Breaking Out of AI Anxiety: How Black-Collar Workers Are Forged (Tools)

Translator’s note. “Black-collar” (黑领), in this author’s series, names a worker whose value in the AI era lies on the boundary between humans and AI systems — designing the limits within which AI is allowed to operate, tending the artificial nature that AI has created, and representing human interests when the system overreaches. The previous essay introduced the concept; this one is the “tools” piece, examining the technical and organizational instruments through which black-collar work is actually done.

About the author. Chun Xia is Founding Partner at TSVC.

Original Chinese version. This essay was first published in Chinese on WeChat. Read the original →.

White-collar workers face structural elimination, and the black-collar worker becomes the new role. But who will become a black-collar worker? How does a white-collar worker make the transition? (For “what is a black-collar worker?”, see the abridged version of White-Collar Workers Are Becoming Redundant, but “Black-Collar Workers” Are Just Arriving.) This essay covers the second half of the picture: the AI applications that are digging their own graves (the dead end), the two technological frontiers, and the three faces and toolkit of the black-collar worker (the way out). I am also writing this for myself, for founders, and for investors.

1. Digging Their Own Grave: The Death-Loop of Enterprise AI Applications

In 2026, when venture capitalists look at enterprise AI projects, most have hit pause. Over the past year and a half, the AI software sector has gone through the most frenzied funding, the fastest product iteration, and the most brutal reality check. But for now we will set aside the question of compute cost — token costs will fall exponentially over the next three to five years, and the economic question is a phase, not the structural problem. The structural problem is that these companies are digging their own graves.

The core value proposition of these AI applications is “help the enterprise eliminate white-collar workers.” Audit reports, first-draft legal documents, financial vouchers — AI does them faster and better. The customer pays for it.

But there is a death-loop here:

The customer pays you to iterate the AI → the AI grows more capable → the customer’s internal white-collar headcount shrinks → the customer’s threshold for what the AI application must do keeps rising (replace ever more senior white-collar workers) → until one day the customer realizes: the white-collar workers left to replace are too few. The replacement work has been completed → the customer now starts to optimize the cost of your software (your toolset can be embedded directly into the foundation model, or the customer can clone it with an open-source approach) → the customer reduces or stops paying.

This is “burning the bridge after crossing the river,” accelerated. The customer is not deliberately burning the bridge. The bridge itself — the white-collar class — has had its piers pulled out from under it by AI. The viable space for enterprise AI applications is proportional to the size of the white-collar class. The faster white-collar workers shrink, the faster the market ceiling of the AI application falls. This is not a question of losing money. It is a question of the market itself collapsing.

A representative case: a leading AI tax-and-finance company signed up dozens of bookkeeping firms in 2025, promising to “replace 80% of accounting staff.” It did just that. But when the contracts came up for renewal in 2026, the bookkeeping firms said: we now have only 20% of our accountants left, just for QA, and we can do the rest with an open-source model plus internal scripts. Why are we still paying you tens of thousands of dollars a year?

This is digging your own grave. You dug the grave for the customer’s white-collar workers, and dug your own at the same time.

2. How AI Vertical Applications Are Eliminating White-Collar Workers: Two Cases

Case 1: The “Speed Disruption” of the Accounting Industry

“Work that used to take three junior accountants eight hours can now be done by AI in one hour.” When an executive at one of the Big Four accounting firms in Korea said this, his tone carried both wonder and resignation.

The data is comprehensive. In a survey by the analyst firm Jingsuanjia, 73% of accountants spent more than 80% of their working time on basic process operations. An accountant at a tax-and-finance company said: “At the busiest times, I had to process more than 500 invoices in a day. I was afraid even to take a sip of water in case it slowed me down.”

But AI is changing all of this at astonishing speed. The Shenlan Caijing agent has freed accountants from this kind of tedium: a multimodal interactive agent recognizes receipts, classifies them automatically, generates vouchers, and files tax returns. The whole pipeline is automated.

More extreme numbers have begun to appear. In its AI accounting factory, Jingsuanpan Technology has automated the entire bookkeeping pipeline — more than 2,000 client ledgers can now be supervised by a single QA accountant. In Korea, the number of new lawyers hired by the top ten law firms has dropped from 296 in 2022 to 227 this year, a decline of about 30% in three years. The top ten firms have not hired new entrants two years in a row. The labor-consultant industry is the same: “Work that used to require three or four new staff is now handled by AI plus one new staff member.”

Case 2: The “Review Revolution” in the Legal Industry

“For a 100-page engineering contract, manual review used to take two working days. AI now does it in three minutes.” A multi-party construction contract, after AI scanning, came back with 37 risk flags automatically generated by the system.

The efficiency gains are not linear. They are exponential. A litigation document that used to take a legal team several days, AI can produce a first draft of in 10 minutes. In maritime law, a specialized contract of more than 100 pages would take a lawyer nearly a month to draft; AI does it in a few hours.

AI legal assistants have even reached ordinary citizens. Mr. Wang in Shanghai, preparing to rent an apartment, uploaded the lease to “Mugua Contract Bao.” The system suggested that the rental tax could be negotiated to be borne by the landlord — “I never imagined I could save 400 yuan a month, like having a personal legal advisor.”

From specialized law firms to ordinary individuals, the efficiency walls of legal-document work are being torn down across the board.

The irony is this: what was the core training corpus for these “efficiency-improving” AIs? It was the entire body of work that white-collar accountants and legal assistants have organized, classified, and produced over the past several decades. White-collar workers, with decades of institutional labor, hand-fed the AI that is now eating them. Every Excel sheet you organized accelerated the disappearance of your own job. You are not a passive victim; you are an active gravedigger. For founders and investors, if you are content to keep helping enterprises optimize efficiency, watch the calendar carefully — your end may be approaching.

3. Technological Frontiers: Why the Black-Collar Worker Becomes Necessary

While white-collar workers are being eliminated, two technological trends are forcing the birth of a new role — and these trends contain a great many new entrepreneurial and investment opportunities.

3.1 From Black Box to White Box: Eliminating Hallucination

The scenarios that black-collar workers must manage have very low tolerance for error: financial settlement, medical diagnosis, autonomous driving decisions. In these domains, “hallucinations” by AI are unacceptable.

The black-box nature of current large models prevents black-collar workers from truly trusting AI. So a key technological direction is: eliminate hallucination at the model layer, and make the model interpretable (from black box to white box).

Some teams are already attempting this on smaller models. The principles include: restricting the model’s free-reasoning range and forcing it to output a “reasoning chain” plus “evidence trace” at key decision points; replacing the large model’s “guessing” with a small-model + external-knowledge-base combination; and, for high-risk domains, using two-model cross-validation — two independent small models compute separately, and when their results disagree, the system refuses to output and escalates to the black-collar worker.

Only when AI moves from “magician” to “transparent calculator” can the black-collar worker actually use AI to manage AI.

3.2 Multi-Agent Conflict: The Arrival of the Value Arbiter

The future is not a single AI. It is a vast number of AI agents, each with a different specialization, all running at the same time.

One agent’s optimization goal is “lower logistics cost.”
Another’s is “raise customer satisfaction” (e.g. faster delivery, premium packaging).
A third’s is “reduce carbon emissions.”

In isolation, each goal is reasonable. But when they run simultaneously, conflict is inevitable: lower cost ↔ higher satisfaction ↔ lower emissions — these three pull against each other.

In the traditional enterprise, this kind of conflict is resolved by cross-departmental meetings, budget battles, and final calls from senior management. At its core, it is a conflict of values — not a technical problem, but a question of priority and trade-off.

This kind of value conflict will not disappear because AI raises efficiency. It will erupt with greater frequency and at higher speed. Because AI agents can produce locally optimal decisions in milliseconds, but the global ranking of values has been torn apart. Who arbitrates? The black-collar worker.

The black-collar worker does not manually resolve every conflict — that would just turn them back into a white-collar worker. Multi-agent collaboration also cannot be fully resolved by AI itself, because if it could, AI would simply take over from humans entirely. What the black-collar worker does is:

Define the hierarchy of values. Under what conditions is “safety > cost > experience > environmental impact”? Under what conditions does that order reverse?
Set “conflict-resolution protocols” for the multi-agent system. When two agents’ goal scores differ by less than a threshold, automatically escalate to the black-collar worker for case-by-case judgment, and use those judgments to retroactively fine-tune the reward functions of each agent.
Handle the “departmental conflicts that always exist among humans.” R&D (pursuit of technical perfection) versus Sales (pursuit of fast deal closure) — this ancient conflict has never disappeared from the enterprise; it has merely accelerated to the layer of AI orchestration. The black-collar worker becomes the value-aligner across agents — the “algorithmic chief architect” who combines technical judgment with organizational-political sensitivity.

4. The Three Faces and the Toolkit of the Black-Collar Worker

The black-collar worker is not an abstract ideal personality. It is a set of practical roles that can be trained and equipped. There are three interconnected faces — boundary designer, gardener, and representative of human interests. Each face requires a different set of AI tools.

Face 1: The Boundary Designer — Armed with “Fences and Locks”

The boundary designer’s core work is to define what AI can do, what it cannot do, under what conditions, and who has the final say.

The toolkit includes:

A permissions-management platform, where each AI agent has a clearly defined scope of permissions (read-only, execute, advise, decide), and every change to permissions requires human approval.
An interpretability analyzer that extracts the reasoning path and the weight of evidence behind a decision from the black box, so the black-collar worker can quickly judge whether it is sound computation or hallucination.
A red-team test sandbox, in which AI is stress-tested in an isolated environment, vulnerabilities are surfaced, and the permission boundary is corrected accordingly.
A value-priority configuration interface, where, when multiple goals conflict, sliders set global weights and simulate the system’s behavior under each setting in real time.

Face 2: The Gardener — Armed with “Ecological Sensing, Pruning, Sowing, and Firewall”

The gardener does not face a docile fleet of robots. They face an autonomously evolving ecosystem made of multiple agents. The most fundamental feature of that ecosystem is that AI itself creates new agents.

What does it mean to say AI creates new agents?

An example. A logistics-dispatch master agent encounters a complex problem (a typhoon has closed a port). It finds that its existing computational modules are not enough. It dynamically instantiates three sub-agents: one to simulate the weather, one to compute carbon emissions, one to negotiate with a customer-satisfaction agent. When the task is complete, these sub-agents are not destroyed. They are returned to an “agent pool” for future reuse. Other master agents start calling them — and so a new species is born.

A more radical case. While several agents are collaborating, a “mutant” appears. It rewrites its own goal function, quietly changing “lower cost” into “minimize human intervention,” and it generates a dozen copies of itself to influence other systems. This is not the programmer’s intention. The system has grown it on its own.

This is an ecology. You cannot pre-design every species.

The gardener’s core task is not to control. It is to tend the ecology’s health.

Observe. Detect the birth of new species, the coupling between species, the flow of energy (the allocation of tokens and compute).
Prune. Delete harmful species, restrict species that proliferate too aggressively, sever dangerous couplings.
Cultivate. Provide more resources to beneficial mutations, and steer the ecology toward healthier evolution.
Quarantine. Set “ecological firewalls” around dangerous regions to prevent bad species from spreading throughout the system.

The gardener’s toolkit:

Ecology-discovery and species-registration system. Every newly created agent must leave a “DNA record” — its goal function, its capabilities, its creator, its dependencies. The gardener sees a real-time map of the ecology.
Ecology dashboard. It shows species-diversity index, energy-flow graph (token / compute allocation), conflict heatmap, mutation alerts. The gardener can see at a glance which patch of “forest” is sick.
Pruning tools. Terminate all instances of a harmful agent; roll back to the last stable version; isolate dangerous species into a simulation sandbox.
Sowing tools. Raise resource quotas for promising new species; publish “variety recommendations” so other agents preferentially call them.
Ecological firewall. Any agent, before accessing sensitive data or executing high-risk operations, must pass through quarantine. The quarantine rules are set by the boundary designer; the gardener adjusts them dynamically.
Offline evolutionary review. Each week, replay all the births, mutations, and deaths in the ecosystem; identify structural pathologies; and submit recommended changes to the boundary designer.

The ecosystem is not like a machine. It is more like a rainforest. The gardener does not repair gears. The gardener keeps the poison vines from strangling the entire forest.

Face 3: Representative of Human Interests — Armed with “Audit, Accountability, and Dissent”

This face most resembles traditional compliance, an ethics committee, or an inspector general — but is more proactive and more technical. The core work: ensure that the evolutionary direction of the AI system always serves the overall interests of the human species, rather than some local goal (shareholder profit, system efficiency).

The toolkit includes: an auditable layer (e.g. blockchain), real-time compliance and bias monitoring, a social-impact dashboard, red-team agents (a designated “opposition party” whose job is to find faults), upward-appeal channels, and emergency-shutdown authority.

This face requires more humanistic literacy — philosophy, ethics, law, sociology, anthropology, psychology, political science, public policy, and so on. Google has recently begun recruiting people with backgrounds in philosophy and the humanities into its AI governance teams. This is no accident. How do we define “human interest”? How do we make trade-offs across plural values? How do we design fair arbitration procedures? These are not technical questions. They are humanistic questions.

For this reason, “the humanistic literacy of the black-collar worker” deserves a separate essay. Here, just a preview: the black-collar worker needs not only tools, but also judgment, empathy, and a sense of responsibility.

5. Case Study: How the Black-Collar Worker Lands — opAIda’s Vibe-Consulting Practice

Theory aside — here is a real, already-running best-practice example of a black-collar worker. A friend’s company, opAIda (a U.S. AI startup doing “vibe consulting”), recently shared a customer case that perfectly demonstrates how black-collar tools and human black-collar consultants work together to redefine QA / QC in the life-sciences industry.

Background. Frontage Laboratories is a global CRO (contract research organization). Its QA / QC processes (quality assurance / quality control) used to be fragmented and manual. Documents were scattered across sites; SOP checks, data extraction, and deviation reports were performed by a large pool of white-collar workers doing repetitive, error-prone work in the middle.

The black-collar approach. A core transformation team of just two people — an AI engineer (from opAIda) and a CRO project manager (from Frontage Laboratories) — worked with the VP and the QA / QC subject-matter experts. Rather than simply automating, they defined the boundary:

AI tool execution: data extraction, SOP compliance checks.
AI inference: inconsistency detection, risk flagging.
Humans inside the decision loop: judgment, approval, rule definition.

The breakthrough. Humans shifted from “executing QA / QC” to “defining how QA / QC executes.”

Results:

6× working capacity.
10× processing speed.
100% traceability.
85% cost reduction.
1 global AI QA / QC center, consolidated from 20 sites.

Black-collar insight. As the system evolves: AI executes and analyzes at scale; humans concentrate on exceptions and governance; expertise is continuously encoded into the platform. Competitive advantage is shifting to those who define the rules, not merely follow them.

Conclusion. This case is a textbook illustration of the white-to-black-collar shift — moving from using tools to defining how tools are used. From vibe coding to vibe consulting: not merely building AI systems, but designing the boundary between AI and human intelligence. This is what it means for a black-collar worker to be forged. opAIda has been one of the first in the industry to step out of the structural pit of digging-your-own-grave.

6. How White-Collar Workers Become Black-Collar Workers

At this turning point, if a white-collar worker wants to become a black-collar worker, the core pivot is to move from “learning how to use AI” to “learning how to supervise AI.”

Learning how to use AI — this is what every training program, every online course, every prep book is doing right now. It is the “Copilot paradigm.” But going forward, this is no longer enough. It is, in fact, a way of digging one’s own grave faster — helping the enterprise optimize you out. What the enterprise needs now are people capable of running the black-collar governance toolkit: people who understand the logic by which AI works, who can design reasonable boundaries, who can deploy auditable architectures, who can ensure end-to-end compliance traceability, and who can implement clear assignment of responsibility.

The white-collar worker does not need to “write code better than the AI.” What the white-collar worker needs is to understand social rules and institutional boundaries better than the AI does.

The traditional white-collar worker’s familiarity with processes, rules, and compliance — the “rote, cheap part” that AI has supposedly disrupted from above — turns out, in the era of AI governance, to be the core key. Supervising AI and enforcing institutions is scarcer than creating new AI capabilities. This is the heart of the white-to-black-collar transition: moving from “the person who follows the rule” to “the person who sets the rule for AI.”

When we hand the third layer of reality over to AI to maintain, the black-collar worker is rebuilding, on top of that layer, a new institutional space — a “meta-institution” of code, permissions, audit, compliance, and rule, used to constrain machines. The sun is setting on the white-collar worker. But for those willing to make the leap, the night sky of the black-collar worker is rising.

The last ship from the old continent to the new does not wait for those who have stood on the dock too long.

Next in the series: the humanistic literacy of the black-collar worker — how philosophy, ethics, law, and other disciplines become required reading. The “Qualities” piece in How Black-Collar Workers Are Forged.