The Automation Hype Ignores the Bill: What Really Decides AI Labour Substitution

The most consequential question about artificial intelligence (AI) and work is not whether a model can perform a task, but whether a company can replace a worker and still save money after counting electricity, computing, system integration, failed outputs, regulatory obligations and human review.

A study, "Capability Determinism, Energy and AI Labour Substitution," authored by Will Mbioh of the University of Kent and published in AI & Society, challenges what it calls "capability determinism": the tendency of governments, consultancies and media narratives to move directly from what AI can do to predictions about which workers it will replace, without testing whether substitution is economically, operationally or legally viable.

According to the author, firms replace labour only when the full cost of producing an acceptable unit of work with AI falls below the equivalent human cost. The calculation must include not only the headline price of an AI service, but also the electricity and computing required to run it, the expense of integrating it into an organisation, and the human supervision that many professional and regulatory systems will continue to demand.

The Automation Debate Is Mistaking Exposure for Replacement

Many forecasts identify tasks AI can perform and then map them onto occupations, which reveals where pressure may emerge, but not whether firms will remove workers.

Jobs are bundles of activities. A lawyer may research, draft, negotiate and accept professional responsibility. A doctor may interpret evidence, communicate uncertainty and remain accountable for treatment. AI may automate selected components while leaving the worker responsible for everything around them.

The author separates augmentation from substitution. When an employee uses AI to work faster, the technology is a productivity tool. Genuine substitution occurs only when a firm restructures the workflow around machine-produced output and removes the human labour cost.

Governments are discussing retraining, income support and tax reform for a future with less work, often before the economic case for substitution has been tested. Nearer-term effects may include job redesign, tighter monitoring, reduced entry-level hiring and a shift of tasks from junior to senior staff.

Occupational "exposure" should therefore not be treated as a forecast of job losses. Labour-market assessments must also consider firm size, task volume, wage levels, infrastructure, liability and deployment costs.

Cheap AI Still Rests on Expensive Energy Infrastructure

AI is commonly presented as software, but every output depends on specialised chips, data centres, cooling systems, electricity grids and semiconductor supply chains. Each time a model generates an answer, it performs an inference. More complex reasoning requires more tokens and computation. Agentic systems, which plan, act, check and revise across several stages, can consume far more power than a simple chatbot exchange.

It creates a structural cost floor. Algorithms may become more efficient, but AI cannot escape the price of electricity. Data centres must power processors and remove the heat they generate, making their economics sensitive to energy prices, grid capacity and utilisation.

The author also questions whether current API prices reflect the long-run cost of providing AI services. Providers may be charging aggressively while investors fund losses in anticipation of future scale. If capital markets demand sustainable returns, or operating costs remain high, businesses may face higher prices than current automation models assume.

The paper identifies a paradox: methods that improve AI performance can also increase computing requirements. Better capability can make substitution more attractive in theory while raising the cost of each completed task.

For developing economies, the trajectory may differ from global forecasts. Where electricity is costly or unreliable, infrastructure is weak and advanced computing must be imported, replacing workers may be less economical. Lower wages further reduce potential savings. AI adoption could therefore widen geographic divides. Firms with cheap power, strong grids and abundant capital may automate faster, while lower-income markets use AI mainly to complement workers.

Integration and Supervision Can Erase the Savings

A low-cost subscription is not an enterprise automation system. Replacing work at scale requires organisations to connect AI to internal data, clean records, establish permissions, build audit trails, redesign processes, train staff and satisfy security and compliance requirements.

Those fixed costs can be substantial, and viability depends on scale. A large customer-service operation may have enough repeated work to justify the investment. A smaller firm with varied, low-volume tasks may not. This helps explain why substitution is more plausible in standardised activities than in complex professional work.

Agentic systems add uncertainty. Multi-step tools may use far more tokens than a single response, while errors can compound across stages. The relevant figure is not the cost of one successful model call, but the average cost of completing an acceptable task after failures, retries and human intervention.

Supervision may be even more decisive. In law, medicine, audit, finance and safety-critical software, qualified professionals remain responsible for final outcomes. AI can produce a draft or recommendation, but a human must still review and approve it. This creates a difficult business case. The highest-paid occupations appear to offer the greatest savings, yet they also face the strongest oversight requirements. Legal and professional accountability cannot simply be transferred to a model.

In some cases, identifying subtle AI errors may require nearly as much expertise as producing the work. The technology may reduce drafting time while leaving judgement, liability and sign-off with the professional. The paper's cost framework therefore rests on three components: variable inference costs, fixed integration costs and residual supervision costs. Any one can prevent substitution even when the model is technically capable.

The Future of Work Will Be Uneven, Not Automatic

The paper replaces a simple capability narrative with a realistic deployment test. It does not deny AI's disruptive potential; it explains why disruption will differ across firms, sectors and countries.

Substitution is most likely where work is repetitive, volumes are large, integration pathways are mature and oversight is limited. It will be slower in professions built around judgement, trust, licensing and legal responsibility. It may also follow the geography of cheap electricity and computing infrastructure.

The analysis has limitations. It is conceptual, not a firm-level empirical study, and it does not estimate how many jobs will be lost or retained. Some assumptions could change as chips improve, models become more efficient and enterprise tools mature. The argument should therefore be read as a corrective framework, not a definitive forecast.

Future research should compare fully loaded AI and human costs across sectors, firm sizes and countries. It should measure supervision time, failed-task rates, integration expenditure and whether pilot projects become genuine workforce reductions. Developing economies deserve attention because wages, grid reliability and access to capital may produce different outcomes.