A set of research findings made me sit for a while after reading them. The researchers discovered that employees who repeatedly use AI for summarization, rewriting, and information synthesis are gradually losing their understanding and judgment of original materials. This is not an isolated phenomenon but an organization-wide trend. An even more striking conclusion appears in another article: companies that lay off employees to let AI take over their work, if they fail to capture employees' organizational knowledge before the layoffs, will only have AI "mess things up faster."
These conclusions come from several in-depth articles published continuously by CIO.com from June 22 to 23. When viewed together with the "Humagent" organizational concept proposed by Inspur Information at the AIEC conference, enterprise AI management is hitting three walls. These walls are not technical issues, but management issues.
The more capable AI becomes, the less organizations will think
Taryn Plumb's article on CIO.com has a straightforward title: Your AI strategy may be training employees to stop thinking. The core argument comes from researchers—when employees repeatedly use AI to summarize, rewrite, and synthesize information, their ability to understand original texts deteriorates.
I have experienced this myself. After getting used to having AI summarize a 50-page report, your ability to grasp the details of the report does indeed decline. You receive the "conclusions" distilled by AI, but you lose the opportunity to discover unexpected information in the original materials and form independent judgments. This process is gradual, and you may not even notice that you are becoming "lazy."
The problem is that the more "successful" the AI deployment, the faster this degradation occurs. In an organization where everyone uses AI for information processing, efficiency appears to rise on the surface, but the organization's original knowledge density is actually declining. When the day comes that you need employees to make judgments on the conclusions provided by AI, you find they no longer have the foundation for making such judgments.
The counterintuitive part is this: The higher the AI adoption rate and the greater the usage, the higher the risk of organizational thinking ability degradation. That upward-sloping AI usage curve on your dashboard may simultaneously be drawing another downward curve — the organization's independent judgment.
In another article, Karan Gupta pointed out the related pitfall: measuring AI maturity by employee token consumption is "rewarding activity rather than value." Your dashboard shows that employee AI usage has tripled, and you think the deployment is successful. But it might just be that employees have outsourced tasks they could have done themselves to AI, without touching the areas that truly need process modernization. Evaluating AI effectiveness by token consumption is like evaluating a programmer by the number of keystrokes — the data looks impressive, but the meaning is close to zero.
Knowledge was hollowed out before AI took over
The title of Kash Mehli's article is even more pointed: Preventing Organizational Amnesia in the Age of AI. The core argument is: using layoffs to let AI take over work is a trap, if you haven't yet captured employees' organizational knowledge.
What is organizational knowledge? It’s not the SOP written in documents, nor the operation manuals on the Wiki. It’s the veteran employee who has worked in the supply chain for ten years, knowing with their eyes closed which supplier’s delivery dates are unreliable and which client always submits requirement changes at the last minute. This tacit knowledge has never been structurally recorded—once the person leaves, it’s gone.
Let these people go and let AI take over their jobs. What AI takes over is only the "process," not the "judgment." Without context, AI will only make even dumber decisions at a faster speed. Mehli's original words were, "AI will only mess things up faster."
This sentence deserves to be posted on the wall of every CIO's office. The correct order should be: first capture knowledge, then deploy AI, and only finally consider workforce adjustments. Many companies get the order reversed — they cut staff first to save costs, then let AI fill the gaps, only to find the system frequently makes errors, and end up spending more money to rehire people who cannot recover the lost knowledge.
Naren Gangavarapu proposed a framework called "Rewire or Rebuild" on CIO.com. He said don't choose between "patching existing systems" and "building from scratch." The right approach is to use quick AI wins to fund and plan major architectural changes. Specifically: Don't start with a company-wide AI overhaul. First, find a scenario with clear data ownership and a well-defined business pain point, use AI to make measurable improvements, and then leverage that small win to drive larger transformations.
Data ownership is more critical than technology
Scott Smeester's article title hit the pain point precisely: why your AI pilot died at the data ownership meeting, not in the demo.
The scenario he describes should be very familiar to many CIOs. The AI pilot demo runs perfectly, the boss is satisfied, and it's ready to be rolled out. Then it gets stuck. Not stuck on technology, but stuck when meetings reveal that data ownership is unclear. Does customer data belong to the sales department or the marketing department? Does production data belong to the factory or IT? Who has the authority to feed data into the AI model? These kinds of meetings can drag on for months, completely draining the project's momentum.
| segment | Technical team perspective | Actual organizational resistance |
|---|---|---|
| AI System Setup | Build a RAG system in 1-2 weeks | No resistance |
| Data Authorization | Just sign at a meeting | 3-6 cross-departmental meetings |
| Data usage boundary | Written in the configuration file | A three-way game among legal, business, and IT departments |
| Launch promotion | Technical Readiness | The momentum has been exhausted. |
I agree with Smeester's assessment: the technical threshold for AI is rapidly decreasing, but the threshold for data governance within organizations hasn't dropped at all. A tech team can set up an RAG system in a week, but getting three departments to agree on data usage boundaries might require six meetings. The faster the technology, the more glaring the slowness of management becomes.
The root of this problem lies not in AI, but in the fact that enterprises have never truly solved the issue of data governance. AI has simply turned this from a "can be delayed" problem into a "must be solved" one. Previously, data ownership was unclear, and everyone used their own data without conflict. Now that AI needs to access data across departments, unclear ownership directly brings everything to a halt.
Humagent and Your AI Management Checklist
At the AIEC conference, Peng Zhen, Chairman of Inspur Information, proposed the "Humagent" (Human+Agent) organizational concept: enterprises evolve from managing Humans to managing Humagents, redefining positions, roles, permissions, responsibility boundaries, and performance evaluations.
Midea Group has given this concept some tangible reality. Wei Xiaogang, Vice President of Midea Cloud Intelligence, stated that Midea has built 13,000 Factory Agents, covering the entire chain of R&D, production, supply, sales, and service. These 13,000 intelligent agents operate within the factory, performing multi-scenario, multi-role, and multi-business chain collaboration.
13,000 Midea Factory Agents covering the full chain of R&D, production, supply, sales, and service
The judgment of Shan Zhiguang from the National Information Center is also worth remembering: Token is transforming from a unit of large model technology into a new economic unit that connects electricity, computing power, model services, and application value. In the future, the basic unit for measuring a company's AI operational costs may not be servers or person-days, but Token consumption.
Peng Kaiping from Tsinghua University offers a more humanistic perspective: intelligent agents handle standardized execution, while humans focus on aesthetic creation, strategic decision-making, and emotional connection. This division of labor is correct, but the prerequisite is that your organization has truly achieved "knowledge capture first, AI deployment later." Otherwise, if you let AI handle standardized execution, you may find that it hasn't even learned the standards.
Returning to the practical level of the CIO, a few suggestions:
Conduct an "AI Dependency Audit". List which positions have employees using AI for information processing (summarization, rewriting, analysis), and assess whether their independent judgment abilities are deteriorating. For high-risk positions, set a "No AI Day"—requiring at least one day per week for manual processing of core information to maintain the organization's original thinking capacity. This is not moving backward, but preventing the atrophy of the organization's thinking muscles.
Conduct knowledge capture before layoffs. If you are considering replacing certain positions with AI, spend two to three weeks on structured knowledge extraction before layoffs: have employees articulate their decision-making logic, document exception handling processes, and organize informal collaboration networks. This tacit knowledge is a necessary prerequisite for AI to take over. Mehli put it clearly: without this context, AI will only mess things up faster.
Put data governance as the first step in AI projects. Before the technical team writes the first line of code, first convene a data ownership meeting with legal, business, and IT departments. Clarify the owner, usage boundaries, and AI training authorization scope for each type of data. This meeting may take longer than the development itself, but it determines whether the project can be implemented. Smeester's observation is accurate—AI pilots don't die in demos; they die in data ownership meetings.
Redesign performance evaluation. With agents in the organization, the criteria for evaluating employees can no longer be just "the number of tasks completed"—AI can help you complete tasks. What you need to evaluate is the quality of judgment, the ability to handle exceptions, and contributions to knowledge transfer. Midea has 13,000 Factory Agents running, and Midea's way of evaluating employees must also be changing. If you don't change, the more agents there are, the less people will know what they are doing.
The three walls of AI management are ultimately not an AI issue, but a management issue. AI merely amplifies management deficiencies that have always existed but were overlooked. Those who first confront these deficiencies will be the ones to truly reap the rewards in this wave of AI.
