As AI accelerates industrial restructuring, can the adaptive capacity of the economy and society keep up?

By | Duoda NewMedia, Authors | Chen Ting, Ryan Li

Artificial intelligence is changing a long-standing economic question: after a major technological innovation emerges, why does the entire economic system often take a long time to complete the corresponding restructuring?

At the same time, the rise of financial AI Agents brings new questions: when large numbers of automated systems begin making decisions independently or collaboratively, are the original boundaries of responsibility, governance approaches, and risk transmission mechanisms still effective?

This means that what we need to pay attention to may not only be the speed gap between technological development and institutional adjustment, but the new relationship forming among technological change, system adaptation, and risk transmission.

I. The History of Electrification: Why Is a Technological Revolution Not Equal to a Productivity Revolution?

In September 2026, Kevin Bryan, a professor at the Rotman School of Management at the University of Toronto, published an article titled "AI Is a Technology, and Driving Its Diffusion Is a Management Problem."

Using electrification and container shipping as examples, he explained that between technological innovation and the full realization of economic value, there is a complex process of organizational and industrial restructuring.

In the age of steam power, factories usually organized production around a central power unit. Machines were connected through drive shafts and belts, and factory building structures, equipment layouts, and production processes were all constrained by the power system.

After the electric motor appeared, companies could replace steam power with electricity, but this replacement itself could not unlock the full value of electrification.

Real productivity improvement required rearranging machine positions, changing factory layouts, and adjusting production sequences and the division of labor. Independent electric drive enabled factories to organize production according to work processes rather than power transmission methods, and also created conditions for new large-scale production methods.

Bryan pointed out that from the commissioning of Edison's first power station to the electrification of about half of U.S. manufacturing, about 40 years passed.

Container shipping is a similar example.

The standardized container itself is not a particularly complex technology, but to truly transform the global logistics system, ports, railways, warehousing, ships, and related enterprises needed to adjust together. Technological invention was only the starting point; reorganizing industrial activities around the technology was the more difficult process.

These two historical cases reveal an important pattern:

The economic value of major technologies comes not only from technological capability itself, but also from complementary innovation, organizational adjustment, and industrial ecosystem restructuring around the technology.

When enterprises deploy advanced models, it does not mean that existing workflows, job configurations, decision-making mechanisms, and business models will automatically translate into higher productivity.

With the development of AI Agent capabilities, some digital work that previously required multiple people and multiple departments to collaborate may receive a higher degree of automated support.

One is the speed at which new technologies are applied within enterprises; the other is the speed at which enterprises redesign production and organizational methods around new technologies.

The latter deserves particular attention.

In the past, for an enterprise to carry out complex business transformation, a large amount of manpower was needed to complete process analysis, system development, testing, training, and organizational coordination. If AI can take on part of this work, it may reduce restructuring costs and shorten the implementation cycle of certain stages.

This capability may even cross enterprise boundaries and affect the division of labor among software vendors, service providers, platforms, and customers.

However, this is still a judgment that needs to be verified.

Industries with a high degree of digitalization may change faster, while physical infrastructure, regulatory approvals, cross-institutional collaboration, and worker skill transitions remain subject to different constraints.

The speed at which the industrial structure is changed does not mean that a new economic and social order can form stably at the same speed.

III. Is the Industrial Ecosystem Really Restructuring at an Accelerating Pace?

More meaningful evidence should come from actual changes in organizational structures, industrial division of labor, market entry points, and value distribution relationships.

In the information services field, such changes have already begun to appear.

A 2025 Pew Research Center study of Google search behavior found that among searches with AI-generated summaries, the proportion of users clicking traditional search results was 8%; when there were no AI summaries, this proportion was 15%.

As users increasingly obtain answers directly within search platforms, the way traditional content providers rely on search engines to gain traffic may be affected.

This is not only a question of search efficiency, but also involves information distribution entry points, user relationships, and value distribution between platforms and content producers.

At the level of corporate organization, the fintech company Klarna provides another window for observation.

According to its public disclosures, in 2025 its AI assistant handled about 80% of customer service chats, and the company estimated that it took on a workload equivalent to more than 700 full-time customer service staff.

But this change is still mainly occurring within the enterprise and cannot directly prove that the entire financial services industrial ecosystem has completed restructuring.

Research in the software development field provides a more valuable comparison.

In 2026, a study released by the NBER analyzed more than 500,000 GitHub developers and data on their use of AI tools.

The study found that autonomous coding Agents increased code commit activity cumulatively by about 240%, but this gain weakened significantly in subsequent production stages: the number of projects increased by about 80%, while actual software releases increased by about 30%.

The study also found that in the major software markets observed, the number of new applications increased significantly, but overall usage did not grow in tandem.

On the other hand, bottlenecks in subsequent stages still exist, indicating that the distance between technological capability and final economic value has not disappeared.

IV. Financial AI Agents: When Organizational Restructuring Touches the Boundaries of Responsibility and Control

Traditional bank digitalization has mainly revolved around financial institutions' own business systems and service channels.

However, with the development of third-party AI Agents, customers may increasingly obtain financial information, compare products, manage funds, and even authorize transactions through personal intelligent agents.

Banks may also internally use multiple agents to separately handle customer identification, risk analysis, product matching, and transaction processing.

This change is not merely the addition of a service tool; it may alter the interaction relationship between financial institutions and customers.

When a complete business process is jointly completed by multiple Agents, who is responsible for the final result?

When model capabilities are upgraded, tool permissions are expanded, or the collaboration methods among Agents change, do the original authorizations and risk assessments still apply?

If an erroneous transaction occurs between a customer's intelligent agent and a bank system, how should responsibility be allocated among the bank, the technology provider, and the customer?

These issues have become real governance challenges that the financial industry is currently facing.

On October 19, 2026, at the Bund Summit roundtable on financial AI governance, Liu Jun, President of the Industrial and Commercial Bank of China, pointed out that commercial banks have moved from computer system construction into digital transformation, and that digital transformation has not yet been fully completed while AI transformation has already arrived.

This reveals another kind of temporal mismatch: the previous round of organizational and technological transformation has not yet ended, while the requirements of the next round of transformation have already begun to pile up.

Arthur Yuen, Deputy Chief Executive of the Hong Kong Monetary Authority, emphasized that as Agent applications expand, governance needs to cover the complete lifecycle of model selection, deployment, updating, incident handling, and change management.

When different intelligent agents are responsible only for local tasks and human oversight covers only some stages, overall responsibility may become blurred.

More importantly, when technological capabilities continue to change, can financial institutions adjust their business processes, decision-making authority, responsibility mechanisms, and risk control methods in tandem?

At the Bund Summit roundtable, Lorenzo Bini Smaghi, Chairman of the European 50 Group, proposed that discussions of financial AI cannot remain at the local perspective of a single institution; they also need to consider the overall market behavior that emerges after large numbers of intelligent agents interact with one another.

Suppose the agents of multiple financial institutions use similar data, models, and risk judgment standards.

When adverse signals appear in the market, they may form similar judgments within a relatively short period and recommend or execute similar transactions.

If these actions further affect market prices and liquidity and in turn trigger the risk control measures of other intelligent agents, a feedback loop may form.

Herding behavior and procyclicality in financial markets are not new phenomena.

Different models, trading limits, human intervention, and liquidity buffers can all reduce the risks brought by synchronized actions.

Value realization at the enterprise level and risk control at the financial system level must be observed separately.

But reality may be more complex.

On October 15, 2026, the Financial Stability Board released the final report on Robust Practices for the Responsible Adoption of AI.

The report sets out 12 robust practices, which explicitly include organizational adaptability, continuous risk assessment, performance monitoring, effective human oversight, and management of third-party AI dependency risk.

However, the existence of institutional documents does not mean that governance is already effective.

If, after major changes to a model, an enterprise cannot promptly reassess risks; if anomalies have been detected but there is no clear responsibility for decision-making and execution; if control measures have been implemented but their effects cannot be verified, then governance may still have real gaps.

Therefore, organizational adaptability needs to be implemented as a set of mechanisms that can operate and be tested:

change identification → impact assessment → decision adjustment → action execution → result verification.

It must answer not only who is responsible, but also who has the authority to adjust, when action must be taken, what controls to adopt, and how to determine that the problem has been resolved.

This capability also needs to be continuously updated.

Research on the diffusion of general-purpose technologies has long pointed out that complementary organizational change exists between technological innovation and economic value realization.

Research on organizational dynamic capabilities has also long focused on how enterprises perceive environmental changes, restructure resources, and adjust themselves.

Financial stability research has already established an analytical foundation regarding common risk exposures, procyclicality, and systemic risk transmission.

But rapid AI iteration, AI-assisted organizational restructuring, and autonomous interaction among Agents may change the operating conditions of the above mechanisms.

First, structural change.

Second, system adaptation.

Third, risk transmission.

When the speeds of change in different systems are inconsistent, how is local imbalance buffered, absorbed, or further diffused through financial, industrial, and social relationships?

There is no simple linear relationship among these three processes.

Accelerating technological change does not necessarily mean that the adaptation gap will inevitably widen.

Improved enterprise adaptability does not necessarily mean that systemic risk will decline in tandem.

Rising industrial productivity certainly does not mean that workers, consumers, and society as a whole will necessarily share the gains at the same speed.

For example, a bank may rapidly complete the automation of customer service, but customers may still lack adequate mechanisms for appeals and error correction in complex decisions.

At the social level, enterprises may complete job adjustments relatively quickly, but it may take longer for workers to relearn, change careers, and restore income stability.

More importantly, it is about understanding: how technology changes structures, how systems adjust themselves, and how the interactions among these changes affect value realization and risk evolution.

Conclusion: What truly needs to be accelerated is not only technology

The history of electrification tells us that the economic value of a major technology requires a production system and organizational restructuring built around that technology.

But the current evidence also shows that improvements in the efficiency of individual tasks do not necessarily translate into proportional growth in final products, market demand, and the overall value of the industry.

It may accelerate certain adjustments, but it may also make changes across different systems more uneven.

Rather, it is: