The most convincing argument for artificial intelligence in the workplace is also the reason dependency deserves more attention.
When AI works well, people can do more with less effort. New employees can perform tasks that once required months of experience. Experienced workers can move faster through routine analysis, drafting and information retrieval. Companies can standardize work that previously depended on individual expertise, reduce costs and handle larger volumes without adding the same number of employees.
Those are not hypothetical benefits. They are already visible in empirical research. The difficult question begins later, after organizations have had enough time to redesign themselves around those gains.
If an AI system becomes unavailable for a few hours, the answer is usually obvious: wait, switch tools or do the work manually. But what happens after several years of adoption, when staffing levels, training programs, software systems and internal procedures have all evolved around the assumption that AI assistance will normally be there?
At that point, losing AI is no longer the same thing as losing a tool. The organization itself may have changed.
This does not mean companies are already forgetting how to work without AI at scale. There is not yet evidence strong enough to support such a broad claim. What does exist is a growing body of research showing several pieces of the mechanism: AI can substantially raise productivity, dependence on a small number of external providers is increasing in some sectors, sustained automation can weaken skills that are no longer practiced, and organizational knowledge can disappear when it is no longer retained in people, routines or structures.
The question is what happens when those trends meet.
Dependency begins with productivity
There would be little reason to worry about AI dependency if AI were not useful.
One of the strongest real-world studies comes from customer support. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered introduction of a generative-AI assistant among 5,172 support agents. Published in The Quarterly Journal of Economics in 2025, the study found that access to the system increased issues resolved per hour by 15 percent on average. The gains were substantially larger among less experienced and lower-skilled workers, and the authors found evidence that AI was helping transmit practices associated with more capable agents. Newer workers also moved down the experience curve faster.
That result is important because it shows why organizations have a rational incentive to reorganize work around AI rather than merely place an optional chatbot beside existing employees. If a less experienced worker can perform closer to the level of an experienced colleague, the economic value is not confined to saving a few minutes. It can change hiring, training, supervision and staffing decisions.
A separate preregistered experiment involving 758 knowledge workers from Boston Consulting Group found a similarly powerful but more complicated effect. For 18 consulting tasks that fell within what the researchers called AI’s “jagged technological frontier,” workers with GPT-4 completed 12.2 percent more tasks and worked 25.1 percent faster on average while producing higher-quality results. But on a complex task deliberately selected outside AI’s capabilities, workers using AI were 19 percent less likely to reach the correct answer than workers without it. The study, published in Organization Science in 2026, therefore offers both the business case for adoption and a warning against assuming that assistance is uniformly beneficial.
The productivity gains are real enough that the relevant organizational question is no longer whether AI will enter knowledge work. It is how deeply companies will redesign work around it.
Using AI is not yet the same as depending on it
The distinction between adoption and dependency is easy to miss because they can look identical during normal operations.
A company may use an AI assistant hundreds of times a day and still remain resilient if employees retain the relevant skills, the underlying processes are documented and alternative ways of completing critical work remain practical. Another company may use AI for fewer tasks but become much more vulnerable because one external model sits inside a process that nobody can easily replace.
Financial services provide one of the clearest current examples because regulators have already begun measuring the dependency.
A 2024 survey by the Bank of England and Financial Conduct Authority received responses from 118 regulated firms. Seventy-five percent said they were already using AI, and another 10 percent planned to do so within three years. One third of reported AI use cases were third-party implementations, up from 17 percent in the regulators’ 2022 survey. The top three model providers accounted for 44 percent of the third-party model providers named by respondents, while 46 percent of firms using or planning to use AI said they had only a “partial understanding” of the AI technologies deployed in their operations. Respondents expected critical third-party dependency to be among the systemic risks increasing most over the following three years.
Those numbers apply to UK financial services, not to companies in general, and they should not be extrapolated across the economy. They nevertheless demonstrate something concrete: in at least one heavily regulated sector, AI adoption is already being accompanied by greater reliance on external providers and concern about concentration.
Dependency therefore has at least two dimensions. One is technological: whether a firm relies on a particular provider, model, cloud platform or data source. The other is organizational: whether the firm still retains the people, knowledge and processes required to perform essential work when that technology is unavailable.
The second form is harder to see because it develops slowly.
Skills change when work changes
Concerns about automation-induced deskilling are much older than generative AI. Human-factors research has documented the “out-of-the-loop” problem in highly automated systems for decades: when people stop performing or closely monitoring a task, they can become less prepared to intervene when automation fails.
Generative AI extends that question from highly specialized operational environments into ordinary knowledge work.
A 2025 study presented at the ACM CHI conference surveyed 319 knowledge workers who used generative AI at least weekly and collected 936 examples of real workplace use. Participants reported that AI reduced the cognitive effort associated with many activities, while the nature of critical thinking shifted toward checking information, integrating AI outputs and supervising the task. Higher confidence in AI was associated with less reported critical-thinking effort, although the study relied on participants’ self-reports and did not demonstrate that their underlying skills had actually deteriorated.
That distinction matters. Reduced effort during a task is not the same thing as permanent deskilling.
More direct evidence exists in some specialized settings, although it remains limited. In 2025, researchers examined what happened to experienced endoscopists after routine exposure to AI-assisted polyp detection. The multicentre observational study covered four centers in Poland and compared non-AI-assisted colonoscopies performed before and after AI tools became part of regular practice. Across 1,443 standard colonoscopies, the adenoma detection rate in procedures performed without AI fell from 28.4 percent before AI exposure to 22.4 percent afterward. The study involved 19 experienced endoscopists.
The finding is striking, but its boundaries are just as important as the result. It was an observational before-and-after comparison in one clinical task, not proof that AI generally causes doctors—or workers more broadly—to lose competence. Patient characteristics and other changes over time can complicate interpretation, and replication in other domains is necessary.
A 2026 peer-reviewed systematic review in The DATA BASE for Advances in Information Systems placed such findings into a broader organizational framework. Atiya Avery, Michael Dinger and Christian Maier reviewed research across automation and AI and defined “technology-driven skill degradation” as the depreciation of still-required essential skills caused by sustained reliance on technology. Their synthesis connects skill degradation to organizational resilience because workers may be less able to improvise effective workarounds when the technology on which normal work depends is unavailable.
That paper is particularly important for the argument here because it means the basic proposition—technology dependence can erode human fallback capability—is no longer merely speculative. It is now an explicit subject of organizational research.
It also means that proposition should not be claimed as a new discovery.
The strongest counterargument is that AI can preserve skills and knowledge too
There is a danger in telling this story only in one direction.
AI does not merely remove opportunities to learn. In some settings, it can accelerate learning, distribute expert knowledge and preserve information that would otherwise remain trapped in individual workers.
The customer-support study offers unusually clear evidence. Lower-skilled and newer employees benefited the most from AI assistance, and the authors found evidence consistent with AI disseminating the practices of more productive workers. They also found indications of durable worker learning rather than pure rote dependence on suggestions. In that workplace, AI appears to have reduced rather than widened some differences in effective skill.
Research on knowledge management points in the same direction. A 2026 case study in the Journal of Knowledge Management examined a generative-AI knowledge system in an Italian manufacturing company and reported improvements in knowledge accessibility and information flows, although the evidence came from a single organization and should not be generalized widely. Other research has argued that generative AI can help codify expert knowledge, improve retrieval and transfer information more efficiently across an organization.
There is therefore no simple law saying that AI adoption inevitably produces deskilling. The effect depends on how work is redesigned.
An AI system can replace thinking, or it can support learning. It can hide the logic of a process, or it can make previously inaccessible knowledge easier to retrieve. It can reduce the need for junior workers to practice basic tasks, or it can expose those same workers to expertise that was previously difficult to obtain.
This makes organizational design more important than the technology alone.
The problem may be what the organization stops practicing
Individual deskilling is only one part of the problem. Companies themselves can forget.
Organizational-memory research long predates AI and shows that knowledge is retained in several places: people, routines, processes, technologies and formal structures. When those repositories change, organizations can lose capabilities even when no individual person has literally forgotten them.
A 2025 study of information-technology projects across 94 U.S. federal agencies found that organizational structures capable of retaining task knowledge—particularly centralized authority and standardized processes—reduced project delays associated with employee turnover. The study is about workforce turnover rather than AI, but it supports a broader principle: an organization’s ability to continue functioning depends partly on where operational knowledge has been stored.
Other empirical work on organizational forgetting has found that knowledge gains can depreciate over time and that the rate of loss varies depending on whether knowledge is embedded in people, routines or technology. Research involving automotive suppliers, for example, found measurable erosion in quality-related knowledge over time. These studies should not be treated as estimates of AI-related knowledge loss; they show only that organizational knowledge is not automatically permanent once it has been acquired.
This distinction changes the way AI dependency should be assessed.
Suppose a company once trained junior analysts to gather information, compare sources, construct a model and write an initial recommendation. After several years of AI adoption, the workflow might instead ask the employee to describe the problem, review a machine-generated analysis and approve or modify the result. The new process could be dramatically faster and produce better routine work.
But if the original process is no longer practiced, documented or taught, the organization has not simply added AI. It has moved part of its operational knowledge into the AI-mediated workflow.
The problem becomes visible only when the workflow is disrupted.
Technical recovery and organizational recovery are not the same thing
This leads to the part of the argument that remains a hypothesis rather than an established empirical result.
An AI service can return after an outage, a cloud provider can restore capacity, or a company can replace one model with another. The technical side of the disruption may therefore be temporary.
If years of organizational redesign have simultaneously weakened human skills, removed manual procedures, changed training pathways and reduced staffing around the old process, restoring the technology does not restore those lost capabilities. Rebuilding them may require retraining, hiring, process reconstruction and renewed practice.
There is not currently a reliable general estimate for how long such rebuilding would take, and no large-scale study has established a universal “AI recovery gap.” It would therefore be misleading to claim that human organizational capability necessarily takes months or years to return after an AI disruption.
The stronger and more defensible claim is narrower: technical recovery and organizational capability recovery are different processes, and an organization that allows important fallback capabilities to decay cannot assume that restoring the technology restores the fallback as well.
This is the recovery asymmetry companies should care about.
It is also why measuring AI risk only in terms of model accuracy, cybersecurity or provider uptime is incomplete. A model may be highly reliable while the organization around it becomes less resilient to the occasions when it is unavailable.
Operational resilience offers a better model than technological nostalgia
The solution is not to preserve every pre-AI job or force workers to perform inefficient manual tasks indefinitely.
That would misunderstand both the evidence and the economics. AI can substantially improve productivity and can sometimes accelerate worker learning. Deliberately refusing those benefits merely to preserve old methods would impose real costs.
Operational-resilience frameworks provide a more useful way to think about the problem.
The Basel Committee’s principles for operational resilience, written for banks rather than AI-dependent firms generally, require institutions to identify critical operations and map the people, technologies, processes, information and facilities on which those operations depend. They also call for business-continuity exercises under severe but plausible scenarios, assessment of whether critical third-party providers can be substituted and consideration of alternatives such as bringing a service back in-house.
Those requirements are specific to banking regulation, but the logic translates well to AI dependency.
A company does not need a manual duplicate of every AI-assisted workflow. It needs to know which workflows are genuinely critical, what happens when AI support disappears, which tasks can stop safely, which can be performed at lower capacity and which still require human competence.
That means asking practical questions before an outage rather than during one.
Which functions depend on a particular model or provider? Which employees can still perform essential tasks without it? How long could the organization operate in a degraded mode? Does an alternative provider actually substitute for the first one, or does it depend on the same cloud infrastructure? Are the manual procedures still documented? Have employees ever tested them? If junior workers are no longer learning a foundational task, where will the next generation of experts come from?
These are not arguments for maintaining obsolete work. They are arguments for knowing what has been made obsolete before discovering that it was still necessary.
Resilience may require preserving capabilities rather than jobs
This distinction is important because debates about AI often become trapped in the wrong unit of analysis.
A company does not necessarily need to preserve a particular headcount. It may need to preserve a capability.
Sometimes that capability can live in a small group of experienced employees. Sometimes it can be embedded in documentation, standardized processes or simulation exercises. Sometimes a smaller local model can provide enough functionality to maintain essential operations if a more powerful external service fails. In other cases, regular human practice may remain necessary because the skill itself decays when it is not exercised.
The correct design will vary by task.
This is also where the strongest pro-AI evidence should change how resilience is designed. If AI can accelerate the learning of less experienced workers, as the customer-service study suggests, then AI itself may be useful for training the human fallback rather than merely replacing it. If AI can help codify expert knowledge, organizations can use it to preserve institutional memory instead of allowing expertise to disappear when experienced workers leave.
The important distinction is whether AI expands the organization’s underlying competence or merely hides its absence during normal operations.
That is not something productivity statistics alone can tell us.
The real warning sign is invisible success
Organizations tend to notice dependency after something breaks. AI may make that especially difficult because the technology can work extremely well for long periods.
The better the system performs, the easier it becomes to stop rehearsing the old process. The more reliable the provider appears, the less urgent an alternative seems. The greater the productivity improvement, the stronger the financial pressure to remove capacity that now looks redundant.
None of these decisions is irrational in isolation.
The risk emerges from their combination.
An organization may gradually discover that a tool has become infrastructure, that infrastructure has become concentrated outside the company, and the internal capabilities that once substituted for it have atrophied. By the time that dependency becomes visible, rebuilding the fallback may be substantially harder than keeping a limited form of it alive would have been.
This should not be confused with a prediction that AI services will frequently collapse or that companies will soon become incapable of functioning without them. Evidence for such a sweeping conclusion does not exist.
The more modest conclusion is enough: as AI moves deeper into core workflows, organizations need to manage not only the reliability of the technology but the capabilities they stop maintaining because the technology is reliable.
That is a different kind of risk from hallucination, bias or cybersecurity. It belongs to organizational memory and business continuity.
The question companies should ask is not whether they can work without AI forever
Almost no modern organization could function normally without electricity, networks, cloud software or telecommunications, and resilience does not require pretending otherwise. Dependence on infrastructure is part of economic progress.
AI will probably become another layer of that infrastructure.
The relevant question is therefore not whether a company can return permanently to a pre-AI world. It is whether the organization can continue its essential operations long enough, and at sufficient quality, when part of that infrastructure becomes unavailable.
A company that uses AI extensively but knows its dependencies, retains essential competence and regularly tests degraded operations may be more resilient than a company that uses much less AI but has no idea what happens when one critical service fails.
The danger is not dependence itself.
It is unexamined dependence combined with disappearing fallback capability.
If companies understand that distinction early, AI can increase both productivity and organizational knowledge. If they do not, they may eventually discover that the easiest part of an AI failure is getting the technology back online.
AI Dependency & Resilience
This essay is part of a three-part series on what happens when machine intelligence becomes ordinary infrastructure: how energy constrains it, what organizations lose when human fallback disappears, and how scarce compute is allocated.
01 — Energy & Availability
When Society Depends on AI, Can It Withstand an Energy Crisis?
02 — Human Fallback
What Happens When Companies Forget How to Work Without AI?
03 — Compute Allocation
When AI Compute Becomes Scarce, Who Gets It First?