AI has a strong encoding capability that creates derivatives from an object without hardcore friction or learning curves, which indicates inefficiency premium may no longer exist in the future, regardless software, SaaS or white-collar jobs.

Wage Premium

In labor market, wages are composed of skills and frictions. Here are lists that the productive payrolls and frictions that enterprises have to pay.

Human Labor Moat

Why labor supply is limited

Most skills require years of accumulation of uncompressed, enumerative experience duo to

• Complexity: aviation pilots, firefighters, surgeons, who operate complex equipments.
• long tail & assurance: soldiers, bookkeepers, infrastructure engineers, who are responsible for rare cases
• Long feedback-cycle: financial, biologics, law & tax, policy, as the results may delay after years.

With time constraints, the labor supply is limited, and the society has been working on labor shortage: education, immigration, prolonged retirement.

Diminishing Returns on skill acquisition & upgradation

The diagram shows the most common case between the skill (the bar) and wage(the line). The skill follows a concave learning curve to capability, while the wage follows a S-shaped line depending on the demand relationship.

• Training: From the begin of career, beginners will take 1-3 years to get workspace training on incorporation into their workspace culture and internal processes, and the wages show underpaid as the internal tools & processes are not transferable.

• Reprice: After the skill accumulated, the wage might soar alias with the job market demand, as the new companies must pay premium to get an experienced labor instantly.

• Plateau: When the skill stabilized, individuals may not continue on investing on their skills for higher pays, as the career opportunities may have stagnated: lack of high-end positions, politicized roles, leaderships, work & family balance, primary accumulation of capital, ageism.

Inefficiency Premium in Organizational Frictions

In corporates organizations, employees are not always working on core productivity, instead they must follow on the rigmarole management framework (like L2C, IPD, Agility) as cogs inside a machine.

Operational Frictions and Modern Process Frameworks

The modern frameworks decompose the complex, abstract problems into standardized, observable, executable process stages, making it possible to help domain isolated employees working togeher, but with more frictions and delays.

1
2
3
4
5
6
7

Workflow
 ├─ Financial: Record to Report
 ├─ HR: Recruit to Retire
 ├─ Sourcing: Source to Pay
 ├─ Sales&Service: L2C/LTR
 ├─ R&D: IPD/DevOps/Agile
 └─ Machine–Machine (data pipelines, BI)

For those problems, there has been mitigation on productivity, like digital office tools, ERP SaaS, data lakes or other IT automations in corperates. Most of them are “the Big Excels”.

1
2
3
4

Mitigation
 ├─ Human–Machine (ERP, CRM, tickets)
 ├─ Human–Human (meetings, chat, Email)
 └─ Machine–Machine (data pipelines, BI)

These services may leverage on AI for less operational frictions, as plenty of SaaS companies have declared their agentic solutions.

Gate-keeping as Friction

Despite AI-acceleration productivity spikes, decision roles are not always with maximizing speed for compliance reasons. When forcely applying AI-based processing on the gate-keeping works, there could be

• governance backpressure on volumes: uncertainty, traceability, accountability and liability
• Transparency & trust conflict: entrepreneurs prefer the single source of truth (SSOT), but employees may be unwilling to be supervised. There should be a gray zone on nuances and seemingly roles.
  • duo record systems - employees may keep hiden Excel files for true work records, but input delay or AI-reshaped data for KPI
  • uncontrolled computing at the edge - junior staff can produce executive-level analysis & dashboards
• Career incentives: for who preparing from retirement, the only consideration is obedience, while new graduates may focus on innovations.

These kind of intentional frictions may be at a prisoner’s dilemma

• Resistant to AI => Redundancy: bureaucrat, left behind other companies & countries
• Accept AI => Taylorism: KPI dashboards, permanent auditability, accountability, screwed, micro-monitoring and peer pressures.

Polarization on Jobs

Verification Bottleneck on biological limitation (Unresolved)

With AI involved in the job, code or document generation may be cheaper and faster, human work may shift into the search and verification, which may be similar to EDA replacement on handcraft circuit in 80s.

Unlike EDA in a scoped mathematical system with format verification, AI can’t guarantee same outputs in nature language. The question arises when it comes to the lag between high throughput AI productivity with human biological limitations.

Which is the right sweat spot for human when trusting or verifying AI artifacts durability?

• the performer test - thousands of blackbox test cases
• disposable artifacts - when verification cost > replacement cost
• ritualization - the worst case like aviation crash

Jobs middle collapse (Facts)

In the last few months, I have almost attempted to transfrom my job into agentic flows and skills in the coding development. Here are something interesting.

• Information Security
  • Static taint analysis: Digging hundreds of XSS/SSRF risks from source code with qwen3.5-27b
  • But comprehensive defensive armor is always required, like supply chain protection over axios vulnerability
• IT outsourcing
  • For entry & middle level engineers, especially who only understand one language but not capable on others skills like linux, they may be constrained inside their own circle of competence despite working with advanced models.
  • Consulting business: outcomes must be proven rather than metaphysical ideas.

After all, all patten match jobs that own limited responsibilities may be distilled first, or the profitability may stagnate.

Long term jobs (Future)

As AI tokens are as cheap as cellular data, jobs may shifts towards high compression ratio or abstraction tasks

• steeper learning curve engineer: exception handling, unpredictable,self-reinforcing feedback loop, slow feedback system.
• MoE knowledge base: learn philosophy, being sophisticated, generalist for problem framing capabilities

In general, we are manage digital employees like the book High Output Management

One time digital transformation on connectivity on business

• AI empowers domain experts to bypass engineering bottlenecks.
  • software maintenance: software engineers may maintain the AI-generated code by experts.
  • bounded automation: liability engineering, explainability or liability insurance
• AI infrastructure
  • build, test, maintain, gateway, sandbox, firewall, multi-level agent management.
  • Less expensive models: Replacing repeated works with cheaper models
  • data aggregation and bi-verification among isolated IT platforms, which will make the corperate more transparent.

Despite AI leveraging, not all friction and uncertainty is unavoidable, but structurally necessary.

• Time decay knowledge with strong penalty: continuous data maintenance on data aggregation

  • Insurances: security, health
  • Laws: tax, legal precedents, visa apply

• Time itself

  • Emotional & ritual Labor: it has happened on Japan even without AI enrollment.

• AI entertainment that may replace video or gaming

Bounded automation on AI in enterprise

In organizations, speed is not always everything. Customers buy confidence, not code. From my limited AI-driven developer view, there will be

• Limited productivity on AI-powered taylorism
  • Ideal for starting from scratch for elite teams
  • but when handling legacy problems, it’s not just an engineering problem, for instance, lack of reason on blackboxed-code.
• The control tower is the GOAT: AI is changing the game (increases observability), but not the rules
  • AI do the paper works first, then gets reviewed by certified expertises for nuances
  • A certificate old school GUI is better than AI generated dialog

Questions (TBD)

• is a failure experience compulsory?

  • learning swimming without entering water
  • driving a car in an emulator

• prior knowledge limitation

  • can you outsource your thinking?
  • can you verify outdated outputs?
  • can you prove your innocence of reasonable reliance on AI generated artifacts like laws & taxes?

• leadership & management bandwidth on human & AI

  • verification: can you replicate yourself & your workflow with AI?
  • entropy: can you manage 6 or more real-world human employees?
  • are you responsible for unlimited responsibilities in limited background knowledge?