Summary
Consider a senior strategist who has worked intensively with artificial intelligence (AI) for two years. Their AI environment knows the company’s terminology, has encountered previous strategies and failed initiatives, and has been corrected repeatedly on which recommendations will never survive implementation. The strategist has taught it which sources to trust, which metrics deserve skepticism, and which questions expose weak thinking, and over time the combination becomes remarkably productive.
Then the strategist leaves. The company retains the AI license, the documents, maybe even a folder full of prompts. It has not necessarily retained the capability. That gap creates a question few AI strategies address: when exceptional performance depends on the combination of a skilled person and a deeply developed AI environment, how does the organization scale what is valuable? The answer cannot be “copy the prompts.”
The Valuable Unit Has Changed
Organizations traditionally think about employee capability as something located primarily in the person. A strong employee brings expertise, judgment, relationships, and institutional knowledge, and technology supports that capability from the outside. AI changes the boundary, because a highly developed AI user distributes pieces of their working method across several places at once: some knowledge stays in their head, some lives in company documents, some appears only in previous AI conversations, and some exists only in the corrections they repeatedly make when the model gets something wrong.
The productive capability now spans a human-AI working system, and that distinction matters because organizations usually transfer its components separately. IT transfers the software, knowledge management transfers the documents, and HR handles the role transition, but the integrated working capability can still disappear between the seams.
Prompts Capture Instructions. They Rarely Capture Expertise.
Prompt libraries became an early answer to this problem, and the logic made sense on its face: find employees getting strong results, save their prompts, give everyone access. That can improve performance on repeatable tasks, and it can also create false confidence about what has really been transferred.
Consider an expert asking AI to evaluate a proposal using the company’s standard criteria, and to stay skeptical about the revenue assumptions and any implementation dependencies the team has historically underestimated. The prompt contains useful instructions, but it does not explain why those revenue assumptions deserve skepticism, and it does not contain the previous projects that taught the expert to worry about implementation dependencies in the first place. It does not capture how the expert recognizes an AI critique that sounds intelligent but misunderstands the organization. Another employee can copy the prompt and copy an artifact of the method. They have not necessarily acquired the method itself.
Personalization Makes the Gap Larger
Modern AI products are increasingly built to accumulate context around individual users. OpenAI’s own guidance on ChatGPT personalization recommends that users provide recurring context about their role, responsibilities, and working requirements specifically so the system becomes more useful and consistent over time, and its memory feature is explicitly designed to carry that context forward across conversations without the person repeating themselves.
That design creates genuine productivity value, and it also complicates capability transfer. Two employees using the identical underlying model can end up with very different effective AI environments, because the experienced user’s system has encountered more relevant context, better instructions, and a richer source collection, and knows when to distrust a response and how to recover when it fails. Giving another employee the same model does not reproduce any of those conditions, because standardized access and transferred capability are two different things.
Your Best Thinker May Have Created Undocumented Organizational Infrastructure
This is where the issue grows larger than personalization. Experienced employees carry enormous amounts of tacit organizational knowledge: which documented process everyone quietly modifies in practice, which customer exception looks insignificant until it reaches the executive team, and who needs to be consulted even though the process map never mentions them.
Harvard Business Review has highlighted this exact problem in the context of AI agents. One June 2026 article on agentic system design found that as companies rush to deploy AI agents, they are discovering that much of their most important organizational intelligence lives outside formal systems and documented processes entirely. A companion piece from the same month argues that as AI takes on more complex work, the binding constraint is no longer access to technology but an organization’s ability to make its own decision-making explicit, since judgments about risk, exceptions, and tradeoffs have traditionally existed only as tacit knowledge inside experienced employees.
AI can expose that hidden layer, because experts have to explain parts of it to get useful results out of the system. That creates a real opportunity, and it also creates a new form of fragility: if the expert’s corrections and operating logic stay trapped inside their personal AI interactions, organizational knowledge has moved into another silo. The silo just happens to be an intelligent one.
The Goal Is Not to Clone the Expert
This distinction needs to be explicit. Organizations should not try to reproduce a person’s complete AI history across the workforce, since much of it will be irrelevant, some will be personal, and some will reflect outdated assumptions or individual preferences that should never become organizational standards. The expert can also simply be wrong about some things, and scaling their environment without judgment would institutionalize those errors right alongside the useful knowledge.
The real objective is identifying which elements of exceptional AI-supported performance deserve to become organizational capability. That requires decomposing the performance rather than exporting it wholesale.
Start With the Work the Expert Does Better
Exporting chat histories is the wrong starting point. Performance is the right one: identify the AI-supported work where the person consistently produces unusual value, whether that is stronger competitive analysis, faster diagnosis of troubled implementations, or account strategies that consistently catch risks other employees miss.
Then ask what creates the difference. The answer may involve domain expertise, a proprietary methodology, superior source selection, or a distinctive sequence of AI interactions the person has refined over time. Their advantage may come from knowing which questions to ask, or from recognizing which AI answers should be rejected outright. The organization needs to understand the performance mechanism itself before it can attempt to scale it.
Separate Knowledge From Judgment
This is one of the most important steps in the whole process. Some of what makes the expert effective can be captured as knowledge: product history, customer information, prior decisions, process documentation, and known exceptions. Other elements depend on judgment: which exception matters here, how much evidence is enough, when a recommendation should be challenged, and which risk deserves executive attention.
AI can help make some of that judgment criteria explicit, but the organization should still respect the distinction between what a repository can hand someone and what only firsthand experience builds. A knowledge repository can tell an employee what happened previously, though it takes something closer to lived judgment for that history to turn into good decisions under pressure.
Sterling Phoenix’s AI-Era Institutional Knowledge System exists around exactly this boundary, governing how knowledge used by humans and AI is identified, validated, connected, retrieved, refreshed, transferred, and retired. The scaling problem begins with institutional knowledge, and it keeps going well past that first step.
Capture the Decision Logic
The next layer is harder. Organizations need to capture enough of the expert’s reasoning method to make it usable by others, which is the same challenge Harvard Business Review describes for AI agents more broadly: firms need to translate tacit judgment into structured guidance if they want AI systems to perform consistently on complex work, building what the article calls “judgment infrastructure” that makes institutional knowledge portable and scalable rather than locked inside one person’s head.
The same principle applies to scaling expert AI use directly. Ask the expert what they examine first, what causes them to distrust an apparently strong result, which exceptions override the normal rule, and which mistakes AI repeatedly makes on this particular kind of work. Those answers start turning tacit reasoning into reusable operating knowledge, and the organization ends up capturing decision logic alongside the instructions.
Capture the Corrections
AI corrections may become an unusually valuable source of organizational knowledge in their own right. Think about what happens when an expert repeatedly works with AI: the model makes a recommendation, and the expert explains that it won’t work here because the sales cycle behaves differently, or that two customer categories the model treated as equivalent require different handling. Those corrections are signals. They reveal the exact distance between generic model knowledge and organizational reality.
When an expert repeatedly corrects AI about how the business really works, those corrections may contain knowledge worth institutionalizing instead of knowledge that evaporates once the conversation ends.
Convert Repeated Expertise Into Shared Context
Once useful knowledge has been identified and validated, it needs a governed home: approved knowledge repositories, role-specific context, decision criteria, exception libraries, verified examples, or reusable AI skills. The exact technical implementation will vary by organization, but the architectural principle should stay stable: knowledge required for organizational performance should not depend exclusively on one person’s AI history.
This is where a custom AI application, a governed enterprise workspace, or a reusable skill can become more valuable than another prompt library, because it lets the organization supply important context intentionally instead of hoping employees accumulate it independently over years.
Build the Method Into the Workflow
The next step is operationalization. Suppose an expert consistently performs exceptional acquisition screening with AI, and the company identifies what makes the work strong: the evidence requirements, the evaluation criteria, the recurring exceptions, the challenge questions, and the conditions that require legal, finance, or executive review. Those elements can now become part of the workflow itself, so the AI system receives the required context automatically, the output follows a defined structure, and known failure modes trigger review before they become problems.
Employees can still exercise judgment on top of this. They begin from a stronger organizational foundation instead of rebuilding it from scratch every time, and that is how individual AI capability starts becoming institutional capability rather than staying a personal advantage that leaves when the person does.
Preserve Human Judgment in the Transfer
There is an obvious danger here. An organization can become so focused on scaling the expert’s method that it freezes the method in place while expertise, markets, and technology keep moving. AI can make stale knowledge look unusually polished, which is precisely the mechanism Harvard Business Review describes in its 2026 work on organizational “knowledge decay”: AI-generated output that looks finished but contains errors or thin reasoning forces colleagues downstream to spend time verifying or redoing it, and as those errors compound across teams, the organization’s collective knowledge base deteriorates rather than improves.
Sterling Phoenix’s AI-Era Institutional Knowledge System includes renewal and retirement for exactly this reason. Someone must determine when scaled knowledge is still valid, challenge the methodology periodically, add new exceptions, and retire assumptions that no longer hold. Scaling expertise requires an ongoing lifecycle rather than a one-time export.
The Expert Still Matters
A company can capture an expert’s methodology without capturing the expert, and that is a healthy outcome rather than a loss. Organizations should avoid the belief that enough documentation and AI context make experienced people interchangeable. Expertise includes pattern recognition developed through exposure, the ability to notice when a situation falls outside known categories, judgment under genuine uncertainty, and relationships the org chart will never capture.
Research on AI and knowledge work reinforces the continuing importance of that human layer. The same Microsoft study cited earlier in this series found that generative AI shifts critical thinking toward verification, integration, and task stewardship rather than eliminating the need for it. AI can increase the reach of an expert and help transfer pieces of their expertise, but it cannot guarantee that the person receiving that transfer has developed equivalent judgment. The organization needs both transfer and capability development, and neither one substitutes for the other.
This Creates a Succession Problem
Traditional succession planning asks who can perform this person’s role. AI-era succession needs a second question: what human-AI capability would disappear if this person left tomorrow? That question can reveal dependencies that never show up on an organizational chart, whether it is a marketing leader’s sophisticated research environment, an operations director’s AI-supported diagnostic method, or a financial analyst’s workflow for interrogating assumptions that no formal procedure has ever captured.
The employee may leave behind every conventional file, and the replacement can still lose substantial productive capability the moment that person walks out the door. That is a continuity risk most organizational charts never surface.
AI Capability Concentration Can Become Organizational Fragility
Sterling Phoenix already addresses this through Organizational Capability Resilience, whose architecture includes Minimum Viable Independence, Capability Continuity Envelope, Capability Debt, and Recovery Architecture as mechanisms for preserving critical capability as AI changes how work gets done. The master content architecture identifies AI capability concentration as an organizational risk in its own right, whenever too much capability concentrates in too few people, vendors, or systems.
The expert-AI relationship is one form of that concentration, and it is easy to underestimate precisely because the capability appears to belong to an employee who is performing extremely well. High performance can conceal dependency, and the organization typically discovers the dependency only when the employee changes roles, leaves, loses access, or stops maintaining the system. By then, it is too late to start understanding how the capability worked in the first place.
Scale the System in Layers
Organizations can approach expert AI transfer through six layers, moving from the personal to the institutional. Start with the individual layer, identifying what the expert knows, how they work with AI, and where human judgment remains essential. From there, the knowledge layer captures validated organizational knowledge, examples, corrections, and exceptions, while the method layer makes recurring reasoning patterns and challenge questions explicit.
The workflow layer embeds those transferable elements into repeatable AI-enabled work. Once that is running, a capability layer trains other people to use, evaluate, challenge, and improve the resulting system, and a governance layer assigns ownership for access, validation, updates, and eventual retirement.
This is where scaling stops being a personal favor and becomes organizational design. The company moves past copying an employee’s tool configuration and starts building a capability that can survive beyond the individual who first developed it.
The Junior Talent Problem Still Exists
Scaling expert AI creates another responsibility organizations should not skip. AI is already changing the early-career work through which people traditionally developed expertise in the first place. McKinsey’s 2026 analysis of early-career talent found that research, documentation, data cleanup, and preliminary analysis, the exact activities through which young employees have historically built instincts and earned the right to take on more, are now being streamlined or absorbed into AI systems.
An organization can therefore make two mistakes at once: failing to capture the expertise of today’s strongest people, while simultaneously automating away the experiences that would have created tomorrow’s strongest people. A good transfer architecture addresses both problems together. The expert’s AI-supported methodology can become a teaching system, where junior employees examine why particular evidence matters, compare their own reasoning against the expert-derived method, and make decisions under controlled conditions with real feedback. The objective is producing stronger thinkers who understand why the workflow works, rather than employees who can merely operate it.
Five Questions to Ask About Your Best AI Users
Leaders can begin without launching a major knowledge-management initiative. Identify several people producing unusually strong results with AI and ask what this person can accomplish that comparable employees cannot reliably reproduce, what organizational knowledge their AI-supported work depends on, what methods and decision logic they have developed, which parts could become shared capability without copying personal or outdated context, and what capability would disappear if this person and their AI environment became unavailable tomorrow.
The answers reveal where AI capability is becoming concentrated, and they also reveal where institutionalizing it now could create real leverage later.
The Competitive Advantage Is Larger Than the Model
Companies can buy access to the same frontier models. Competitors can hire from similar talent pools, and most organizations can eventually acquire similar AI infrastructure without much difficulty. The harder asset to copy is the accumulated combination of organizational knowledge, validated methods, decision logic, exceptions, human expertise, and AI-enabled workflows built through actual work over actual time.
That combination can become genuine proprietary operating capability, and it can also disappear one employee at a time if nobody recognizes it while it still exists. This changes how leaders should think about their strongest AI users. They are more than early adopters who discovered better prompts; some of them are building prototypes for how the whole organization could work. The leadership task is figuring out which parts deserve to scale: capture the knowledge, make the method explicit, preserve the judgment, build it into the work, teach others how to challenge it, and keep it current. That is how an organization scales its best thinker’s AI without pretending it can clone its best thinker.
