以AI留存員工知識：企業不能迴避的法律與治理問題

Before your organisation relies on an AI agent to absorb a departing employee's knowledge, two questions need answers. Most companies are asking neither.

Where this is coming from

In the spring of 2026, a GitHub project called colleague.skill went viral across Chinese tech company chat groups. The tool was built to extract an employee's digital presence — their Feishu and DingTalk messages, their shared documents, their decision logs — and construct an AI agent capable of standing in for them. The README's own guidance was direct: prioritise collecting long-form content the person actively wrote, decision-making responses, then daily messages. The quality of the replacement depends on the quality of what was captured.

The reaction was swift. Days later, a counter-tool called anti-distill appeared on the same platform. Its purpose: feed it your skill file, and it returns a version that looks complete and professional but has the core knowledge stripped out. The author's framing was precise — "your company asks you to write down your work experience as an AI skill? This is essentially distilling you — turning you into a replaceable component." Anti-distill was the counter-weapon.

Both tools were enabled by a specific feature of Chinese enterprise platforms: administrators can bulk-export member documents and communication records from the backend. The employee does not need to cooperate. The company already holds the data.

This is presented as a Chinese phenomenon. It is not. Every major Western enterprise collaboration platform has the same admin export capability. Slack, Microsoft Teams, and Google Workspace all allow administrators to export employee messages, documents, and files in bulk. Any organisation with API access to a large language model can build a functionally identical tool. The ingredients are identical; only the label on the jar is different. Western companies are already doing versions of this under more palatable names: AI onboarding systems, institutional memory tools, knowledge graph platforms, departure documentation requirements. The underlying logic — capture what the employee knows before they leave, so the organisation retains it — is the same.

The business case is straightforward and spreading. Capturing institutional knowledge in AI agents before restructuring reduces rehiring risk, sustains operational continuity with fewer people, and provides a rationale for headcount reductions that frames the decision as process improvement rather than cost-cutting. The practice will grow. The legal and governance requirements have not kept pace with it.

Two questions that must be answered first

Before any organisation proceeds with AI knowledge transfer as a precursor to restructuring, two questions must be answered — in parallel, not sequentially, and before the restructuring decision is finalised, not after.

The legal question: does this organisation have the legal authority to process employee data in this way? Not whether it technically can — it almost certainly can — but whether it has identified a lawful basis, met its transparency obligations, and satisfied the requirements that apply in each jurisdiction where affected employees are based.

The governance question: does the AI agent actually contain what this organisation thinks it contains? Not whether the tool ran successfully — it almost certainly did — but whether what was captured is sufficient to rely on for the decisions the organisation is about to make.

Most organisations proceeding with this practice have not answered either question. The legal layer is skipped because the technical capability exists and the business pressure is real. The governance layer is skipped because nobody has thought to ask it — the tool ran, the files were produced, the project is complete. Both failures are consequential. The second is harder to recover from than the first.

What doing this right actually delivers

This is not primarily a risk mitigation exercise. Organisations that build a proper legal and governance framework for AI knowledge transfer before restructuring get something the others do not: a restructuring that actually delivers what it promised.

The efficiency case for AI-enabled workforce restructuring depends on one assumption — that the AI agent is a functional substitute for the departing employee in the domains that matter. If that assumption is wrong, the restructuring does not deliver the expected efficiency. It delivers a short-term cost reduction followed by operational degradation, rehiring, and the cost of rebuilding institutional knowledge that could not be recovered. The organisations that keep rehiring after layoffs — a pattern well documented in the 2025–2026 restructuring wave — are often the ones that cut before validating what the AI actually contained.

Doing the work outlined in this article delivers six things that the shortcut approach does not:

The legal layer — what authority do you actually have?

When in-house legal first encounters this question, the instinctive response is often: these are work products. The employer owns them. The employee cannot refuse to hand them over. That is correct — and it answers the wrong question.

Work product ownership is a property and contract law question. Data protection authority is a separate legal question, governed by a separate legal framework, that operates independently of who owns the underlying content. An employer that owns all of an employee's documented output still requires a separate lawful basis to process that material through an AI system. Understanding why requires walking through what actually happens technically.

What the AI actually does — the four steps where legal questions arise

The analogy that makes this concrete: an employer owns an employee's performance reviews. That does not mean the employer can publish them, sell them, or feed them to a third-party AI system without separate legal justification under data protection law. Ownership of the document and authority to process the personal information it contains are different questions answered by different legal frameworks.

In-house legal that stops at "these are work products" has answered Step 1. The data protection question lives in Steps 2 through 4.

The lawful basis problem

Data protection frameworks across jurisdictions share a common requirement: personal data may only be processed if the organisation can identify a valid legal ground for doing so. The grounds most likely to be considered in this context are contractual necessity, legitimate interests, and consent. None of them is straightforward here.

Transparency obligations

Independent of lawful basis, most data protection frameworks require that employees be informed about how their personal data is being processed — what is being collected, for what purpose, by whom, and for how long. Standard employment contracts and privacy notices almost never cover AI training as a permitted use of employee data. An organisation that has been exporting employee communications and using them to train AI agents without disclosure has almost certainly failed its transparency obligations, even if it could establish a lawful basis for the processing itself.

The practical question — whether to tell employees that their data is being used to build an AI agent that may replace them — is not just a legal one. It is a strategic and ethical one. But the legal minimum in most jurisdictions is that some form of notice is required. What that notice says, and when it is given, requires careful thought and specialist advice.

The third-party information problem

Employee communications and work documents do not contain only the employee's information. They contain information about everyone the employee ever communicated with — clients, counterparties, external advisors, regulators, colleagues at other organisations. When an employer exports those communications and feeds them into an AI system, it is processing personal data and confidential information belonging to people who have no relationship with the employer, no knowledge that this is happening, and no opportunity to object.

This is a separate and in many respects harder problem than the employee data question. For employee data, the organisation at least has a legal relationship from which to construct a potential lawful basis. For third-party data, there is no relationship, no applicable basis, and in many cases a positive legal obligation running in the opposite direction.

The third-party problem has a practical implication that most organisations have not confronted: a clean solution may not exist. Unlike the employee data question — where a lawful basis can in principle be constructed with sufficient care — the third-party data problem may not be solvable by better compliance work alone. For certain categories of data, the right answer may be that those communications cannot be included in the AI training set at all. A scoping decision that excludes client communications, privileged material, and counterparty correspondence from the export may be the only legally defensible position.

That scoping decision has direct implications for what the AI agent actually captures — and therefore for the governance question addressed in the next section. An AI agent trained only on internal communications and non-confidential documents will be materially less complete than one trained on the full communications archive. The legal constraint and the completeness problem are not independent. They compound each other.

Third-party AI providers and cross-border transfer

If the AI agent is being built using an external LLM provider — which is the case for most organisations outside China — employee data is being transferred to a third party and potentially across national borders. This triggers additional requirements in most jurisdictions: data processing agreements with the AI provider, cross-border transfer mechanisms where required, and in many frameworks a specific privacy impact assessment before the transfer occurs.

The practical minimum before proceeding is not long, but it requires doing the work: audit your employment contracts and privacy notices to identify what they actually cover; identify the lawful basis you are relying on and document it; conduct a privacy impact assessment; determine what disclosure is legally required; and put data processing agreements in place with any AI provider receiving employee data. None of this is technically complex. Most organisations have not done it.

The governance layer — what did the AI actually capture?

Assume for a moment that the legal layer has been addressed. The organisation has a lawful basis, has conducted a privacy impact assessment, has given appropriate notice, and has the right contracts in place. The AI knowledge transfer process has run. The files have been produced. The AI agent exists.

The governance question is: what is actually in it?

The explicit/tacit distinction

In 1958, the philosopher Michael Polanyi published what remains the most precise observation about the limits of knowledge documentation: "we can know more than we can tell." He called the knowledge that resists articulation — the knowledge that experts carry but cannot fully express — tacit knowledge. His example was riding a bicycle: the rider can maintain balance through every shift of gradient and momentum, but cannot describe the physics of how they do it in a way that would teach someone else.

The same distinction applies directly to what an AI knowledge transfer tool captures and what it misses.

This is not a limitation that better tools will solve. It is structural. An AI agent trained on an employee's documented output can reproduce their stated reasoning, their communication style, their procedural knowledge. It cannot reproduce the judgment that operated beneath those things — the knowledge that determined when the procedure applied and when it didn't, when the standard response was right and when it wasn't, when a situation required escalation that wasn't written into any policy.

The reason this matters for restructuring decisions is direct. The employee whose knowledge is being transferred was not just executing documented procedures. They were exercising judgment — in client interactions, in risk assessments, in decisions about what needed escalation and what didn't. If that judgment was a material control in any domain, the AI agent is not a substitute for it. And the organisation making the restructuring decision may not know which domains those were, because the judgment was never documented.

The irreversibility problem

Process failures are recoverable. A system that fails can be fixed; a procedure that breaks can be rewritten; a tool that underperforms can be replaced. A restructuring decision made on the assumption that an AI agent contains what an employee knew is not recoverable in the same way. Once the employee is gone, the organisation cannot conduct a gap assessment. The knowledge that is missing is missing precisely because it was never written down, which means there is no record from which to reconstruct what was lost.

This is the due diligence argument. Before a board approves a restructuring that relies on AI knowledge transfer as a risk mitigation, it should require evidence that the knowledge transfer has been validated — that the AI agent has been tested against real scenarios, that known gaps have been identified and assessed, and that human oversight has been designed for the domains where the original employee's judgment was a material control. Approving the restructuring without this is approving a decision whose operational risk profile is unknown.

How to proceed — a practical framework

The argument above is not against AI-enabled knowledge transfer. It is for doing it in a way that is legally defensible and operationally sound. The seven steps below are the minimum. They are not technically complex. The challenge is organisational: building the discipline to do this work before the restructuring decision is made, when business pressure is pushing toward speed.

The question to ask before the decision is made

The AI knowledge transfer tools work. They produce output. They run successfully. They generate files that look comprehensive and professional. The problem is that "the tool ran" is not the same as "the organisation knows what it has."

The most diligent employees — the ones who document everything, who write detailed post-mortems, who explain their reasoning in long Slack threads — are the ones who produce the most complete-looking AI agents. They are also the ones whose tacit knowledge is most deeply embedded and most thoroughly absent from what the AI captured. The output looks complete precisely because they were good at documentation. It is not complete because documentation has never been the same as knowledge.

The question the board and the GC need to ask before the restructuring decision is finalised is not "did the knowledge transfer run?" It is "do we know what the AI did not get — and have we designed for that gap?" If the answer to the second question is no, the organisation is not making an informed decision. It is making an assumption dressed as one.

The companies that will get this right are not the ones that move fastest. They are the ones that pause long enough to answer the two questions this article is built around — before the employee is gone and the answer becomes impossible to find.