Worked example · enterprise AI strategy
History supports workflow redesign, governed use and organizational complements. It does not justify a numerical enterprise-AI uplift, a uniform three-to-five-year realization schedule or durable adopter-margin capture.
Why this example: one familiar analogy produces four different claim outcomes. Countercase keeps those outcomes separate instead of turning “AI is transformative” into one undifferentiated forecast.
The decision path
Evidence stays attached throughout. Human judgment remains the final authority.
Name the decision, the claim, the horizon and what would change your mind.
Build a reference class with included, excluded and negative cases.
Test mechanisms, decisive disanalogies, complements and falsifiers.
Leave claim-level verdicts, evidence limits and an advisory recommendation.
01 · Decision under challenge
The packet does not ask whether “AI works.” It asks which inference survives: comparability, timing, complements or durable margin capture. Each claim carries its own evidence, estimate tag and confidence.
Enterprise AI is a general-purpose technology whose diffusion should produce material productivity and margin gains within the decision horizon when complementary organizational investments are present.What changes now
02 · Claim outcomes
Evidence is allowed to move each claim differently. The strongest result is a bounded complement mechanism; the timing and margin claims remain withheld from forecast use.
Earlier general-purpose-technology histories can transfer mechanisms and measurement risks, but their lag coefficients and uplift estimates do not transfer to enterprise AI.
Use the historical cases to challenge mechanisms and boundary conditions; refuse a numerical lag forecast.
Specific workflows show early gains and historical digital systems show adjustment periods, but the evidence does not establish a firm-wide three-to-five-year realization base rate.
Model governed workflow cohorts and explicit scenarios instead of one enterprise uplift date.
Redesign, skills, integration and control work recur as realization conditions and testable moderators; the package does not prove that one universal complement bundle is necessary in every workflow.
Budget and measure organizational complements with the tool, with workflow-specific falsifiers.
No eligible source in this package measures both enterprise-AI operating margin and persistence after implementation, vendor, control, failure and price-pass-through costs.
Do not change durable-margin assumptions until a separately governed value-capture bridge is evidenced.
03 · Reference class
Retain every frozen candidate, assign its comparison role after source review, require adoption timing, a complementary change, an observable outcome and an allocation-of-gains question, and record decisive disanalogies before using the case.
This purposive twelve-case set is too small and too heterogeneous to estimate a numerical enterprise-AI base rate. It transfers mechanisms, boundary conditions, lag shapes and value-capture questions only.
c. 1780–1900 · moderate confidence
Aggregate contribution rose only after long diffusion; establishment and powered-machinery changes matter more than one invention date.
Gains depended on establishment scale, capital and location; steam availability alone did not identify the beneficiary.
Physical motive power, fixed capital and geography differ from metered probabilistic software with rapid model change.
improved engines · powered machinery · capital deepening · larger establishments · coal and transport access
c. 1880–1930 · high confidence
System-wide diffusion took decades, while individual adopters could improve faster after factory redesign; those are different clocks.
The useful asset was a redesigned production system, not merely electricity consumption.
Aggregate electrification history is partly a diffusion measure and cannot set a waiting period for one AI-enabled workflow.
unit drive · factory layout redesign · new machinery · managerial learning · grid access
c. 1840–1910 · moderate confidence
Value followed the operating institution—messages, roles, confirmation and managerial routines—not transmission speed alone.
Coordination capability accrued through a managed network and operating protocols.
Deterministic message transport does not reproduce generative error, model drift or judgment substitution.
standard messages · specialist operators · authority rules · confirmation · management information
c. 1955–1985 · high confidence
Invention, standardization, infrastructure adoption and firm usage form separate clocks.
Ports, carriers, shippers, workers and consumers experienced different gains and losses.
Trade coefficients from a physical network standard cannot transfer to cognitive-work productivity.
standards · ports and cranes · ships and rail interfaces · customs practice · network adoption
c. 1960–2000 · high confidence
Long-difference firm estimates exceeded one-year effects, consistent with adjustment and organizational co-investment.
Technology producers and a minority of complement-rich adopters may capture value before the average user firm does.
Historical computing was embodied capital; AI is often metered external software acting on harder-to-measure cognitive output.
skills and training · teams · distributed decision rights · process redesign · organizational capital
c. 1985–2010 · moderate confidence
Operational performance can deteriorate at deployment, improve through learning and remain sensitive to maintenance and flexibility costs.
Vendors and implementers receive earlier, more certain revenue than customers receive residual value.
ERP is deterministic transaction infrastructure; generative AI can remain an overlay and adds stochastic factual and judgment errors.
cross-functional authority · data standardization · internal experts · consultants · training · integration capacity
c. 1990–2015 · moderate confidence
Micro-level price and transaction effects preceded large measured macro effects and sustainable profits for many adopters.
Consumers could receive surplus while branded incumbents retained advantage and weak retailers discounted without durable profit.
E-commerce reorganized an external market channel; much enterprise AI operates inside the firm and changes judgment quality.
fulfilment · payments · inventory integration · brand and trust · customer service
c. 2005–2025 · moderate confidence
Compute efficiency improved through learning over several years, with persistent firm and divisional heterogeneity.
Capable users can reduce waste while concentrated providers retain rents through usage pricing and switching barriers.
CPU utilization is an input-efficiency measure, not labour productivity, revenue, margin or output correctness.
architecture expertise · FinOps · observability · demand forecasting · portability · contract governance
c. 1970–2025 · high confidence
Firm, industry and local-labour-market estimates describe different units and can point in different directions.
Productivity and lower prices can coexist with lower labour share, local job loss or gains at selected adopter firms.
Robots are embodied, task-bounded capital; generative AI changes cognitive work and quality assurance across many workflows.
production redesign · capital integration · maintenance · skills · product demand
c. 1975–1995 · moderate confidence
XCON became useful before full autonomy, but accuracy improvement and knowledge maintenance were multi-year, continuing work.
A bounded system could create value, but the defensible asset was the maintained knowledge, test and production institution around it.
XCON used explicit deterministic rules in a bounded verifiable task; generative AI operates over ambiguous language and external model change.
domain experts · knowledge engineers · quality assurance · software integration · automated tests · human redundant checking
c. 1940–1980 · moderate confidence
Outcomes varied with location-specific diffusion of varieties, irrigation, inputs, research, policy and market access.
Consumers can capture gains through lower prices while producer outcomes depend on input costs, land, institutions and access.
Biological innovation and agricultural infrastructure are too remote for pooled AI effects; the case only tests package and distribution logic.
crop varieties · irrigation · fertilizer · research systems · policy · market access
c. 1950–1990 · moderate confidence
Performance reflects an operating system of flow, problem exposure, learning and response—not adoption of one tool.
Operational gains and worker welfare are distinct outcomes and can vary with implementation context.
TPS is an organizational system rather than a technology diffusion event; use it to test governance, not to forecast AI uplift.
just-in-time flow · jidoka · andon · standard work · supplier coordination · worker problem solving
04 · Mechanism challenge
Historical cases transfer mechanisms and failure tests. They do not transfer published coefficients into an enterprise-AI forecast.
Comparable firms realize durable end-to-end gains from production AI without measurable training, workflow, data, integration or control changes.
Production cohorts show immediate, stable and organization-wide gains with no implementation dip, learning curve or cross-unit heterogeneity.
Ungoverned AI use outperforms otherwise comparable governed workflows on output, error, rework and auditability over repeated production cycles.
Observed workflow productivity gains translate one-for-one into durable adopter operating margins after prices, implementation, vendor, control and failure costs.
Decision implication
No source in the evidence package measures both enterprise-AI operating margin and persistence. The commercial claim therefore remains withheld.
Gross productivity−price pass-through−implementation−vendor rents−controls−failure loss
05 · Evidence lineage
A citation is only the pointer. The worked packet preserves the claim-relevant excerpt, its source version and its limits, then binds that evidence to the exact claim it can support.
Paul A. David · 1989 · 24 preserved words
Although the analogy between information technology and electrical technology would have many limitations were it to be interpreted very literally, it nevertheless proves illuminating.
Historical technology analogies can illuminate diffusion and measurement mechanisms when their limits are explicit.
AI and electrification equivalence, a transferable numerical lag or a productivity uplift.
ca5b778d340d8c…91e20cead3336177cdb4b6…bcdd6868Erik Brynjolfsson, Daniel Rock and Chad Syverson · 2018 · 24 preserved words
General purpose technologies (GPTs) such as AI enable and require significant complementary investments, including co-invention of new processes, products, business models and human capital.
An AI-as-general-purpose-technology model requires complementary intangible investment and can create delayed measurement effects.
Proof that every enterprise-AI deployment is a general-purpose technology, a fixed realization horizon or adopter margin capture.
2b03c6749768e1…5034024b065bd67716a84a…10bb6cceMIT Initiative on the Digital Economy funding acknowledged by the authors.
Erik Brynjolfsson and Lorin M. Hitt · 2003 · 23 preserved words
the productivity and output contributions associated with computerization are up to five times greater over long periods (using five to seven year differences).
In a large firm panel, measured computer contributions differed materially between short and longer specifications, consistent with time-consuming complements.
A universal causal effect, an enterprise-AI lag coefficient or profitability.
46b9b55c93bc56…b5e479e6a35fc3dfccc645…9f4800daLorin M. Hitt, D. J. Wu and Xiaoge Zhou · 2002 · 24 preserved words
there is a slowdown in business performance and productivity shortly after the implementation … Due to the lack of mid- and long-term post-implementation data
ERP adopters showed favorable measures alongside an implementation slowdown and insufficient long-run post-implementation data.
Causal attribution, long-run durability, transfer to AI or durable operating-margin uplift.
53f52f5cb0a0ed…82eeee16fb4c4a2f811384…539aaad9Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond · 2023 · 17 preserved words
Access to the tool increases productivity, as measured by issues resolved per hour, by 14% on average
A workflow-embedded assistant raised a specific quality-adjusted throughput measure for 5,179 support agents, with material worker heterogeneity.
Enterprise-wide total factor productivity, broad adoption, three-to-five-year durability or operating margin.
c53bb4db549516…e1c9e7e135162d9527b6bf…3868be69One company, one workflow and a proprietary system bound external validity.
Eleanor W. Dillon, Sonia Jaffe, Nicole Immorlica and Christopher T. Stanton · 2025 · 25 preserved words
treated workers … spent two fewer hours on email each week … we do not detect shifts in the quantity or composition of workers’ tasks
Across 66 firms and 7,137 workers, individual time savings appeared without detectable broader task reallocation during the six-month experiment.
Productivity, profitability, persistence or a three-to-five-year firm forecast.
b2b5355292552c…5a3917ac9facc153b51eb0…53f66062Some authors were Microsoft employees; Microsoft performed privacy review and the authors report retaining discretion over results.
Anders Humlum and Emilie Vestergaard · 2025 · 24 preserved words
precise null effects on earnings and recorded hours at both the worker and workplace levels … task reorganization—including … AI oversight, and AI integration
Administrative Danish evidence shows task reorganization and new complementary work can precede visible earnings or hours effects.
Null productivity or margin effects at three-to-five years; follow-up is about two years and outcomes are earnings and hours.
3eb64361362af8…f46438ab7c0839eade7cea…90550b12The authors acknowledge support from the Center for Applied Artificial Intelligence and the Polsky Center for Entrepreneurship and Innovation.
Kathryn Bonney and coauthors · 2026 · 19 preserved words
Among adopters, scope remains limited … Regression results show a positive relationship between firm performance and AI integration breadth.
Nationally representative linked data show that firm adoption often covers few functions and integration breadth correlates with performance.
Causality, durable productivity, operating margin or a defined three-to-five-year cohort forecast.
9946482f04c35c…e7d017e11cff339568376e…e3b937c5The performance result is observational; public-data status does not license the working paper.
Fang and coauthors · 2026 · 16 preserved words
GenAI adoption increases sales in most workflows, with effects ranging from no detectable impact to 16.3%
Seven randomized workflow experiments within one retailer found heterogeneous short-run sales effects.
Enterprise-wide productivity, implementation and inference costs, operating margin, persistence or external validity.
88e380436cb6a3…40b2189ad951cbcbda22b3…dcbd9b8bTwo authors consulted for the platform and one was its employee.
UK Competition and Markets Authority · 2026 · 22 preserved words
The investigation identified limits to customer choice as a result of data egress fees and barriers to interoperability restricting switching and multi-cloud
Cloud infrastructure can impose switching costs and bargaining frictions relevant to the packet's vendor-rent deduction.
AI-specific rents, adopter productivity or adopter operating margin.
c1bc185061d59d…01b36b7b13b6fc47c33307…803baf0f06 · Mechanical receipt
Structural completeness, deterministic lineage and release-bound inputs. It does not prove historical judgment, specialist acceptance, customer suitability or publication authority.
f65dd3e9e30c24…5bf577611b28e48eac4f27…64ada6a4d05df0e2eb6e94…aa48396bHuman review boundary
The software can confirm structure and provenance. It cannot confer subject-matter judgment, methodological approval or decision authority.
Audit appendix
Metadata, exact pointers and research paraphrases only. Full source documents are not redistributed or represented as locally snapshotted. The register is retained as audit depth; it is not needed to understand the packet outcome.
Nicholas Crafts · 2004
Steam's aggregate productivity contribution was delayed and rose after complementary capital and applications diffused.
A transferable numerical lag or uplift for enterprise AI.
22bf8d4914bff7…17b580e2Jeremy Atack, Fred Bateman and Robert A. Margo · 2006
Steam adoption was associated with establishment scale and labour-productivity change rather than a stand-alone machine effect.
That factory organization originated with steam or that all adopters captured equal returns.
819e9aa8e583a4…a129c88ePaul A. David · 1990
Aggregate productivity effects arrived after electric power diffused and factories reorganized around its capabilities.
A fixed forty-year wait for individual AI adopters.
01c66aa0fecf05…0340236fWarren D. Devine Jr. · 1983
Unit drive changed the useful organization of machinery, space and work beyond replacing the prime mover.
That electricity alone caused the measured gains.
781a5025d51a0d…c2075549Alexander J. Field · 1992
Telegraphy created value through standardized information, coordination and management practices as well as transmission speed.
A clean comparable productivity coefficient for AI deployment.
84c7ddafa8a755…0bd568a0Alfred Vail · 1914
Early telegraph operation depended on a human and procedural system around the wire.
Modern enterprise productivity or profit effects.
713ef760b98633…8f337e2fGisela Rua · 2014
Containerization diffused through interacting adoption decisions, infrastructure and network effects rather than the box alone.
That one adoption date describes all ports, carriers or firms.
eb0fa61b37f15d…f864d1eaDaniel M. Bernhofen, Zouheir El-Sahli and Richard Kneller · 2013
Standardized intermodal infrastructure had large trade effects in the studied setting.
Transfer of the paper's model-specific trade coefficient to AI productivity.
2191a4eb70d3d8…77a9978fErik Brynjolfsson and Lorin M. Hitt · 2003
Measured contributions were materially larger over five-to-seven years than over one year in the studied large-firm panel.
A five-to-seven-year numeric forecast for current enterprise AI.
b88d95bd077030…9b43eb4eErik Brynjolfsson, Lorin M. Hitt and Shinkyu Yang · 2002
Teams, distributed decision rights, training and process change were associated with disproportionate value alongside computing investment.
That hardware spending itself generated the observed market-value association.
064805983f4ece…2d87f29aLorin M. Hitt, D. J. Wu and Xiaoge Zhou · 2002
ERP adopters showed productivity differences, but timing, endogenous adoption and post-implementation evidence limited causal interpretation.
A guaranteed adopter return or a stable post-implementation uplift.
1e01d5ad7403dd…6a36a0a6Andrew McAfee · 2002
One enterprise-system deployment produced an immediate performance dip followed by learning and later improvement.
That all enterprise deployments follow the same curve.
70d5e5b4e59561…1b581648Erik Brynjolfsson and Michael D. Smith · 2000
Lower online prices and menu costs coexisted with dispersion and branded-retailer advantage in the studied categories.
Total adopter profitability or economy-wide productivity.
4d385d39ef88e9…fa53415dStephen D. Oliner and Daniel E. Sichel · 2000
Visible early e-commerce adoption could coexist with small measured macro productivity effects and weak retailer profitability.
A precise counterfactual efficiency gain from early e-commerce.
a492790a101e20…b476ce4aJames Brand, Mert Demirer, Harrison Finucane and David Kreps · 2025
Cloud compute efficiency improved through learning, stabilized over several years and remained heterogeneous across firms and divisions.
Labour productivity, revenue or operating-margin effects.
354804a6e271ef…6d4f0b9bUK Competition and Markets Authority · 2025
Provider concentration, switching barriers and software ties can affect where cloud rents accrue.
Adopter productivity or enterprise-AI market structure.
38cce2e0be825b…abf0263fDaron Acemoglu and Pascual Restrepo · 2017
Robot exposure can produce distributional effects on employment and wages even when automation raises output elsewhere.
A universal firm-level productivity or employment effect.
30dfa848b18fb8…57e03f91Georg Graetz and Guy Michaels · 2018
Industrial robots were associated with productivity and price effects alongside changes in labour composition in the studied economies.
That aggregate gains equal adopter margin capture or worker welfare.
db45b27df607eb…da40b283Judith Bachant and John McDermott · 1984
XCON became useful before full accuracy, retained human checks and required continuing knowledge-base maintenance.
An independent causal productivity effect or general-assistant capability.
fcbfdd9dbac7a4…ef495cbeVirginia E. Barker, Dennis E. O'Connor, Judith Bachant and Elliot Soloway · 1989
The configuration-system family paired useful returns with volatile rules, long developer ramp-up and a substantial support organization.
An audited XCON-only return or a general enterprise-AI business case.
9c0f3f9ef486a5…e72b1596Prabhu L. Pingali · 2012
Green Revolution outcomes depended on a package of varieties, inputs, irrigation, research, policy and market access.
A pooled numerical effect for digital or enterprise-AI adoption.
c0990b21b0a93c…c9b94262Robert E. Evenson and Douglas Gollin · 2003
Agricultural innovation affected productivity and prices through location-specific diffusion and complementary inputs.
Direct comparability to cognitive software or adopter operating margins.
9175e56970d7bf…fdab91a4Toyota Motor Corporation · 2026
Toyota describes production as an operating system combining flow, problem exposure and human response rather than isolated tools.
Independent productivity, welfare or transfer estimates.
5110b8033d5cc1…7d5ff8faY. Sugimori, K. Kusunoki, F. Cho and S. Uchikawa · 1977
The production system linked technical flow controls to operating rules and workforce practice.
A universal productivity effect or uniformly positive worker outcome.
1502ecc4101d50…048f32af