As AI begins to appear inside devices, office software, and everyday workflows, the management question is no longer whether a company “uses AI.” The harder question is how deeply the company uses AI, how it changes work, and how human capability is being prepared. One notable point in new data from the Federal Reserve Bank of New York is that AI adoption is spreading quickly across firms, but the immediate effect does not follow the narrative of mass layoffs. The clearer signal is that work is being partially redesigned, while the need for retraining is increasing.
This article treats AI as a work-management issue, not only a technology issue. As tools become more common, the risk is not merely that a company buys software too slowly. The risk is that teams use AI in fragmented ways, without output standards, quality controls, or a clear understanding of which parts of work should be delegated to machines and which parts still require human judgement.
What does the new data show?
In an analysis published on September 1, 2026 in Liberty Street Economics, researchers at the Federal Reserve Bank of New York reported that the share of firms using AI in their regional surveys had increased rapidly over three years. Among service firms, AI adoption reached 61% in 2026, up from 40% in 2025 and 25% in 2024. Among manufacturers, adoption reached 51% in 2026, nearly double the 26% recorded in 2025 and more than triple the 16% recorded in 2024. This is no longer a narrow experiment inside a few technology teams.
Yet investment levels remain relatively cautious. According to the same data, about three-quarters of service firms and more than 90% of manufacturers described their AI investment as minimal to modest. Only 15% of service firms said they had devoted significant resources to AI; no manufacturers in the survey reported that level. About 5% of service firms described AI as a major strategic investment.
This points to a managerial paradox: AI is spreading very quickly, but most firms are still deploying it in partial, learning-by-doing ways. AI has entered work, but it has not necessarily entered organisational design, control mechanisms, and capability systems in a complete way.
This is not only a story of immediate replacement
An important detail is that the labour effects in the New York Fed data do not confirm a short-term story of mass layoffs caused by AI. Among firms already using AI, only 4% of service firms said they had laid off workers because of AI; the figure for manufacturers was 0%. A larger share said they had hired fewer workers than they otherwise would have, but some firms also hired more workers because of AI-related needs. Specifically, 15% of service firms said they hired fewer workers, while about 13% hired more due to AI.
Therefore, if managers view AI only through the lens of “machines replacing people,” they may miss the more important shift: work is changing shape. A marketing employee may keep the same position, but the way they research customers, draft content, test messages, and prepare reports has changed. An operations employee may still own the same process, but parts of checking, summarising, and exception alerts may now be AI-supported. A job title that remains unchanged does not mean the required capability remains unchanged.
The gap is implementation capability
The New York Fed data also show that the share of workers actually using AI within adopting firms remains modest. The median share is 17% for service firms and 7% for manufacturers. In other words, many organisations “have AI,” but AI has not yet become a broad working capability across the workforce. This is an important implementation gap.
The reasons are not only about tools. Among firms that have not adopted AI, about half say their work does not lend itself to AI; roughly a quarter say AI is not good enough; more than a third cite concerns about data, security, and confidentiality; a similar share worries about accuracy and reliability; and about a third lack technical staff. These are all management issues: identifying use cases, controlling data, designing processes, setting review standards, and developing capability.
Retraining is the infrastructure of AI transformation
One positive signal is that firms have begun retraining workers. The New York Fed reported that just over a third of service firms and more than 20% of manufacturers are retraining workers to use AI. Training topics include basic AI literacy, tool-specific instruction, automation of routine tasks, prompt engineering, function-specific applications, responsible AI use, output verification, bias, security, and avoiding over-reliance.
But AI training should not be reduced to tool instruction. If employees are only taught how to write prompts, the company may end up with many individual experiments but little system-level improvement. Training needs to be tied to real work: which process is being supported, what output is acceptable, who checks the result, which errors are unacceptable, what data must not be entered into a tool, and when humans must make the final decision.
Management lens
A company that wants to use AI in customer service should not begin with the question “which chatbot should we use?” A better question is: which types of requests consume the most time, which requests can be standardised, which situations must be escalated to humans, what criteria will be used to check answers, and what customer data must not be entered into the system. Once these questions are clear, tool selection becomes meaningful.
Implications for Vietnamese businesses
For Vietnamese businesses, the data carry a practical implication: companies should not wait for a “large AI project” before preparing. AI is entering everyday work tools, so capability gaps will emerge early between teams that know how to integrate AI into workflows and teams that use AI only as a personal convenience. This gap affects not only productivity, but also decision quality, customer response speed, and organisational learning.
Over the next 6–12 months, a practical roadmap can start with three actions. First, choose several high-frequency and moderate-risk processes for AI experimentation, such as report summarisation, customer-feedback classification, content drafting, or support for basic data analysis. Second, build output standards and review rules instead of allowing every individual to use AI in their own way. Third, train by work role: managers need to frame problems and control risk, specialists need to use tools to produce better outputs, and leaders need to measure impact rather than merely count the number of tools deployed.
Conclusion
AI is spreading across firms, but the new data show that the story is not simply about immediate labour replacement. The more urgent issue is redesigning work and retraining the workforce. A company that only buys tools will have more tools. A company that redesigns how work is done, standardises outputs, controls risk, and develops human capability has a better chance of turning AI into operating capability. At this stage, the advantage belongs not to the organisation that uses AI the most, but to the organisation that can turn AI into a controlled way of working.
References
Abel, J., Deitz, R., Emanuel, N., & Montalbano, N. (2026, September 1). Businesses are using AI to transform work, not cut jobs. Liberty Street Economics, Federal Reserve Bank of New York. https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/
Microsoft. (2026, May 5). Agents, human agency, and the opportunity for every organization. Work Trend Index. https://www.microsoft.com/en-us/worklab/work-trend-index
