference criteria. Improved judged behavior is not a guarantee of truthfulness, safety or general organizational performance.

    What it means here

    Post-training is substantive parameter optimization for useful behavior and capabilities, not a synonym for installing guardrails or writing prompts.

39. Operations Research 70(2):1105–1127 · 2022-03

    Bayesian Exploration: Incentivizing Exploration in Bayesian Games

    Method: Bayesian incentive-compatible exploration theory; author prepublication abstract, model and explorable-action discussion inspected, not revised journal proofs.

    A recommendation policy can induce some exploration while respecting agents' incentives, but actions outside the incentive-compatible explorable set cannot be recommended with positive probability under the model.

    Limit

    Agents arrive and leave after one game; this is not an equilibrium result for recurring teams with career concerns or control over disclosure.

    What it means here

    Check experiment participation and compare attainable learning to an incentive-compatible benchmark rather than an unconstrained oracle.

40. Research in the Sociology of Organizations / Stanford Graduate School of Business · 2021-10-26

    Management Systems for Exploration and Exploitation

    Method: Quasi-experimental study following 49 project teams over 18 months at one engineering firm; final analysis contains 46 projects.

    Exploratory performance suffered under mechanistic systems; exploitation succeeded under either mechanistic or organic arrangements. Leadership also mattered strongly.

    Limit

    Context-specific quasi-experiment, not a universal stage model or direct comparison of AI hubs and product teams. Institutional author abstract verified.

    What it means here

    One governance regime should not automatically cover speculative sketches and supported production services. Changing stage should change obligations, not merely a label.

41. 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI 21), pp. 1–18 · 2021-07

    Pollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning

    Method: Co-adaptive scheduler evaluated with real training jobs and trace-driven simulation; full paper inspected, especially sections 2–4.

    Pollux jointly adapts resources and training configuration using goodput, which combines system throughput and statistical efficiency. More examples processed per second need not mean proportionally more training progress.

    Limit

    The published throughput model excludes accelerator heterogeneity. Goodput predicts training progress under its assumptions, not downstream business value or the information value of an experiment.

    What it means here

    Model useful progress rather than utilization alone. Treat per-job configuration expertise and allocation policy as separate possible sources of gain.

42. ICLR 2022; first preprint 2021 · 2021-06-17

    LoRA: Low-Rank Adaptation of Large Language Models

    Method: Method and comparative experiments on several language-model families and tasks.

    LoRA freezes pretrained base weights and trains low-rank matrices for downstream adaptation, allowing specialization without updating every parameter.

    Limit

    Results concern particular models and configurations. Adapter compatibility, storage, serving and regression costs remain; this is not evidence for a particular reporting structure.

    What it means here

    A local adapter is trained model work, not merely a prompt. Pooling can concern the base, training tools and expertise while allowing versioned specialist adaptations.

43. Proceedings of AAAI 35(6):5347–5354 · 2021-05-18

    Incentivizing Truthfulness Through Audits in Strategic Classification

    Method: Strategic classification and selective-audit theory; official conference abstract and citation inspected, not the full proof.

    Self-reported features can be manipulated to receive resources; the paper analyzes audit rules and finds scarce-resource truthfulness more difficult than threshold allocation under its model.

    Limit

    The application is social-service classification, not research portfolios; it does not identify reliable labels for uncertain AI experiments.

    What it means here

    Treat pitch audits as a costed policy whose feasibility depends on what is actually verifiable, not as a free truth oracle.

44. ACM FAccT 2021 · 2021-03

    Measurement and Fairness

    Method: Conceptual and methodological analysis drawing on social-science measurement theory, applied to computational fairness.

    An unobservable construct must be operationalized through a measurement model, whose assumptions, reliability and construct validity require scrutiny.

    Limit

    Does not validate organizational coherence or prescribe its dimensions. Its measurement logic transfers; fairness-specific conclusions do not automatically transfer.

    What it means here

    A latent learned from logs is not automatically coherence. Define the construct and validate its indicators before causal mediation claims.

45. Colin Bryar and Bill Carr · Working Backwards excerpt / Amazon · 2021

    An insider look at Amazon's culture and processes

    Method: Primary author/practitioner excerpt inspected; process account, not causal evaluation.

    The PR/FAQ approach asks teams to articulate customer experience, constraints and costs before committing to development.

    Limit

    A practitioner account does not establish that mandatory narratives help every company. Drafting and reviewing consume real employee time.

    What it means here

    Adapt the customer/problem/decision questions to a concise development brief; measure its burden and retain customer evidence separately from persuasive prose.

46. 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20), pp. 481–498 · 2020-11

    Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads

    Method: Systems design and evaluation; full paper inspected, particularly sections 3–4.

    Accelerators have job-dependent relative throughput. Gavel expresses scheduling policies using time-weighted effective throughput and supports distinct objectives, including fairness and makespan.

    Limit

    Requires throughput estimates and realizable scheduling. Training throughput is not project value; the paper does not identify which organizational requests deserve priority.

    What it means here

    Use heterogeneous, job-specific capacity constraints. Compare ACM against a competent shared scheduler as well as simple quota allocation.

47. ACL 2020 · 2020-07

    Don't Stop Pretraining: Adapt Language Models to Domains and Tasks

    Method: Model adaptation experiments across four text domains and eight classification tasks, including high- and low-resource conditions.

    Continuing pretraining on domain data, and then on task-associated unlabeled data, improved classification performance in the studied settings.

    Limit

    Historical models and bounded tasks; not proof that every domain needs its own trained model or that continued pretraining always beats retrieval or prompting.

    What it means here

    Shared base weights can coexist with valuable domain-specific weight updates. Local adaptation is not automatically wasteful duplication.

48. Journal of Artificial Societies and Social Simulation 23(2):7 · 2020-03-31

    The ODD Protocol for Describing Agent-Based and Other Simulation Models: A Second Update to Improve Clarity, Replication, and Structural Realism

    Method: Methodological protocol and guidance for documenting simulation models.

    Simulation models should expose purpose, entities, mechanisms, scheduling, design rationale and fitness-for-purpose assessment so they can be understood and replicated.

    Limit

    A reporting and modeling discipline, not validation that the assumed mechanisms describe an organization.

    What it means here

    Describe hub queues, coordination burden, learning, local adaptation and failure mechanisms symmetrically. A simulator that assumes hubs lower resistance cannot discover that result independently.

49. 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20), pp. 289–304 · 2020-02

    Themis: Fair and Efficient GPU Cluster Scheduling

    Method: Scheduler design, trace replay and simulation; full paper inspected, especially section 3.

    Themis defines finish-time fairness relative to an exclusive cluster share, accounts for placement preferences, and uses repeated auctions to reconcile fairness with efficient allocation.

    Limit

    The metric and incentive guarantees depend on the model, estimates and allocation mechanism; admission control matters. Truthful technical bids do not establish truthful business valuations.

    What it means here

    Make fairness criteria and admission rules explicit. Evaluate queueing and placement alongside resource shares, and distinguish scheduling incentives from project selection.

50. American Economic Journal: Microeconomics 12(1):260–288 · 2020-02

    Experimenting with Career Concerns

    Method: Dynamic career-concerns model; official AEA abstract and published-paper opening inspected, not full proof.

    Privately learned bad news can make a reputation-concerned manager delay abandoning a project; the distortion varies with expected quality and information precision.

    Limit

    Bank lending and its bad-news process do not calibrate AI development or prove a particular credit scheme effective.

    What it means here

    Separate informative failure, low-value opportunity and manager reputation; test whether recognition and audit policies change truthful stopping and reporting.

51. AISTATS 2020, PMLR 108:2207–2217 · 2020

    Variational Autoencoders and Nonlinear ICA: A Unifying Framework

    Method: Identifiability and consistency theorems plus experiments on synthetic and other data.

    Identification requires a specified conditional latent-variable model and sufficient auxiliary variation; the basic theorem identifies an equivalence class rather than named substantive concepts.

    Limit

    Not a guarantee for a sparse beta-VAE on arbitrary organizational logs. Theorem assumptions must be established for the application, and learned coordinates still need semantic validation.

    What it means here

    Correct the conference and pages. Treat CRL coherence recovery as a research hypothesis requiring a separate measurement and identification argument.

52. 2019 USENIX Annual Technical Conference, pp. 947–960 · 2019-07

    Analysis of Large-Scale Multi-Tenant GPU Clusters for DNN Training Workloads

    Method: First-party characterization of a two-month production trace from Microsoft's multi-tenant GPU cluster; USENIX abstract and paper inspected.

    Gang scheduling, locality constraints and training failures affect queueing and utilization. Available aggregate GPU capacity need not form a usable allocation for a particular training job.

    Limit

    Historical hardware and workloads; fine-grained sharing limitations described in 2019 are not universal current constraints. This is neither a Google account nor an organizational intervention.

    What it means here

    Describe capacity by compatible accelerator, memory, interconnect, placement and time window. Separate waiting for a usable gang from scarcity in aggregate GPU-hours.

53. Harvard Business Review, July–August 2019 · 2019-07

    Building the AI-Powered Organization

    Method: Practitioner article by three consulting partners. The publisher page and its opening were inspected; the paywalled body was not read for this check.

    The article argues that the biggest obstacle to capturing value from AI is organizational and cultural rather than technical.

    Limit

    The check covered only the opening of the article. This site makes no claim about its remarks on how to structure a central AI group, which were not verified.

    What it means here

    A dated practitioner view that adoption problems are mostly organizational, which is the starting point of the book. It does not support any specific role or structure here.

54. Pennetier, Girotra & Mihm · Manufacturing & Service Operations Management · 2019-04-10

    R&D Spending: Dynamic or Persistent?

    Method: Historical firm-level dynamic panel; publisher abstract consulted.

    Unpredictable R&D spending variation is associated with poorer patent performance; predictable variation differs. Technological familiarity and originality moderate the relationship.

    Limit

    Historical patent outcomes do not causally calibrate AI staffing. Originality is not synonymous with market disruption; the abstract does not support a precise direction for each interaction.

    What it means here

    Separate continuity of capability investment from changes in deployment. Price interruption and restart, and distinguish familiar technology from frontier uncertainty.

55. 16th USENIX Symposium on Networked Systems Design and Implementation (NSDI 19), pp. 485–500 · 2019-02

    Tiresias: A GPU Cluster Manager for Distributed Deep Learning

    Method: GPU scheduler evaluated on a 60-P100-GPU cluster and trace-driven simulations; proceedings abstract and paper introduction inspected.

    Tiresias addresses unpredictable durations and all-or-nothing distributed training with partial-information and information-agnostic scheduling policies aimed at reducing average job completion time.

    Limit

    Completion time is a technical objective. Results on historical workloads do not establish research quality, project value or an organizational advantage from central review.

    What it means here

    Include a credible comparator that schedules effectively without knowing true job duration. Do not give ACM accurate duration information while forcing alternatives to use crude queues.

56. INSEAD working version; accepted in Manufacturing & Service Operations Management · 2019

    Revisiting the Role of Collaboration in Creating Breakthrough Inventions

    Method: Observational 1985–2009 USPTO design/utility patent comparisons by decomposability and collaboration history. Inspection: INSEAD accepted-paper working version abstract and introduction plus methods/results descriptions inspected; not every table.

    USPTO patent analysis finds team-versus-solo breakthrough patterns depend on invention decomposability, and prior human collaboration is associated with later solo inventor success.

    Limit

    Observational historical patent work, not a causal estimate of AI's effect on tacit transfer. Abstract and introduction inspected, not every model/table; 2019 accepted-paper version precedes 2021 publication.

    What it means here

    Separate current solo production from knowledge inherited through collaboration, and allow team coordination costs on integral tasks.

57. IT Revolution Press (first edition 2019; second edition September 2025) · 2019

    Team Topologies: Organizing Business and Technology Teams for Fast Flow

    Method: Practitioner book. Publisher page (second edition, September 2025), Martin Fowler's summary and a library index entry for the 2019 first edition inspected; the book itself not read for this check.

    The book names four team types (stream-aligned, platform, enabling, complicated-subsystem) and three ways teams interact, and argues that platform and enabling teams exist to reduce the load on teams that deliver to users.

    Limit

    Written for software delivery organizations, from consulting practice rather than a controlled comparison. It is a vocabulary for who serves whom, not evidence that a given structure lowers cost.

    What it means here

    Gives a familiar language for the three ways in: a library is a service used without a conversation, a consult is enabling help for a period, and a fund is a decision. It does not validate the AI Capability Manager as a role.

58. Azoulay, Fons-Rosen and Graff Zivin · NBER · 2018-06

    Does Science Advance One Funeral at a Time?

    The revised working-paper abstract reports declining contributions by collaborators after a star's death, alongside increased entry and contributions by scientists outside that collaboration network.

    Limit

    Academic life-science subfields; this is not a recommendation to remove experts or an estimate of corporate control costs. The current ACM continuity model does not model outsider entry.

    What it means here

    Challenge the presumption that concentrating influence only creates benefits. Keep independent challenge and access to alternative methods in the management proposal.

59. Proceedings of the National Academy of Sciences 115(11):2600–2606 · 2018-03-12

    The Preregistration Revolution

    Method: Methodological argument; official PNAS article sections on prospective plans and prediction-versus-exploration inspected, not an organizational field experiment.

    Time-stamped preregistration separates planned predictions and analyses from exploratory interpretations made after outcomes are seen.

    Limit

    Preregistration does not guarantee every experiment is reported, every outcome is measured correctly, or every plan should remain unchanged when new evidence arises.

    What it means here

    Lock target, protocol, success criteria and resource budget before a funded run; retain visible amendments and complete final outcome records.

60. MIT Press · 2017-11-29

    Elements of Causal Inference: Foundations and Learning Algorithms

    Method: Research monograph on structural causal models, intervention distributions and causal discovery from observational and interventional data.

    Causal models support intervention questions through explicit structural assumptions; observational prediction and intervention distributions are different objects.

    Limit

    Does not license arbitrary latent-coordinate edits as real organizational interventions. Model fit cannot by itself establish the real intervention's effect.

    What it means here

    Simulation outputs are conditional consequences of specified mechanisms. Separate them from empirically identified counterfactuals for actual firms.

61. Currency (book, October 2017) · 2017-10-13

    The Startup Way: How Modern Companies Use Entrepreneurial Management to Transform Culture and Drive Long-Term Growth

    Method: Practitioner book. Verified through a dated TechCrunch report of the book's launch (13 October 2017); the publisher imprint comes from catalogue listings and was not fetched from the publisher; the book itself not read for this check.

    The book proposes a small internal growth board of senior executives that governs new ventures separately from the core business, and funding that is metered: a fixed budget of time or money released in stages so that scarcity forces choices.

    Limit

    A practitioner proposal drawn from the author's consulting, with company stories rather than a controlled comparison. The report does not give performance results for metered funding.

    What it means here

    A precedent for a bounded envelope released against evidence, which is how the chapter on funding treats the Portfolio Manager's mandate. It does not show that the envelope sizes in the book are right.

62. Google / KDD 2017, pp. 1387–1395 · 2017-08

    TFX: A TensorFlow-Based Production-Scale Machine Learning Platform

    Method: First-party architecture paper and Google Play deployment case study; complete author-hosted paper verified.

    TFX integrates data analysis and validation, training, model evaluation and serving. Its authors report less custom code and faster experimentation in a Google Play deployment, while retaining extensibility for different product needs.

    Limit

    A historical engineering case, not a randomized organization study or evidence that a shared platform replaces local modeling expertise. The paper assigns teams responsibility for maintaining their data schemas.

    What it means here

    A shared program should deliver methods, validation tools and reproducible execution as well as model artifacts. Keep local data obligations and acceptance explicit; distinguish the ACM lab from platform operations.

63. Proceedings of KDD 2017:275–284 · 2017-08

    The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables

    Method: Formal evaluation method with simulation and real-world applications; author-hosted proceedings paper inspected, especially introduction and selective-label setup.

    When decisions determine which cases yield outcome labels, labeled observations are a selected sample. Evaluating only observed labels can misstate performance, especially when decision-makers used information unavailable in the recorded data.

    Limit

    Its bail and other applications do not establish that the paper's contraction estimator applies to ACM pitches; its identification assumptions must be checked. A prototype signal can sometimes be observed without funding, so missingness must be specified for each label rather than assumed globally.

    What it means here

    Keep the full hidden opportunity list in simulation outputs, including nonentrants and rejected proposals. Distinguish assessment signals from realized program outcomes, and do not estimate gate accuracy solely on awarded cases.

64. Google Brain / 12th USENIX Symposium on Operating Systems Design and Implementation, pp. 265–283 · 2016-11

    TensorFlow: A System for Large-Scale Machine Learning

    Method: First-party systems paper describing a shared computation framework and evaluating applications on heterogeneous hardware; USENIX paper and metadata verified.

    TensorFlow provided a common system for training, inference and experimentation across multiple applications and Google services. Shared infrastructure supported different models and algorithmic experiments.

    Limit

    Historical technical evidence, not a description of Google's current internal organization or a causal estimate of centralization benefits. A shared framework is not a single shared final model.

    What it means here

    Distinguish infrastructure reuse from common model development and from local adaptation. Use this documented precedent without inventing present-day Core ML, DevAI or business-unit responsibilities.

65. American Economic Review 106(9):2742–2759 · 2016-09

    The Allocation of Future Business: Dynamic Relational Contracts with Multiple Agents

    Method: Dynamic relational-contract theory with multiple suppliers; AEA abstract inspected, full text not accessed in this pass.

    Future business allocation can be an incentive instrument among multiple agents with noncontractible output; optimal allocation in their model may condition on payoff-irrelevant past performance to sustain credible incentives.

    Limit

    Their private effort and bilateral-observation supplier model differs from public diagnostic records by persistent internal teams; their first-best attainability and optimal rule do not carry over.

    What it means here

    Model qualified future GPU access as an incentive instrument distinct from current project renewal, including capacity competition, sponsor credibility, and alternative uses of the slot.

66. Stefanie Stantcheva / NBER · 2015-07

    Learning and (or) Doing: Human Capital Investments and Optimal Taxation

    Method: Dynamic mechanism-design model of training time, work and taxation; NBER abstract.

    Work and human-capital investment can be substitutes or complements, with different incentive implications.

    Limit

    The ACM extension models a session that displaces delivery. Productive learning while doing remains an omitted technology, not an impossibility result.

    What it means here

    State which work stops during mentoring; a different training method needs a different production model.

67. Benson Garner · Strategyzer · 2015-03-19

    How to Create a Crystal-Clear Business Model Canvas

    The method guidance recommends an appropriate level of detail, explicit connections, separation of present and intended states, and a visible distinction between evidence and assumptions.

    Limit

    Practical design guidance, not evidence that the ACM arrangement works. The ACM handoff worksheet is an original tool, not the Business Model Canvas or a validated derivative.

    What it means here

    Make the proposed commitment, surviving responsibilities, unknowns and next test visible together. Filling in a worksheet does not establish feasibility.

68. D. Sculley and coauthors / Google · 2015

    Hidden Technical Debt in Machine Learning Systems

    This systems paper identifies maintenance risks from entangled models, data dependencies, undeclared consumers, feedback loops and configuration debt; shared model outputs can create hidden coupling across systems.

    Limit

    A foundational practitioner paper predating generative AI, not recent evidence for a Capability Manager role or an estimated amount of waste.

    What it means here

    Qualify the original claim that one solution can simply serve every team. Reuse creates support obligations: declare consumers, version interfaces and data, monitor behavior, and fund migration and retirement.

69. American Economic Review 104(12):3779–3813 · 2014-12

    Optimal Allocation with Costly Verification

    Method: Mechanism-design theory; publisher abstract and citation inspected, not the full proof or all assumptions.

    The paper studies allocation of one object among agents with private principal values when verifying a report is costly and monetary transfers are unavailable.

    Limit

    Its optimal mechanism relies on its specified information and verification technology; an AI project's true value may not be perfectly checkable even ex post.

    What it means here

    Compare selective costly review with unverified ranking under a finite one-award constraint; model audit accuracy and cost explicitly.

70. Science 345(6203):1502–1505 · 2014-08-28

    Publication Bias in the Social Sciences: Unlocking the File Drawer

    Method: Empirical follow-up of 221 known completed social-science survey experiments; PubMed-indexed abstract inspected, not the full paper or coding appendix.

    In that cohort, null findings were substantially less likely to be written up and published than strong findings; much of the observed bias occurred before submission.

    Limit

    Academia is not an internal AI lab; non-write-up does not prove deliberate concealment, and a null finding is not necessarily an informative failed development experiment.

    What it means here

    Keep an independent ledger of all authorized and run experiments, including undiagnosed and null outcomes, before estimating learning from reported successes.

71. Econometrica 81(6):2463–2485 · 2013-11

    An Efficient Dynamic Mechanism

    Method: Dynamic mechanism-design theory; Stanford publication abstract, publisher metadata and author manuscript abstract/introduction inspected, not every proof.

    The authors construct efficient dynamic incentive mechanisms with private information arriving over time and quasilinear transfers; budget balance and self-enforcement need additional independent-type, patience and Markov conditions.

    Limit

    The result does not make an ACM pitch gate truthful without its transfers and informational assumptions, nor establish a central organizational advantage.

    What it means here

    Check each team's report, experiment and participation incentives at every continuation state; state exactly which future consequences a current report changes.

72. Theoretical Economics 7(3):425–463 · 2012-09

    A principal-agent model of sequential testing

    Method: Dynamic information-acquisition and reporting theory; journal abstract inspected, full text unavailable through the browser redirect.

    With costly hidden testing and hidden signals, their optimal baseline mechanism rewards messages consistent with the underlying state; payments for a correct good-state announcement increase with time.

    Limit

    The mechanism uses state-consistent monetary payments under specific signal and observability assumptions, not a scarce-resource award for an AI record whose eventual usefulness may be difficult to verify.

    What it means here

    Define what makes a disclosure diagnostically correct or useful, when it can be verified, and how its delayed reward affects incentives; keep report count separate from information value.

73. Research Policy / Harvard Business School · 2012

    Exploring the Duality Between Product and Organizational Architectures: A Test of the Mirroring Hypothesis

    Method: Matched software-product pairs comparing tightly coupled commercial organizations and loosely coupled open-source communities.

    Loosely coupled organizations produced more modular architectures in the pairs studied, measured through dependency and change-propagation structure.

    Limit

    Organizational form was not randomized; selection, product history and other differences may matter. Does not prove the direction of causation or superiority for every task.

    What it means here

    A hub that routes every decision through itself may create technical coupling. Examine dependency and communication patterns, not just reuse counts.

74. David Aaker · Prophet · 2012

    10 Steps to Build a Successful Brand Portfolio Strategy

    Method: Author's practitioner article; indexed primary-page text inspected, direct full fetch unavailable.

    Portfolio strategy considers context-specific roles and relationships among offerings rather than names in isolation.

    Limit

    Brand architecture does not establish shared technical lineage, common failure risk or model performance; full article not reviewed.

    What it means here

    Give each candidate a task-specific role and identify complements, alternatives and maintenance obligations. The site's operational role labels are its own proposal.

75. David and Brachet · American Economic Journal: Microeconomics · 2011-08

    On the Determinants of Organizational Forgetting

    The study distinguishes organizational forgetting through worker turnover from skill depreciation, using ambulance-company workforce data.

    Limit

    Publisher abstract consulted. Its estimated magnitudes are context-specific; neither channel is an estimated learning law for ACM.

    What it means here

    Separate departure from loss of retained skill and record applicability. A temporary staffing change need not erase every stock, but surviving stocks still require maintenance.

76. International Conference on Machine Learning 2011; author-hosted paper · 2011-03

    Doubly Robust Policy Evaluation and Learning

    Method: Contextual-bandit policy-evaluation theory and empirical comparisons; author-hosted abstract and introduction inspected, not the complete proofs or experiment tables.

    Logged data reveal rewards for selected actions; changing policy shifts the action mix, and model-only or propensity-only evaluation has different bias/variance failure modes. The paper studies a doubly robust estimator under its assumptions.

    Limit

    ACM proposals can also be missing before logging, actions can lack overlap, and outcomes can be delayed or changed by organizational behavior. Doubly robust estimation cannot identify never-tried alternatives without assumptions.

    What it means here

    Log eligible choices, selected actions and propensities; reserve bounded randomized exploration or audits where feasible, and state support and missingness conditions before policy comparison.

77. Matthew Bidwell / Administrative Science Quarterly · 2011

    Paying More to Get Less: The Effects of External Hiring versus Internal Mobility

    Method: Personnel records from one US investment-banking employer, 2003–2009; author manuscript.

    Entry through internal promotion and external hiring was associated with different initial performance, pay and exit outcomes in the studied employer.

    Limit

    One employer and nonrandom entry routes do not establish a universal causal benefit of training. The study does not estimate pre-entry development costs.

    What it means here

    Compare relevant existing knowledge and arrival performance as well as recruitment price; do not import its estimates into AI staffing.

78. Azoulay, Graff Zivin and Wang · Quarterly Journal of Economics / NBER · 2010-05

    Superstar Extinction

    Unexpected loss of eminent life scientists is followed by a lasting decline in their collaborators' quality-adjusted publication output; intellectual proximity matters to the spillover.

    Limit

    The revised NBER abstract was consulted. Academic publication effects do not calibrate AI teams, and death does not identify the full consequences of a negotiated job move.

    What it means here

    Retain a contribution that documentation and training may fail to replace. Test key-person dependence rather than inferring resilience from the number of files.

79. Google Research; KDD 2010 · 2010

    Overlapping Experiment Infrastructure: More, Better, Faster Experimentation

    Method: Google account of its experiment infrastructure; Google Research publication page and abstract inspected, not full paper.

    The authors describe overlapping experimentation infrastructure and say tools and educational processes were both required to use it effectively.

    Limit

    A historical platform case, not evidence that Google currently needs an additional central decision authority or that the same configuration works in another firm.

    What it means here

    For the Google-scale configuration, compare any proposed new expert function with the existing shared tooling and trained local experimenters.

80. American Psychologist / American Psychological Association · 2009-09

    Conditions for Intuitive Expertise: A Failure to Disagree

    Method: Integrative theoretical review/adversarial collaboration comparing naturalistic decision making with heuristics-and-biases research; not a new randomized intervention. Inspection: Original journal PDF hosted by a university inspected through abstract, skilled-intuition conditions, feedback/validity discussion, fractionated expertise, and conclusion; not every citation checked.

    Skilled intuition requires an environment with learnable valid cues and adequate opportunity to practice and receive useful feedback; subjective confidence alone does not establish skill, and expertise is often task-specific.

    Limit

    These are boundary conditions synthesized from prior domains, not a calibrated accuracy threshold or direct evidence for AI curricula. Some AI work has weak or changing feedback, so an impressive expert may lack valid intuition on a novel task.

    What it means here

    Make assessment task-specific and compare judgment with outcome feedback or a suitable rule; do not grant broad autonomy on reputation or confidence alone.

81. The Economic Journal 119(539):1036–1064 · 2009-07

    Strategic Disclosure of Research Results: The Cost of Proving Your Honesty

    Method: Formal sender–receiver model with endogenous research and verifiable selective disclosure; publisher abstract inspected, not full proof.

    Under some assumptions hidden research effort induces extra research and full disclosure to establish honesty, so mandatory disclosure can reduce welfare; other cases feature withheld evidence.

    Limit

    The comparative statics depend on its persuasion environment; they are not evidence about an AI lab or a universal effect of mandatory experiment logs.

    What it means here

    Let disclosure rules alter both the production and selection of experiments, and retain fixtures where mandatory reporting helps and harms.

82. INFORMS Journal on Computing 21(4):599–613 · 2009-05-19

    The Knowledge-Gradient Policy for Correlated Normal Beliefs

    Method: Sequential Bayesian ranking-and-selection theory; publisher abstract and bibliographic metadata inspected, not the full derivation or computational studies.

    For a specific correlated-normal belief model, the authors propose a sequential policy that chooses noisy measurements to improve later selection under a fixed measurement budget.

    Limit

    Its normal-belief assumptions and measurement budget do not map directly to AI-development outcomes, delayed feedback, strategic pitches, expert-hours or risk constraints.

    What it means here

    Evaluate an experiment by how it may improve later feasible choices, while separately charging time, compute, diagnosis and opportunity cost.

83. McKinsey Quarterly · 2008-03-01

    Enduring Ideas: The 7-S Framework

    Method: Official retrospective on the organizational framework.

    Organizational effectiveness depends on coordinating interrelated factors, beyond reporting structure alone.

    Limit

    This framework does not determine ACM headcount, authority or comparative value.

    What it means here

    Connect structure with skills, systems, resources and incentives rather than offering an organization chart alone.

84. Chip Heath and Dan Heath · 2008

    Made to Stick: SUCCESs Model

    The authors' communication framework emphasizes a clear core, concrete detail, credibility and stories that help people imagine action.

    Limit

    Communication guidance does not validate the mathematics or management proposal. Memorable claims still need evidence and stated limits.

    What it means here

    Keep one recognizable business situation, let the reader inspect the decision, and make the loss as understandable as the gain.

85. Journal of Political Economy 114(5):956–995 · 2006-10

    Adaptive Organizations

    Method: Formal organization-design model. Primary institutional record, paper introduction, model overview and propositions1–3/communication discussion inspected; appendix proofs not independently audited.

    Local adaptation trades off against coordination among specialized tasks. Better communication can make adaptation attractive enough to reduce optimal specialization; organizational direction depends on the configuration of the work.

    Limit

    A theoretical mechanism, not a measured ranking of AI operating models or a universal prescription for delegation.

    What it means here

    Review task interdependence, change and communication cost alongside team skill. A better tool does not by itself determine whether more work should be centralized.

86. Garicano & Rossi-Hansberg · NBER Working Paper 11458 · 2005-07

    Organization and Inequality in a Knowledge Economy

    Method: Knowledge-based organizational equilibrium; NBER abstract consulted.

    The model joins knowledge acquisition, heterogeneous skills and organizational assignments. It permits interpretations through firms or consulting and referral markets.

    Limit

    This is not evidence that a particular small team, polymath leader or central department is optimal. The original assumptions are distinct from the proposed AI staffing comparison.

    What it means here

    Model specialist access and integration work explicitly, and allow several institutions to supply the same capability.

87. Long Range Planning 38(3):299–319 · 2005-06

    Failing to Learn and Learning to Fail (Intelligently): How Great Organizations Put Failure to Work to Innovate and Improve

    Method: Authors' research synthesis and organizational framework; publisher abstract and introduction inspected.

    The authors distinguish identifying failure, analyzing it and deliberate experimentation. Their framework says technical and social barriers can prevent an observed failure from becoming an organizational lesson.

    Limit

    A framework, not an estimated probability that an AI experiment teaches. It does not show that central reviewers diagnose more accurately than embedded experts.

    What it means here

    Credit expertise only after attended, validated diagnosis; charge the work of identifying and analyzing failures separately from running a test. Include unexplained failures that produce no learning.

88. Long Range Planning 38(3):281–298 · 2005-06

    Learning from Failures: Why it May Not Happen

    Method: Qualitative study of 14 strategic failures in one large European telecommunications firm; publisher abstract and data-section excerpts inspected.

    The studied firm learned little from the 14 failures despite the visibility of those ventures. Failure occurrence alone did not produce useful change in decisions or routines.

    Limit

    One firm and strategic ventures, not AI model experiments. The case study does not identify a universal learning rate or prove that formal debriefs always work.

    What it means here

    Make 'diagnosed outcome changed a later choice' observable; permit zero learning after a failed program and test what a shared assessment process adds to the counterfactual.

89. Wallace J. Hopp, Eylem Tekin and Mark P. Van Oyen · Management Science · 2004-01-01

    Benefits of Skill Chaining in Serial Production Lines with Cross-Trained Workers

    Models of serial production compare cross-training choices and dynamic worker coordination; overlapping skill chains can provide robust, efficient flexibility within that setting.

    Limit

    Serial production and a constant work-in-process release policy differ from uncertain research. Neither personalities nor a universal generalist staffing advantage are established.

    What it means here

    Separate which additional skill is valuable from how qualified people are assigned when congestion occurs. Our finite model checks reservations; unknown-arrival dispatch remains an extension.

90. MIT Press (2004) · 2004

    The Venture Capital Cycle, second edition

    Method: Academic monograph. Edition, publisher and year confirmed in library catalogue listings; the description of its parts (fund-raising, investing, exits) comes from the distributor's description via search. The publisher's own page returned an access error; the book itself not read for this check.

    The book treats the venture capital cycle as stages: raising a fund from investors, investing it in companies, and exiting, and studies each separately.

    Limit

    A study of independent venture funds with outside investors, fixed lives and carried interest. A company fund has none of these, so the comparison is an analogy for roles only.

    What it means here

    Supports the chapter's habit of treating raising money, deciding who gets it and pitching for it as different jobs with different skills. It says nothing about the right size of a company's AI envelope.

91. Edward P. Lazear / NBER; Journal of Political Economy · 2003-05

    Firm-Specific Human Capital: A Skill-Weights Approach

    Method: Theoretical skill-weights model; NBER abstract and publication record.

    Skills that are transferable individually can form combinations valued differently by different firms.

    Limit

    This theory does not calibrate an AI training course, mentor productivity or a universal internal-development advantage.

    What it means here

    Define the task and context to which qualification transfers; a broad label such as AI skill is insufficient.

92. Noah Gans and Yong-Pin Zhou · Operations Research · 2002-12-01

    Managing Learning and Turnover in Employee Staffing

    Method: Markov decision model of staffing with random nonstationary service requirements, learning and turnover; publisher abstract consulted.

    The authors derive a state-dependent hiring policy and examine when explicitly modeling stochastic learning and turnover matters, including when flexible capacity is costly or unavailable.

    Limit

    Their hiring-policy theorem belongs to their assumptions. The ACM extension does not inherit it and does not model learning or exogenous turnover. Its finite causal-policy certificates are separate results.

    What it means here

    Treat borrowing availability and price as explicit alternatives, and distinguish existing qualified staff from capacity not yet ready to serve.

93. The Journal of Finance 57(5):1891–1921 · 2002-10

    Information Production and Capital Allocation: Decentralized versus Hierarchical Firms

    Method: Theory comparing organizational structures by information generation and capital allocation. Published abstract and May 2000 NBER working-paper record verified.

    Decentralization is more attractive when project information is difficult to credibly transmit; hierarchy gains relative advantage when information can be hardened and passed along without cost.

    Limit

    Stylized theory with a banking application, not a test of AI laboratories. Experimental logs do not automatically make tacit judgment transferable or eliminate incentives to withhold information.

    What it means here

    The Journalist function should make reasoning inspectable while preserving direct contact with local experts. Assess whether central allocation improves evidence and decisions, or merely rewards what fits a report.

94. Organization Science / INFORMS · 2002-06-01

    Deliberate Learning and the Evolution of Dynamic Capabilities

    Method: Conceptual organizational-learning model deriving testable hypotheses about experience accumulation, articulation, and codification across task conditions. Inspection: Publisher abstract and substantial author/institution-hosted preprint sections (introduction, heterogeneity and ambiguity hypotheses, conclusion) inspected; not every section or cited empirical study.

    Zollo and Winter argue that organizational capability evolves through interaction of accumulated experience, knowledge articulation, and codification, with relative value depending on task frequency, heterogeneity, and causal ambiguity; deliberate codification may be more valuable for less frequent, less homogeneous tasks than routine-scale logic suggests.

    Limit

    The article develops a theory and testable hypotheses, not an AI-era intervention or a general proof that writing a lesson causes durable skill. The cited empirical work is supporting context and varies by setting.

    What it means here

    Preserve the reasoning behind failures and test when explicit articulation helps transfer across infrequent or heterogeneous tasks; charge the time spent codifying as part of curriculum cost.

95. Harvard Business School Working Paper 02-078 · 2002-05

    The Option Value of Modularity in Design: An Example from Design Rules, Volume 1

    Method: Formal real-options account of modular design, illustrated through splitting and substitution.

    Value depends on technical potential, independent experiment cost and visibility constraints, less the cost of creating the modular architecture.

    Limit

    A valuation model with assumptions, not proof that any shared primitive creates positive value. Author institution repository abstract verified.

    What it means here

    Shared interfaces may enable competing experiments. Reuse need not mean a single centrally chosen implementation; abstraction must earn back its cost.

96. Avner Bar-Ilan, Agnès Sulem and Alessandro Zanello · Journal of Economic Dynamics and Control · 2002-01

    Time-to-build and capacity choice

    Method: Impulse-control model of irreversible capacity investment under uncertain demand and construction delays; author-institution abstract consulted, DOI 10.1016/S0165-1889(00)00018-X.

    Construction delays alter the effect of uncertainty on investment timing and size. The abstract reports that sufficiently long delays can reverse the usual direction in which uncertainty affects investment timing.

    Limit

    This is a capital-investment model, not workforce evidence. It supports examining the interaction between uncertainty and delay; it does not establish that uncertain AI demand should always accelerate recruitment.

    What it means here

    Vary delay and demand uncertainty separately. A general recommendation to wait for certainty can miss the period when qualified capacity must begin preparation.

97. Administrative Science Quarterly 46(4):597–621 · 2001-12

    Knowledge Specialization, Organizational Coupling, and the Boundaries of the Firm: Why Do Firms Know More than They Make?

    Method: Historical analysis of aircraft-engine control systems, developed into an argument about multitechnology firms and integration, with examples from other industries.

    The authors distinguish outsourcing activities from outsourcing technological knowledge. Systems integration can require knowledge beyond the components a firm produces, particularly when technologies develop unevenly and product interdependencies are unpredictable.

    Limit

    A sector-specific historical study, not an AI-training experiment or evidence that every capability should be centrally owned.

    What it means here

    Using external models need not mean giving up training expertise. Conversely, sharing weights does not eliminate the domain and integration knowledge needed to adapt and evaluate them. Specify what knowledge each side must retain rather than equating asset consolidation with capability integration.

98. Journal of Marketing 65(4):16–32 · 2001-10

    The Impact of Category Management on Retailer Prices and Performance: Theory and Evidence

    Method: Competitive-channel model and before/after intervention analysis of one supermarket chain's laundry-detergent category; publisher abstract and author-uploaded article verified.

    Retail category management coordinates decisions across brands for category performance. The model and empirical application show that this can raise retail prices while reducing sales; benefits are not automatically shared by all participants.

    Limit

    Retail pricing is not model development. The analogy identifies a scope of stewardship, not evidence that an AI capability category, central lab or organizational structure will work.

    What it means here

    Explain the category-management inheritance as responsibility for a related group of investments and their interactions. Define whose outcomes count; a locally attractive result can harm another participant.

99. Edward G. Anderson, Jr. · Management Science · 2001-06-01

    The Nonstationary Staff-Planning Problem with Business Cycle and Learning Effects

    Method: Optimal strategic staffing model with nonstationary stochastic demand and apprentice/experienced staff; publisher abstract consulted.

    The model balances training costs and delays against meeting changing demand. Its results relate hiring and firing policies to business-cycle duration and training time.

    Limit

    The paper's learning mechanism and industry applications are not calibration for AI teams. Our workforce model has stipulated qualification and arrival delays, with no learning curve or productivity estimate.

    What it means here

    Specify when recruited capacity can actually serve and include idle wages. Current demand alone need not justify the best staffing commitment.

100. Jennifer Aaker, Verónica Benet-Martínez and Jordi Garolera · Stanford GSB · 2001

     Consumption Symbols as Carriers of Culture: A Study of Japanese and Spanish Brand Personality Constructs

     Method: Stanford primary abstract inspected; four studies across the United States, Japan and Spain.

     The studies identify common and culture-specific aspects of perceived brand personality.

     Limit

     The constructs cannot be assumed universal or transferred to AI systems without contextual measurement. This is not a taxonomy of machine abilities.

     What it means here

     Treat proposed interaction qualities as hypotheses about users' perceptions, not universal model labels.

101. The Journal of Finance 55(6):2537–2564 · 2000-12

     The Dark Side of Internal Capital Markets: Divisional Rent-Seeking and Inefficient Investment

     Method: Two-tiered agency theory. Journal metadata and abstract verified through RePEc and the DOI record; full proofs not audited.

     When division managers can lobby for a larger share, rent-seeking can subvert an internal capital market: allocation then follows bargaining power as well as merit, and stronger divisions can end up subsidizing weaker ones.

     Limit

     A stylized corporate-finance model of conglomerate divisions, not a measurement of AI request queues. Its mechanism is rent-seeking by managers who also want pay and influence; a free priority label is a similar but not identical device.

     What it means here

     Supports treating the priority label as a cheap way to bid for a share, and supports a Portfolio Manager who allocates against a budget and written evidence. It does not show that any particular committee size or rule works.

102. Journal of Political Economy 108(5):874–904 · 2000-10

     Hierarchies and the Organization of Knowledge in Production

     Method: Formal knowledge-allocation theory. Published paper introduction, model assumptions and relevant comparative statics inspected; not a new audit of every appendix proof.

     Knowledge specialization can economize on learning common versus unusual problems. Communication consumes specialist time even when the specialist cannot answer. Lower communication costs can reduce frontline knowledge acquisition because referral becomes cheaper.

     Limit

     A static knowledge architecture, not an AI-team trial. The model explicitly excludes learning through communication by assuming the same problem does not recur. It therefore does not establish the apprenticeship channel added here.

     What it means here

     Price escalation and expert attention. Distinguish access to an expert from development of independent judgment, and compare specialist support with investment in local capability.

103. American Economic Review 90(4):1034–1054 · 2000-09

     Learning and Forgetting: The Dynamics of Aircraft Production

     Method: Empirical analysis of commercial-aircraft production costs. AEA publication metadata and the author's NBER working-paper abstract verified; author-hosted published PDF is available. NBER working paper 7127 dates to May 1999.

     The studied cost data favor a learning model in which production experience depreciates over a model based only on cumulative output. Some, but not all, experience transfers between aircraft generations.

     Limit

     Evidence from aircraft production, not model development. Neither the estimated depreciation nor transfer parameters calibrate AI-team learning, and the findings do not identify an optimal centralization policy.

     What it means here

     Allow expertise to decay and lessons to transfer incompletely in the research model. Estimate or test those mechanisms in the AI setting; do not assume that more accumulated experiments guarantee monotonically better performance.

104. Marketing Science 18(3):208–229 · 1999-08-01

     Accounting Profits Versus Marketing Profits: A Relevant Metric for Category Management

     Method: Formal model linking category merchandising, store choice and store profits; illustrative estimation using Marsh Supermarkets data. Publisher abstract verified.

     A category's accounting contribution can omit its effects on other categories through consumers' store choice. The proposed measure explicitly incorporates these cross-category effects.

     Limit

     The measurement relies on demand and retailer-optimization assumptions. A retail demand externality is not evidence of an AI knowledge externality, and the paper supplies no ready-made measure of organizational coherence.

     What it means here

     Use this as an analogy for examining portfolio effects beyond one program's ledger. Record who benefits, who pays adaptation costs and how spillovers would be measured before treating them as value.

105. Journal of Law, Economics, and Organization 15(1):56–73 · 1999-04

     Informal Authority in Organizations

     Method: Formal relational-contract model. Author-hosted November1998 manuscript abstract, introduction, model and reneging conditions inspected; published journal metadata checked. No empirical AI effect is estimated.

     A boss can retain formal authority while informally delegating decisions. Delegation depends on a self-enforcing relationship because a later intervention may be attractive. Information can change the incentive to respect that arrangement.

     Limit

     A stylized delegation mechanism, not proof that an ACM autonomy promise is enforceable. The paper does not calibrate learning rates or establish a curriculum.

     What it means here

     Separate the value of delegated decisions from the credibility of leaving them delegated. Check incentives to reverse the promise and the cost of making it effective.

106. Econometrica 67(2):349–374 · 1999-03

     Strategic Experimentation

     Method: Many-agent bandit theory; publisher abstract and bibliographic page inspected, not full derivation.

     Publicly observed experimentation makes information a public good; stationary Markov equilibria reflect both free-riding and an encouragement effect from expected future experiments by others.

     Limit

     The model assumes public feedback and does not solve selective reporting, noisy diagnosis or finite ACM allocation.

     What it means here

     Test whether shared learning crowds out, brings forward or increases private experimentation under the declared continuation incentives.

107. Daron Acemoglu and Jörn-Steffen Pischke / NBER; Journal of Political Economy · 1998-01

     The Structure of Wages and Investment in General Training

     Method: Labor-market theory; NBER abstract and publication record.

     Labor-market frictions and wage compression can change an employer’s incentive to finance general training.

     Limit

     The joined ACM staffing model fixes wages and employee retention; it does not implement this wage-setting equilibrium.

     What it means here

     Employer returns, employee benefits and retention incentives require separate modeling; current company-value results are incomplete welfare claims.

108. The Journal of Finance 52(1):111–133 · 1997-03

     Internal Capital Markets and the Competition for Corporate Resources

     Method: Formal theory of headquarters allocation among competing projects. Published abstract and NBER working-paper record verified; NBER version dates to May 1995.

     Control rights allow headquarters to move scarce resources between projects. Better selection can create value without increasing the firm's total available capital; the model can favor a focused set of projects.

     Limit

     A theoretical possibility, not evidence that headquarters always selects better. Applying financial-capital logic to GPU capacity requires attention to hardware, scheduling and nonfungibility.

     What it means here

     The Internal VC lens needs a real, bounded allocation mandate. Show the next allocation, displaced alternative and continue/stop decision rather than presenting a catalog or recommendation as investment authority.

109. Journal of Political Economy 105(1):1–29 · 1997-02

     Formal and Real Authority in Organizations

     Method: Formal organizational theory; publisher and institutional-author abstracts, publication metadata and indexed communication discussion inspected. Full proof not audited.

     Formal decision rights and effective control differ. Information acquisition, initiative and communication respond to authority; more agent initiative can involve less principal control.

     Limit

     Stylized theory rather than a measured ranking of AI organizations or an optimal centralization percentage. A technically capable center does not automatically possess local information or preserve local initiative.

     What it means here

     Specify who decides separately from who investigates, performs the work or advises. Give a serious federated comparator access to technical evidence and learning; test handoff costs, review delays and lost initiative as well as allocation accuracy.

110. Jennifer Aaker · Journal of Marketing Research / Stanford GSB · 1997

     Dimensions of Brand Personality

     Method: Stanford primary publication abstract inspected; measurement research on perceived brand personality.

     The paper develops a framework for consumer-perceived symbolic brand personality.

     Limit

     Perceived competence is not measured accuracy, robustness or reliability. A system's voice does not demonstrate human intention or character.

     What it means here

     Study interaction style and user perception separately from task evaluation, in a specified user context.

111. The Journal of Finance 51(4):1139–1174 · 1996-09

     The Capital Budgeting Process: Incentives and Information

     Method: Formal model of internal capital budgeting under decentralized information and incentive problems. Publisher abstract and bibliographic metadata verified.

     Information and incentive problems can generate underinvestment in high-productivity opportunities and overinvestment in low-productivity ones. The appropriate budgeting process depends partly on opportunities and information-transfer technology.

     Limit

     Does not identify a universal allocation rule or establish that an internal technical evaluator should control all funding. AI programs have uncertainties beyond those in this model.

     What it means here

     Separate a lab's technical case from approval of its own resource request. Make assumptions, competing uses and review authority explicit; revisit allocations when evidence changes.

112. Strategic Management Journal · 1996

     Exploring Internal Stickiness: Impediments to the Transfer of Best Practice Within the Firm

     Method: Survey of 122 transfers of 38 practices in eight firms; 271 source, recipient and third-party questionnaires; correlational analysis.

     Recipient absorptive capacity, causal ambiguity and difficult source-recipient relationships dominated the measured transfer barriers, rather than motivational explanations alone.

     Limit

     Selected surviving transfers, missing observations and correlational design restrict causal claims. Aborted transfers were difficult to observe.

     What it means here

     Incentives and discovery cannot alone fix transfer. Fund explanation, recipient learning and adaptation; record failed and refused transfers.

113. The Review of Economic Studies 62(4):541–555 · 1995-10

     Credit and Efficiency in Centralized and Decentralized Economies

     Method: Adverse-selection credit and refinancing theory; publisher abstract and bibliographic record inspected, not full proof.

     Sunk investments can create an ex-post incentive to refinance a project known to be low-quality; decentralized credit can harden the no-refinancing commitment, while placing excessive weight on short-term returns.

     Limit

     A credit-market model, not evidence that central AI governance creates soft budgets or that stopping a project is always efficient.

     What it means here

     Keep any disclosure reward transferable to a newly qualified request rather than guaranteed renewal of the diagnosed project; test both rescue of bad work and premature termination of valuable learning.

114. William C. Jordan and Stephen C. Graves · Management Science · 1995-04-01

     Principles on the Benefits of Manufacturing Process Flexibility

     In a production-allocation model with uncertain demand, appropriately connected limited flexibility can capture much of the benefit of complete flexibility.

     Limit

     Plants and products are not people and AI projects. This does not establish that a small ACM should replace specialist departments; our staffing benchmark has known work rather than their uncertain demand.

     What it means here

     Treat the placement of overlapping capability as a design choice. Ask which missing connection would remove stranded capacity before training everyone in everything.

115. The Quarterly Journal of Economics 109(4):1125–1156 · 1994-11

     Subjective Performance Measures in Optimal Incentive Contracts

     Method: Theory of implicit subjective and explicit objective incentives; publisher abstract and NBER working-paper abstract inspected, not full proof.

     Subjective and objective performance measures can be complements or substitutes in an ongoing incentive relationship; the feasibility of a promised subjective reward depends in part on the firm's incentive to honour rather than renege.

     Limit

     The analysis concerns performance pay, not GPU capacity or an enforceable internal grant. It does not imply all discretionary rewards fail or specify the needed commitment technology here.

     What it means here

     Treat sponsor follow-through on a usefulness-contingent future-access promise as an explicit incentive constraint, or state and implement an exogenous committed rule.

116. Journal of Political Economy 100(3):598–614 · 1992-06

     Incentive Contracts and Performance Measurement

     Method: Principal–agent theory of imperfect performance measures; publisher abstract inspected, not full proof.

     Rewarding a measure that differs from the principal's objective generally fails to produce first-best incentives, even with a risk-neutral agent; the paper includes gaming of measures as an application.

     Limit

     The model identifies a proxy-incentive problem but does not supply an operational measure of useful AI diagnostic disclosure or show that every record-count policy is worse.

     What it means here

     Compare access priority awarded for report volume with priority awarded for verifiable diagnostic contribution, including verification error and effort displaced from project work.

117. Organization Science / INFORMS · 1991-02-01

     Exploration and Exploitation in Organizational Learning

     Method: Two formal models of organizational learning and competition.

     Adaptive processes can refine exploitation faster than exploration, producing short-term effectiveness while undermining longer-term learning.

     Limit

     A modeled mechanism, not an estimate of the right experimentation budget or evidence that centralization always suppresses research. Publisher abstract verified.

     What it means here

     A hub rewarded for consolidation can select familiar recipes too early. Separate operational reuse from funded exploration and define when competing variants should converge.

118. American Educator / American Federation of Teachers · 1991

     Cognitive Apprenticeship: Making Thinking Visible

     Method: Conceptual pedagogical framework with illustrated cases, not a randomized AI or workplace trial. Inspection: Primary 1991 author exposition framing/framework/examples inspected; original 1989 chapter excerpt only, full chapter not read.

     The authors' cognitive-apprenticeship framework makes expert reasoning visible through modeling, coaching, scaffolded practice, articulation, reflection, and exploration with fading support.

     Limit

     Pedagogical framework and author's exposition of the earlier Collins–Brown–Newman chapter; neither an AI trial nor an estimated effect size. Original 1989 full PDF was not read.

     What it means here

     Design a curriculum whose support fades and whose learners can independently frame, diagnose, explain, and transfer a problem-solving approach.

119. Administrative Science Quarterly · 1990-03

     Absorptive Capacity: A New Perspective on Learning and Innovation

     Method: Theoretical account of learning, organizational communication and R&D investment, with empirical tests of R&D predictions.

     Using outside knowledge requires related expertise throughout recipient units. Pages 132–134 explicitly warn that centralized gatekeepers can be ineffective during rapid uncertain change; some overlapping expertise supports cross-functional absorption.

     Limit

     Does not establish the optimal distribution of AI specialists. Its mechanism concerns knowledge and learning, not simply code duplication.

     What it means here

     A capable CM cannot replace product-side ability to evaluate, adapt and maintain AI. Redundant implementations and overlapping expertise have different economics. Pool scarce training craft while retaining knowledgeable local partners who can specify needs, assess evidence and maintain an adaptation.

120. Management Science 36(2):140–154 · 1990-02-01

     The Persistence and Transfer of Learning in Industrial Settings

     Method: Empirical analysis using data from multiple industrial organizations; publisher abstract and bibliographic record inspected.

     In the studied production settings, acquired knowledge depreciated rapidly; cumulative output overstated its persistence. There was limited evidence of cross-organization transfer, with little benefit from others' learning after own production began.

     Limit

     Industrial production differs from AI development; observational patterns do not set case-coverage, decay or transfer coefficients for the proposed model. This source adds cross-organization transfer evidence; the separately cited Benkard study covers depreciation in aircraft production.

     What it means here

     Treat retained expertise and recipient transfer as separately measured states and costs. Test whether the assessor actually reuses diagnoses across product contexts instead of presuming every prior case remains valuable.

121. Avinash Dixit · Journal of Political Economy · 1989-06-01

     Entry and Exit Decisions under Uncertainty

     Method: Theoretical entry/exit model; publisher abstract consulted.

     Under stochastic output prices and sunk entry/exit costs, the model produces different entry and exit triggers. The gap makes the current operating state matter.

     Limit

     This is a firm entry/exit model, not a theorem about ACM. Transferring the mechanism requires explicit organizational states, information, switching costs and future value.

     What it means here

     Propose separate service activation and retention conditions rather than treating organizational changes as instantaneous and free.

122. Annual Review of Sociology, volume 14 · 1988-08

     Organizational Learning

     Method: Conceptual synthesis of organizational-learning literature; publisher abstract and bibliographic record verified.

     Organizations learn by interpreting experience and incorporating lessons into routines that guide behavior. Organizational memory concerns encoding, retaining and retrieving those lessons despite personnel turnover and the passage of time.

     Limit

     A review and theoretical perspective, not an AI-team experiment or an estimated learning rate. Retained routines can embody mistaken or outdated interpretations; experience and stored records do not guarantee improved judgment.

     What it means here

     Define useful experimental memory by a traceable change in a later decision or practice. Test whether the ACM team improves the interpretation, retention and use of experience rather than using document or experiment counts as learning outcomes.

123. Daniel S. Hamermesh · NBER Working Paper 2572; later American Economic Review · 1988-05

     Labor Demand and the Structure of Adjustment Costs

     Method: Monthly plant-level evidence on employment adjustment; NBER working-paper abstract consulted.

     The study reports discontinuous employment adjustment: small demand shocks can leave employment unchanged, while larger shocks induce jumps. It connects this pattern to the structure of adjustment costs.

     Limit

     Plant-level historical evidence does not identify AI staffing thresholds. The linked record is the 1988 working paper and lists the 1989 journal publication; no reported magnitude is transferred to our examples.

     What it means here

     Charge for recruitment and release instead of treating workforce size as a frictionless response to every demand movement. Retention and fresh recruitment can have different thresholds.

124. American Journal of Sociology 94: S154–S179 · 1988

     An Economic Approach to Influence Activities in Organizations

     Method: Formal organizational-economics analysis; publisher abstract and bibliographic details inspected, not the full derivation.

     Interested organizational members can have incentives to manipulate private information they supply to decision makers, degrading decisions and diverting effort.

     Limit

     The model does not estimate misreporting in AI teams or imply that participation and advocacy should simply be eliminated.

     What it means here

     Make team-private evidence, submitted forecasts and the cost of influence activity separate variables when comparing allocation policies.

125. The Review of Economic Studies 52(2):173–191 · 1985-04

     Planning under Incomplete Information and the Ratchet Effect

     Method: Dynamic asymmetric-information planning theory; publisher abstract and bibliographic record inspected, not full proof.

     When a planner revises future terms using a firm's observed performance and cannot commit to a revision procedure, the firm may underproduce to avoid a more demanding future schedule.

     Limit

     The agent hides productive capacity through output in a central-planning model; this does not prove that an AI team will hide a failed experiment or quantify the effect of any future-access rule.

     What it means here

     Test whether improved future screening after a candid record reduces the reporter's own continuation payoff enough to suppress disclosure; specify commitment to the full future rule.

126. Journal of Economic Theory 27(2):245–252 · 1982-08

     Rational Cooperation in the Finitely Repeated Prisoners' Dilemma

     Method: Finite repeated-game theory; publisher abstract and author-hosted opening model inspected, not every proof.

     Complete-information finite prisoners' dilemma unravels to stage-game defection, while incomplete information about a committed reciprocal type can sustain cooperation until near the end.

     Limit

     A terminal-stage reporting failure is not a universal result for all finite games; current credit, persistent private types and different stage equilibria can change backward induction.

     What it means here

     Include both a terminal-reporting unraveling fixture and a reputation/current-credit counterfixture; name the precise assumptions behind each.

127. Management Science 28(6):604–620 · 1982-06-01

     Asymmetric Information, Incentives and Intrafirm Resource Allocation

     Method: Formal intrafirm allocation model; publisher abstract and bibliographic metadata inspected, not the complete mathematical derivation.

     Divisional productivity is private information; managers value resources because they substitute for costly unobservable effort. Optimal transfer-pricing mechanisms depend on whether resource-production capacity can bind.

     Limit

     An incentive model, not evidence that managers generally lie or that technical review eliminates strategic claims. Its preferred mechanism is conditional on its effort, information and compensation assumptions.

     What it means here

     Specify what requesters privately know and why overstating need benefits them. Treat urgency inflation as a testable mechanism; compare mechanisms under both elastic and binding capacity.

128. Journal of Economic Behavior & Organization · 1982

     Towards an Economic Theory of the Multiproduct Firm

     Method: Economic theory of organizational knowledge, specialized assets and transaction costs.

     Economies of scope can justify joint production without proving that it must be organized inside one integrated enterprise; alternative arrangements can share inputs.

     Limit

     Analysis of firm boundaries, not a direct test of an internal AI hub. Application to internal organization is an analogy that requires checking actual costs.

     What it means here

     Technical commonality alone does not justify central ownership. Compare governance arrangements after including coordination, adaptation and control costs.

129. Econometrica 49(5):1261–1288 · 1981-09

     Funding Criteria for Research, Development, and Exploration Projects

     Method: Sequential project-funding theory. Author-hosted published-paper opening and institutional abstract for the April 1979 MIT report inspected; full mathematical derivation was not available through the browser.

     Funding criteria must account for subsequent decisions to review and continue a project using new information, rather than treating investment as a single irreversible approval.

     Limit

     Does not supply AI-specific priors, experiment costs or payoff estimates, and does not establish that a centralized team learns faster.

     What it means here

     Separate funding a bounded experiment from committing to full development. Predefine the decision it informs and the evidence for continuing, stopping or changing course.

130. Interfaces / INFORMS · 1974-05-01

     Organization Design: An Information Processing View

     Method: Conceptual organization-design framework linking task uncertainty, cognitive limits and information-processing requirements.

     Organizational responses include self-contained tasks, information systems and lateral relationships; adding coordination capacity is one response among alternatives.

     Limit

     Theory, not an experiment showing that any particular hub improves AI delivery. The author's site reproduces the scanned paper and summarizes its alternatives.

     What it means here

     Diagnose an information or dependency problem before adding a role. A capability map is a way to see work; it is not itself a mechanism that resolves dependencies.

131. Communications of the ACM · 1972-12-01

     On the Criteria to Be Used in Decomposing Systems into Modules

     Method: Worked comparison of alternative decompositions of a software design problem.

     The benefit of modularization depends on how the boundaries are drawn, not merely on splitting the system into named parts.

     Limit

     A foundational design argument and example, not an enterprise organization study. Publisher abstract accessed; PDF fetch unavailable.

     What it means here

     Language, perception and agents are labels, not automatically useful component or ownership boundaries. Test which changes can be isolated behind a stable contract.

132. Bruce Henderson · Boston Consulting Group · 1970

     The Product Portfolio

     Method: Original BCG practitioner essay; full text inspected.

     The growth–share portfolio connects investment needs and cash generation across a collection of products.

     Limit

     Its market-share and growth assumptions are not AI capability measures. A rare but critical fallback cannot be dismissed because internal adoption is low.

     What it means here

     Examine portfolio roles and continuing investment together; do not relabel the original axes as model quality and usage.

133. The Annals of Mathematical Statistics 24(2):265–272 · 1953-06

     Equivalent Comparisons of Experiments

     Method: Statistical decision theory; original abstract and bibliographic metadata verified through the mathematical bibliography, while the publisher full text was access-controlled and not inspected.

     The original comparison orders experiments by attainable decision risk across decision problems; a more informative experiment can be used to reproduce the decisions available from a less informative one under a common decision problem.

     Limit

     This is a gross-information ordering, not a claim that buying a stronger test has positive net value after human/GPU cost, delay, strategic entry, model misspecification or a changed feasible action set.

     What it means here

     Show the gross value-of-information inequality in the finite gate, then subtract assessment cost and lost continuation options. State the likelihood and feasible decision set instead of equating more measurement with better portfolio outcomes.