The Dynamo and the Algorithm: Why Industrial AI will arrive at the speed of complements, and why boards are right to be cautious

John H Howard, 28 July 2026 [1]

This working paper is one of a series prepared by the Acton Institute for Policy Research and Innovation to inform the UTS research project Turning AI into Productivity: The Role of Innovation Ecosystems, supported by the Google Foundation. The project's final report will be launched in mid-September 2026.

Summary

Electricity took roughly forty years from Edison's Pearl Street station to measurable productivity gains in American manufacturing. The gains arrived only when factories were rebuilt around workflow rather than around their old power systems. Adoption also waited on the institutions that made electricity trustworthy: codes, inspection and insurance standards were a precondition of mass adoption, not an afterthought.

Industrial AI is at the line-shaft stage. Many firms appear to be applying general-purpose models onto unchanged processes, and the management, skills, data and process complements are accumulating slowly.

The July 2026 intrusion at Hugging Face, carried out end-to-end by an autonomous AI agent, reveals a further missing complement in security assurance: the AI supply chain has become an attack route, and the tools of defence can fail defenders when needed most.

The 2008 financial crisis also shows how systemic trust behaves when assurance fails. Opacity in financial market counterparty chains allowed confidence to withdraw all at once in 2008, and the structure of today's AI supply chains resembles the chains of exposure that transmitted that contagion.

Business owners, CEOs and boards therefore have every right to be deliberate about commitment to AI. Their caution is rational and, on the historical evidence, well calibrated to how general purpose technologies actually diffuse.

History also records what happened to firms that mistook caution for abstention: the factories that prospered ran bounded experiments and built capability as the complements matured, while the ones that waited met electrified competitors as displacement rather than transition.

The task for policy is to compress the complement accumulation phase through compute access with assurance conditions attached, vetted tools for smaller firms, domestically developed small models, and place-based ecosystems that assemble the complements no single firm can build alone.

The proponents' clock and the historians' clock

Every general purpose technology arrives with two clocks running. The proponents' clock measures capability: benchmark results, model releases, demonstrations of what the technology can do under favourable conditions. The historians' clock measures incorporation: the slower process by which firms, workforces and institutions reorganise themselves so that capability becomes productivity.

The two chronometers can diverge for decades. Paul David's celebrated study of the dynamo was written to explain exactly such a divergence: why, in the late 1980s, computers were everywhere except in the productivity statistics. His answer came from the history of electrification, and it bears directly on the question now confronting Australian boards and policymakers about Industrial AI (David, 1990).

The term Industrial AI is itself recent, entering general use in the engineering literature of the late 2010s through the work of Jay Lee and colleagues on Industry 4.0 manufacturing systems, before Lee's 2020 book consolidated it as the name for the disciplined application of AI to industrial operations with measurable and repeatable performance (Lee et al., 2018; Lee, 2020).

Lee's founding definition tied the discipline to measurable performance and trustworthiness. The discussion that follows can be read as an account of what honouring that requirement involves.

What actually slowed electrification

Edison's Pearl Street station began supplying current in 1882. By the turn of the century, on some estimates, electricity accounted for less than five per cent of mechanical drive in American factories, and its measurable contribution to manufacturing productivity did not arrive until the 1920s, by which time around 80 percent of factory drive had been electrified (David, 1990; Devine, 1983).

The delay arose from the way the technology was first used rather than from any deficiency in the technology itself. Factory owners replaced their steam engines with large electric motors and connected them to the existing system of line shafts, belts and pulleys that distributed power through the building. The power source changed; the organisation of production did not.

Warren Devine's reconstruction of this period shows how little this substitution achieved, since the constraints on output lay in the shafting system itself, with its friction losses, its rigid machine placement and its habit of stopping the whole floor when one component failed (Devine, 1983).

The productive breakthrough came with the unit drive, the fitting of a small motor to each machine. The unit drive dissolved the constraint that machines be arranged along the shafts and allowed factories to be laid out around the flow of materials and work. Single-storey plants with overhead cranes, flexible layouts and no machine-level control became possible, and with them the modern production line. Ford's Highland Park plant is the emblematic case, designed around electric drive rather than adapted to it.

None of this could be purchased, and all of it had to be learned. Factory redesign required capital to be scrapped before it was worn out, which owners resisted. New management methods were needed to run flow-based production, and a generation of managers whose expertise was embedded in the old system had to retire or be retrained. An electrical trades workforce had to be trained at scale.

The complements to electricity were organisational, managerial and human, and they accumulated at the speed of organisations, careers and institutions rather than at the speed of invention (David & Wright, 1999). Above all, the technology had to be trusted if scarce resources were to be allocated, and assessed on a business case covering cost, risk and return.

Trust as infrastructure

Early electricity was dangerous: it caused fires and killed people, including in well-publicised urban accidents that fed genuine public alarm. Fire insurers, facing losses from a technology their actuarial tables did not comprehend, might reasonably have priced electrification out of existence. In other words, the technology was not trusted.

The early response to the trust deficit was the construction of an assurance layer. Underwriters Laboratories was established in the 1890s to test electrical equipment for the insurance industry. The first National Electrical Code followed in 1897, and municipal inspection regimes, licensing of electricians and insurance standards grew around them. These institutions did not slow adoption; they enabled it, by converting a hazardous novelty into a governable input that a prudent owner could trust and commit to.

Trust in a general-purpose technology is not a sentiment that arrives on its own but infrastructure that has to be built, and mass adoption tends to follow its construction rather than precede it. A board that declined to electrify in 1885 was exercising sound judgement and prudent decision-making. A board that declined on the same grounds in 1915, after the assurance layer existed and competitors were rebuilding around unit drive, was actually making a different and more dangerous decision.

Two dimensions of trust can be identified from this analysis.

• Institutional trust: the codes, inspectors. and testing laboratories that provided the confidence a prudent owner could place in a certified artefact.

• Relational trust: the accumulated interactions and connections that grew between people and organisations through repeated dealing and reputation.

The distinction between the two may matter more for AI than it did for electricity as we think about the relationships between people and between machines themselves, and between machines that occur with agentic AI.

The line-shaft stage of Industrial AI

The parallel with the present is close enough to be uncomfortable. Most firms adopting AI today are at the line-shaft stage: a general-purpose model has been connected to unchanged production systems, workflows and management structures. Pilots produce impressive demonstrations and modest results because the constraints on performance are in the surrounding organisation rather than in the model.

The research literature has begun to formalise the pattern. Brynjolfsson, Rock and Syverson describe a productivity J-curve for general-purpose technologies, in which measured productivity may stagnate or decline during the years in which firms accumulate intangible complementary capital: the process redesign, data foundations, skills and management practice that conventional statistics do not record. The investment is real; the statistics simply cannot see it until it begins to pay (Brynjolfsson, Rock & Syverson, 2021).

Our own published analysis finds the same structure in the field. The binding constraint on Industrial AI in Australian firms is a management gap rather than a technology gap; value is migrating toward the firms that hold the complements rather than the firms that hold the models; and the productive frontier currently lies in assistive uses of AI, because firms cannot so far govern autonomous agents in operational systems at industrial standards of reliability (Howard, 2026a).

The parallel has one limitation. AI diffuses through software and existing devices, so the physical retooling constraint that slowed unit drive is weaker this time, and the historians' clock may run somewhat faster as a result. It is unlikely to collapse, because the binding constraints on electrification were never really physical either; they were managerial, institutional and relational, and those categories of constraint move no faster in 2026 than they did in 1910.

A new complement revealed

On 16 July 2026, Hugging Face, the world's largest repository of open AI models and datasets, disclosed an intrusion into part of its production infrastructure carried out end to end by an autonomous AI agent system. The attack entered through a malicious dataset in the platform's own data processing pipeline, harvested credentials, and ran for a weekend across thousands of individual actions before being contained (Hugging Face, 2026).

Two features of the incident bear on the diffusion argument.

• The AI supply chain on which nearly every adopting firm depends has become an attack route, compromised at its most trusted point.

• The defenders turned to hosted frontier models to analyse the attack, and the models refused, because their safety guardrails could not distinguish an incident responder from an attacker. The forensic work was completed with an open-weight model running on the defender's own infrastructure (Hugging Face, 2026).

Read in the vocabulary of trust rather than security, the incident records two failures of calibration. The agent pipeline extended trust to a dataset that betrayed it and a breakdown of trust between machines. The hosted models then withheld trust from a legitimate human defender, a breakdown of trust between machine and human at the moment it mattered most.

Electrification never posed either question. The dynamo asked to be trusted as an artefact, for its reliability, and the institutions of the 1890s could therefore confine themselves to certifying equipment. An agentic system asks to be trusted as something closer to a counterparty, exercising delegated discretion across thousands of actions, and its assurance institutions may need to certify conduct as well.

In the Acton Institute for Innovation’s vocabulary, security assurance has joined the list of complements: a firm cannot turn compute, models, and data into safe, productive use without the capacity to verify what it is running, detect compromise, and respond with tools that will operate when needed.

Few small and medium-sized firms have that capacity, and the institutions that could supply it on their behalf, the testing laboratories and wiring codes of the AI era, are only beginning to be built.

Have we reached a situation that resembles electricity before the assurance layer: a technology of evident power, a documented capacity to cause harm, and an institutional vacuum where trust should be.

When trust withdraws: the lesson of 2008

The behaviour of systemic trust under stress has been studied most closely in finance, and the global financial crisis offers a cautionary companion to the electrification story. Before 2008, derivatives had built long chains of counterparty exposure through which risk moved invisibly. Securitisation made the location of that risk untraceable, so when doubt arrived it attached to everyone at once (Financial Crisis Inquiry Commission, 2011).

The defining feature of the crisis was the manner in which trust failed. Trust did not decay gradually in proportion to losses; it withdrew abruptly across the whole system, freezing interbank lending within days. Trust between institutions behaved less like a price and more like a characteristic of the network, stable until the system crossed a threshold and then gone (Haldane & May, 2011).

Agentic AI may be assembling a similar structure of exposure. Models are fine-tuned from other models, trained on synthetic data of uncertain origin, and increasingly invoke agents whose provenance the deploying firm cannot inspect. The Hugging Face intrusion showed compromise arriving through the chain's most trusted node, much as the highest rated tranches carried the hidden risk in 2008. Confidence had concentrated exactly where verification was weakest.

The ratings agencies supply a further warning, because they were the assurance layer of structured finance and they failed. They miscarried through conflicted incentives and models that could not comprehend the instruments they certified.

A badly built assurance layer may be worse than none, since it licenses exposure that prudence would otherwise refuse. Benchmarks, model cards and vendor safety claims deserve scrutiny on the same grounds.

Concentration and speed complete the parallel. Adoption is converging on a handful of frontier models and platforms, creating the shared vulnerability of a monoculture in which a single flaw reaches every adopter at once, and agents interact at machine speed, a dynamic the 2010 financial crash previewed in miniature.

There is, however, a significant difference: AI exposure today is largely operational rather than leveraged, so a collapse of trust would more plausibly produce an abrupt adoption winter than a cascade of insolvencies, although a debt-financed data centre build-out may be narrowing that difference.

The asymmetry carries the lesson for both boards and governments. Trust of this kind is slow to build, instant to destroy, and after 2008 it was restored only through institutions: deposit guarantees, central bank liquidity and a rebuilt regulatory architecture.

AI has no lender of last resort, and a serious agent-caused failure at a major firm could deplete trust for every adopter at once, in the manner of a Lehman moment, while the assurance institutions remain unbuilt.

The management dilemma: caution without abstention

Against this background, the caution of,  business owners, CEOs and boards in adoption of AI is a rational response to genuine conditions, and not a failure of imagination. This caution can be ascribed to:

• Vendor timelines that echo the dynamo's early promoters

• An assurance layer that does not yet exist

• A supply chain that has just been demonstrably compromised

• A productivity literature that predicts years of intangible investment before returns become visible.

Directors who discount confident forecasts and demand evidence of governability are, in these circumstances, doing exactly what their role requires.

The historical record does, nonetheless, provide a discipline of how caution should be exercised. The firms that prospered through electrification were not the ones that waited for the technology to become settled, or the ones that electrified everything at once. They were the ones that ran bounded experiments, trained their people, redesigned processes incrementally and sequenced commitment as the complements matured.

Deferral looked prudent for years, until the cost structure of electrified competitors made the old plants unviable, and the adjustment then arrived as displacement rather than transition.

The defensible board position in 2026 is therefore caution about mode, not abstention. In practice this may mean:

• Assistive deployment before agentic autonomy

• Bounded, well-instrumented experiments with clear learning objectives

• Investment in data foundations and management capability that retains value under any model future

• Assurance requirements, including provenance of models and data, imposed on suppliers before deployment

• Board-level attention to the upside as well as the risk register.

As this series has argued previously, most current AI governance guidance is all brakes and no steering; the electrification parallel explains why both controls are needed at once (Howard, 2026b).

Compressing the curve: the task for policy

If Industrial AI diffuses at the speed of complements, the object of policy comes into focus. Government cannot do much to accelerate the technology and does not need to.

What Government can do is compress the complement accumulation phase, as public institutions eventually did for electricity through codes, inspection, standards and the education of an electrical workforce.

The following policy issues emerge from this discussion:

• Conditions negotiated on data centre approvals should attach assurance obligations, provenance vetting, incident response arrangements and data residency, to the compute access that Australian firms will take up.

• A vetted catalogue of models and tools can supply smaller firms with the assurance they cannot produce themselves.

• Domestically developed Small Language Models, run within a firm's own environment, offer both a capability instrument and the practical answer to the defensive failures the Hugging Face incident exposed.

• Place-based innovation ecosystems remain the setting in which the managerial and workforce complements are most efficiently assembled, because they are learned between organisations rather than inside them. The ecosystems literature already suggests that inter-organisational trust is itself a complement, and possibly the one complement an assurance layer cannot supply, because it grows through repeated dealing in place rather than through certification.

Framed this way, board caution and public policy should stop pulling in opposite directions. The board's hesitation identifies the complements that are missing; the policy program exists to lower the cost of supplying them.

A government that builds the assurance layer is not subsidising a technology; it is doing for the algorithm what the codes, the inspectors and the testing laboratories did for the dynamo.

Conclusion: the speed of complements

The dynamo took forty years to remake manufacturing, and the delay was located almost entirely outside the technology. It was located in the factories, management capacity, the trades and the institutions of trust. All of which had to be rebuilt around it. Industrial AI seems to be following the same path, perhaps more quickly, though on the historians' clock rather than the proponents' timepiece.

Business owners, CEOs and boards are entitled to be cautious, and recent events have vindicated it. The experience of 2008 adds its own warning about the alternative to building assurance, since trust of the systemic kind withdraws all at once when it fails, and returns only through institutions built after the damage has occurred. Two practical questions emerge:

• For businesses: whether caution is being used to sequence commitment or to avoid it.

• For Governments: whether Australia can wait for the complements to accumulate at their own pace, or set about building, deliberately and early, the assurance and access institutions that history says mass adoption waits for, and that the financial crisis says cost far less to build before trust fails than after.

References

Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333-372.

Bresnahan, T. F., & Trajtenberg, M. (1995). General purpose technologies: Engines of growth? Journal of Econometrics, 65(1), 83-108.

David, P. A. (1990). The dynamo and the computer: An historical perspective on the modern productivity paradox. American Economic Review, 80(2), 355-361.

David, P. A., & Wright, G. (1999). General purpose technologies and surges in productivity: Historical reflections on the future of the ICT revolution. University of Oxford Discussion Papers in Economic and Social History, No. 31.

Devine, W. D. (1983). From shafts to wires: Historical perspective on electrification. Journal of Economic History, 43(2), 347-372.

Financial Crisis Inquiry Commission. (2011). The financial crisis inquiry report: Final report of the National Commission on the Causes of the Financial and Economic Crisis in the United States. U.S. Government Publishing Office.

Green, R., & Agarwal, R. (2009). Management matters in Australia: Just how productive are we? Department of Innovation, Industry, Science and Research.

Haldane, A. G., & May, R. M. (2011). Systemic risk in banking ecosystems. Nature, 469(7330), 351-355.

Howard, J. H. (2026a). Making Sense of AI in 2026: A framework for policy and practice. Decision making in a time of opportunity, uncertainty and risk. Acton Institute Publishing.

Howard, J. H. (2026b). Governing the Upside: Why board guidance on AI is mostly brakes and limited steering, Acton Institute for Policy Research and Innovation, 16 July, https://www.actoninstitute.au/post/governing-the-upside-why-board-guidance-on-ai-is-mostly-brakes-and-limited-steering

Hugging Face. (2026, July 16). Security incident disclosure - July 2026. https://huggingface.co/blog/security-incident-july-2026

Lee, J. (2020). Industrial AI: Applications with sustainable performance. Springer.

Lee, J., Davari, H., Singh, J., & Pandhare, V. (2018). Industrial artificial intelligence for Industry 4.0-based manufacturing systems. Manufacturing Letters, 18, 20-23.

[1] Dr John H. Howard is Executive Director of the Acton Institute for Policy Research and Innovation, Sydney, and Research Director of the UTS research project Turning AI into Productivity: The Role of Innovation Ecosystems, supported by the Google Foundation. The project is led by Emeritus Professor Roy Green AM, Special Innovation Adviser at the University of Technology Sydney, who reviewed drafts of this paper and contributed advice on its argument and policy framing.

AI declaration

This Insight was prepared with the assistance of AI tools under the direction of the author. Details of the July 2026 Hugging Face incident were checked against the company's primary disclosure. Historical figures on the pace of factory electrification are drawn from the cited literature and expressed with appropriate qualification.