When humans are reduced to working as machine quality inspectors, the AI transformation has already gone astray.
Many enterprises regard manual verification as a security solution for AI implementation, only to fall into the automation paradox: the life cycle of verification positions is extremely short, and the value of human labor continues to shrink. Real transformation is not to make people fill in the shortcomings of AI, but to seize the new space opened up by technology, restructure organizations around new decisions and new businesses, and tap higher-dimensional growth opportunities.
"There are always people checking every step in the process."
This sentence is probably the most repeated phrase when enterprises discuss AI implementation. Executives always use it to explain why adopting such systems is prudent and responsible, and to reassure employees that their jobs will not be impacted. But this seemingly reassuring phrase often ends up only assigning the company's most core and valuable employees to guard tasks that are about to be taken over by machines.
Thought leaders often claim that the positioning of human workers in the future is to fill the gaps that machines cannot fill. For example, a leading venture capital firm proposed that in the era of knowledge economy, four main types of human positions will soon emerge: the person who takes final responsibility when AI outputs are untrustworthy; the person who designs the workflow that machines run; the person who maintains interpersonal cooperative relationships that intelligent agents cannot undertake; the person who verifies the content output by machines. This vision of the future now frequently appears in consulting frameworks, academic predictions, and implementation guides from AI vendors. Regulators have even written it into regulations: the EU AI Act requires human supervision of high-risk AI systems, and manual verification has almost become the general shorthand for responsible AI implementation.
The logic behind such predictions is largely the same: first find out the things that AI cannot do for the time being, and then delineate the work that humans need to undertake to fill the gaps. In this line of thinking, humans are a layer of safety protection wrapped around the still imperfect automation. On the freelance platform Fiverr, the fastest-growing demand in the writing category is "manual polishing of AI content" (search volume rose 641% within half a year). Proofreading AI-generated manuscripts has become an independent category of services since 2023. The translation industry is particularly obviously impacted by AI. Nowadays, the business of proofreading machine-translated texts accounts for about a quarter of the revenue of European language service companies, while the scale of traditional translation business shrinks year by year. Even those retained positions established specifically to correct machine outputs have their work content constantly narrowing, requiring fewer and fewer skills, and almost no longer requiring employees to learn new capabilities.
Asking humans to fill the gaps left by machines sounds prudent and reliable, and it also promises job continuity when many people are worried about losing their jobs. But the problem is that this line of thinking starts from the existing work process, picks out the tasks that intelligent agents can take over, and then assigns the remaining trivial tasks to humans. From this perspective, the future organization is nothing more than removing the automated tasks on the basis of today's organization and adding an extra layer of human supervision. Here, humans have never been regarded as a source of brand-new productivity.
This view itself has many hidden dangers. Decades of research in the field of automation have shown that the easier-to-use an automation system is, the more fragile this layer of human safety defense becomes. The higher the reliability of automation, the easier it is for the person responsible for monitoring it to slack off, their skills gradually degrade, and it becomes more and more difficult for them to concentrate. Forty years ago, Lisanne Bainbridge called this phenomenon the automation paradox: the more automation does, the more you rely on human capabilities that are precisely being slowly taken away by automation. And AI further amplifies this problem. When AI makes mistakes, the content it outputs is as assertive in tone, fluent in writing, and complete in structure as when it is correct. Reviewers who only check the readability of the text cannot tell at all whether the content is true or not.
Even if the verification mechanism works and humans successfully catch the errors, every correction will be collected as training data. Every special case handled by humans will teach the system to deal with similar situations on its own; the better the quality of the training data, the smaller the scope that requires manual verification in the future. For example, xAI cut about 500 general "AI coaching specialists" in September 2025, and announced plans to expand the scale of specialized positions tenfold. If you want to pick the human positions that are most likely to disappear in the future, the life cycle of such verification positions is undoubtedly the shortest.
However, organizations built based on the manual verification model have a deeper problem: this top-level blueprint puts the cart before the horse in understanding transformation. It makes enterprises operate around the constantly shrinking boundary that machines cannot do for the time being. But there is a bigger opportunity hidden here: AI breaks the original constraints, making brand-new work combinations, brand-new decision-making authorities, and brand-new value forms possible. Accordingly, enterprises have the opportunity to restructure organizations and recombine businesses around these new decisions, instead of clinging to the shrinking remaining parts of old work. This is exactly the core that this article intends to discuss.
Two Waves of Transformation
Whenever a general-purpose technology enters a certain industry, enterprises will face similar organizational choices. As machine capabilities become stronger, various restrictions that once bound work methods will be broken, allowing people to shift to high-value work that was previously difficult to access.
The lifting of an execution-level constraint will broaden the decision-making space that an organization can undertake and make. Just look at the changes that have taken place in the film industry in recent years. In 2019, *The Mandalorian* proved to the industry that giant LED screens can render and generate virtual scenes in real time, allowing the filming team to shoot complex special effects footage directly in front of the camera. This seems to be just an innovation in special effects technology, but in fact it kicked off the prelude to the fundamental transformation of the organizational model of film production. The changes it brings are actually more far-reaching: many decisions that were originally to be finalized in the post-production stage are moved forward to the shooting site. The style of the scene, lighting, picture composition, and a large number of visual effects in the final film can all be created and confirmed while actors are shooting in real life. Creators do not need to build real sets, nor wait for months of special effects construction period, to switch between rich virtual scenes; the world-building at the level of big-budget productions can also be achieved with the budget of TV series.
The focus of special effects artists' work has also changed from repairing footage after shooting to participating in deciding what the camera should capture from the very beginning. After the constraints disappear, the scope and occurrence position of valuable decisions change accordingly, and the organizational structure of the film industry is also rebuilt around the people responsible for making these new decisions. A number of new positions were born as a result, integrating cinematography, art set design, visual effects and computer technology, such as virtual production supervisor, LED studio operator, and virtual cinematographer.
Similar "recombination" of businesses can also be seen in the evolution of news editorial departments. Traditional editorial departments bind three things together: fact collection, manuscript writing, and content distribution. After generative AI makes writing manuscripts, generating headlines, and making summaries easier, the response of many editorial departments is only to let journalists spend more time reviewing AI-generated manuscripts, while the structure of the editorial department itself remains unchanged.
But the real opportunity for editorial departments is to move upstream: to identify clues worthy of in-depth digging, verify information sources, and produce evidence that others can rely on. A very illustrative example is that the most iconic editorial innovation of BBC in the AI era is not a production line with manual review that writes manuscripts faster, but BBC Verify. In the face of endless AI disinformation, the editorial department specially set up a team of journalists to verify the authenticity of controversial claims based on open source intelligence, satellite images, traceability analysis and data science, rather than simply reposting reports. Reuters has also expanded in the same direction, building a dedicated verification and fact-checking team, and making verified, fact-based reports an independent commercial product.
Management Transformation
Breaking out of the manual verification model does not mean waiting for AI to iterate continuously until manual verification is no longer needed. Leading enterprises will not first think about which tasks machines can take over, but think in reverse: what new positions that did not exist before has this technology created, and then redesign the organization around these positions. To achieve this, you can refer to the following steps:
1. Sort out broken constraints and new capabilities, instead of taking inventory of automatable tasks
The conventional AI transformation roadmap starts with task inventory: which work can be handed over to machines and how many working hours can be saved. This starting point needs to be replaced. For each business module, the question to ask is: which old constraint has disappeared? What new opportunities does this open up? What things were previously too costly, too slow to iterate, or too small in target market to be implemented at all?
In 2021, IKEA's customer service chatbot began to undertake routine consultations, freeing up employees to do higher-value work. Between 2021 and 2023, this chatbot handled 3.2 million consultations, saving nearly 13 million euros in costs. But Ingka Group (IKEA's largest retail operator) went further and provided retraining related to remote interior design for 8,500 employees in its customer service centers. After reallocating manpower to personalized consulting services, IKEA did not only focus on cutting customer service expenses, but instead expanded its sales and service capabilities. By 2022, Ingka Group's remote sales channel reached an annual revenue of 1.3 billion euros. IKEA did not only use automation to compress the cost of existing businesses, but used automation to amplify the value that employees can create.
2. Fund cutting-edge innovation projects, not just efficiency optimization projects
Solutions that improve efficiency often seem more pragmatic and easier to implement, but they will keep enterprises trapped in the thinking framework of "automation for burden reduction", delaying deeper organizational restructuring. When each AI business is established, the team is required not only to state how many working hours it can save, but also to propose a new problem that this technology has only just had the opportunity to solve. In the UK, Garfield Law obtained the qualification from the UK's Solicitors Regulation Authority in May 2025, becoming the first "AI-driven law firm" in the UK. It undertakes small debt recovery that traditional law firms find unprofitable, with service fees as low as 2 pounds, and every case has a licensed lawyer taking full responsibility throughout the process; within one year of launch, it initiated hundreds of lawsuits and won its first controversial court hearing. This is not simply improving efficiency in the original market, but a brand-new market spawned after the change of economic cost structure.
3. Continuously evaluate human capabilities and keep pace with the iteration of AI capabilities
If the job definition is static, it cannot keep up with the AI capability curve that is updated every few months. Johnson & Johnson uses machine learning to calculate the skill status of each person from the work data of thousands of technical personnel, score them against the capabilities required by the enterprise's next stage strategy, and carry out multiple rounds of iteration to continuously update the evaluation. Johnson & Johnson does not regard employee capabilities as a static stock, but builds a system to continuously judge what employees can do at present, what they need to learn, and in which direction their value contribution can extend as technology evolves.
4. Embed governance rules into business processes, instead of adding an extra verification checkpoint at the end
Post-event verification nodes are not only insufficient in stability, but will gradually become invalid as technology evolves. The direction of governance should not be to reduce control, but to change to a new governance paradigm: embed rules into business processes, let the responsible person take charge of control, and carry out continuous verification.
You can look at Morgan Stanley's practices before and after the full launch of the GPT-4 assistant in its wealth management business. In the pilot phase, internal experts first defined the standards for high-quality responses, sorted out about 400 "benchmark questions" with confirmed standard answers to test the system. After that, Morgan Stanley embedded the evaluation link into the formal business link, and ran a set of regression test sets composed of benchmark questions every day to identify the shortcomings of the system, and check whether the version update would damage the response quality and touch the compliance red line. In this knowledge service field with extremely high regulatory requirements, this solution achieves continuous governance during the deployment process, rather than occasionally reviewing the responses given by consultants after the fact.
5. Verify whether you are really promoting transformation
Real organizational restructuring has three core differences from simple task redistribution. First, the enterprise boundary will be redrawn: the enterprise will choose to internalize the learning closed loop and core AI capabilities that drive transformation, instead of always purchasing and renting the capabilities of service providers from outside. Second, organizational units change: teams are built around new problems that have just become solvable, instead of being divided according to the links of the old business assembly line. Third, the center of gravity shifts: the functional department responsible for carrying out this new business is closer to the CEO.
If you only add a number of new job titles without the above three changes, it cannot be called restructuring. It is just redistributing tasks in another form, which still belongs to the organizational model of manual verification.
John Deere's transformation just implements these three points. First, it redrew the enterprise boundary, acquired the computer vision startup Blue River Technology in 2017, and brought the core model iteration capabilities in-house. Second, it adjusted its organizational units in 2020, no longer dividing equipment business by separate product lines, but reorganizing around production systems - that is, building businesses around interrelated tasks and decisions in crop planting and other agricultural work. Third, the 183-year-old company established the position of Chief Technology Officer for the first time. This adjustment changed the center of power: the authority and responsibility of the unified technology stack was assigned to an executive who reports directly to the CEO. This series of decisions changed the capabilities that John Deere independently controls, the way businesses are organized, and the ownership of technology-related decisions. It did not force AI into the existing organizational framework, but redesigned the entire framework along the opportunities opened up by new technologies.
As AI clears the constraints on the execution side, the benefits will flow to enterprises that independently decide how to use this new freedom: what products to make, who to serve, and which problems to tackle. If enterprises only assign their employees to handle the remaining trivial tasks, they will lose more than just transformation opportunities. They mistake business disassembly as the end point, rather than the starting point to inspire brand-new thinking. Over time, those new control rights - model iteration, governance rules, customer relationships, exclusive knowledge that can continuously optimize the system, will gradually fall into the hands of tool vendors and competitors; their own employees, on the contrary, will become labeling trainers who help others cultivate competitive advantages.
Keywords: #AI
Sangeet Paul Choudary | Text
Sangeet Paul Choudary is an executive advisor on AI and platform strategy, a senior fellow at the University of California, Berkeley, and has been selected as the author of "Top 10 Must-Read Articles" by *Harvard Business Review* four times.
Zhou Qiang | Editing and Proofreading
This article is from the WeChat official account "Harvard Business Review" (ID: hbrchinese), written by HBR-China, and published with authorization from 36Kr.