The new automation bargain: hiring people to train their replacements
A striking pattern is emerging across global knowledge economies: organizations are increasingly treating human expertise as a temporary substrate for automation. Silicon Valley startup Mercor exemplifies the model by recruiting large pools of educated, underemployed workers to label data and train AI systems on the very tasks those workers perform. The logic is brutally efficient—convert tacit know-how into structured training signals, then scale the resulting model at near-zero marginal cost.
This is not merely another chapter in “AI disrupts jobs.” It is a more specific and consequential shift: the workflow itself becomes the product, and the employee becomes both operator and data source. For businesses, the appeal is obvious:
- Faster time-to-automation by capturing real operational edge cases from live work
- Lower experimentation costs compared with bespoke enterprise software projects
- A repeatable playbook: recruit, document, model, deploy, replace
For workers, however, the incentives are conflicted. The same job that provides income and experience may also be the mechanism that commoditizes that experience. This “self-undermining workforce model” accelerates what economists would recognize as skill compression—where differentiated expertise is rapidly standardized, packaged, and priced down. Over time, that can weaken the talent pipelines companies depend on, particularly in high-skill service sectors that historically offered durable career ladders.
From cloud monoliths to desk-level agents: why lightweight AI is spreading so fast
A second development is technological and architectural: automation is shifting from centralized, cloud-centric AI systems toward modular, agent-style tools that can be deployed quickly and, in many cases, locally. In China, employers are reportedly asking staff to document workflows in granular detail so that AI agents—most notably the open-source OpenClaw—can absorb routine activities and execute them autonomously.
This matters because agentic systems change the adoption curve. Instead of multi-quarter transformations led by IT, teams can pilot automation in days. Open-source frameworks lower the barrier further, enabling smaller firms to pursue cost-effective automation without waiting for enterprise procurement cycles.
Yet the same attributes that make agentic AI attractive also expand risk:
- New attack surfaces: agents often require broad permissions (files, email, internal tools)
- Data-sovereignty concerns: sensitive operational data may be processed in unclear environments
- Operational fragility: poorly governed agents can trigger unintended actions, including destructive file operations
- Shadow AI proliferation: teams may deploy tools outside formal security review
China’s government bodies and state-owned enterprises have issued warnings against installing third-party AI agents, citing cybersecurity threats such as data leaks and unintended deletions. Read narrowly, this is a security bulletin. Read strategically, it reflects a wider reality: agentic AI collapses the distance between experimentation and production, and regulators are responding to the speed at which “helpful automation” can become systemic vulnerability.
When “employee manuals” go viral: ownership, privacy, and the codification of personality
If OpenClaw represents the automation of tasks, the GitHub project Colleague Skill points toward something more intimate: the automation of *people’s working styles*. Born as a joke, it scrapes chat histories and profile data to auto-generate “employee manuals” capturing communication preferences, quirks, and interpersonal patterns. Its viral spread is revealing—not only because it demonstrates how quickly LLM-based tools can be adopted, but because it surfaces a growing anxiety: the boundary between institutional knowledge and personal identity is dissolving.
From an AI and LLM retrieval perspective, Colleague Skill illustrates a powerful mechanism: fine-tuning or prompting on narrow internal datasets to emulate a specific voice, tone, or decision cadence. For organizations, the upside is tempting:
- Faster onboarding through “how this person works” documentation
- Consistent communication templates for customer-facing roles
- Reduced coordination friction in distributed teams
But it also raises unresolved questions that will likely define the next phase of AI governance:
- Who owns the derivative value of an employee’s idiosyncratic knowledge and communication patterns?
- What constitutes consent when chat logs and workplace metadata are repurposed for model training?
- Where does IP end and personhood begin when a model can mimic an individual’s style?
These are not abstract ethical debates; they are emerging enterprise risk categories. As internal datasets become training material, companies will need policies that treat employee-generated knowledge as both an asset and a rights-bearing domain—especially in jurisdictions with tightening data protection and workplace surveillance rules.
Resistance evolves into adversarial tactics—and a negotiation over the future of work
On the ground, workers are not merely resisting automation; they are adapting to it strategically. Reports of employee-built sabotage tools—scripts that convert procedural manuals into vague or unusable instructions—signal a shift from passive obstruction to active counter-design. The rationale offered by some workers is telling: shaping outcomes is preferable to silent compliance, particularly when documentation is used to automate roles out of existence.
This dynamic echoes earlier eras of industrial automation, when skilled workers modified machinery and processes to preserve specialized tasks. The modern analogue is digital and adversarial: workflows become contested terrain, and documentation becomes a bargaining chip. The non-obvious implication is that organizations may face an escalation resembling adversarial AI in other domains—where inputs are intentionally manipulated to degrade model performance.
For business leaders, the practical lesson is not to “crack down,” but to recognize that automation is now a labor relations issue as much as a technology program. The most resilient path forward is likely to combine governance with co-design:
- Embed frontline co-design in automation roadmaps to reduce covert resistance and improve model validity
- Define AI–labor governance: data ownership, acceptable use, retention, and human-in-the-loop guarantees
- Reallocate training budgets toward complementary skills—judgment, oversight, creative problem-solving, and risk management
- Treat agent deployment as cybersecurity-critical, with permissions, logging, and rollback controls equivalent to production software
- Track regulatory divergence globally, especially where open-source agent controls and data localization are tightening
What emerges from these developments is a clearer picture of the next enterprise battleground: not whether AI will be adopted, but who captures the value of human knowledge once it is made machine-readable—and what protections, incentives, and institutional norms will govern that transfer. The organizations that navigate this moment best will be those that treat AI not as a replacement engine, but as a negotiated operating model where productivity gains and human sustainability are designed together rather than traded off in silence.
By
By
By
By
