Costlier AI would change which tasks a business automates
A compute shortage could change the economics of some tasks. Employment outcomes also depend on workflow design and demand, so a higher token bill is no hiring forecast.

If an automated task becomes more expensive, a business has to decide whether to keep running it. It may switch models if another configuration can meet the service standard at a lower cost. Reducing output is another response. Some work could return to an employee. Each response is possible. A higher compute bill alone does not tell us which one the business will choose.
That makes compute scarcity a useful scenario to examine, with limits. It can expose an automation project that only works under a narrow price assumption. It cannot support a precise forecast that displaced jobs will return once AI providers face enough pressure on electricity supply.
The energy pressure is real enough to study without inventing a global rationing event. The IEA's 2026 assessment of energy and AI reports that data-center electricity demand grew by 17 percent in 2025. Its central outlook keeps demand growth near the trajectory in the previous report. These are system-level observations and projections, with uncertainty about how supply and demand develop.
A local power constraint is not a retail price
A data center needs an electricity connection in a particular place. A shortage at that location can delay capacity even when another region has power available. The IEA's October 2026 report on integrating data centers examines their interaction with local electricity systems. This is a more specific issue than saying that the world has run out of energy for AI.
An application buyer experiences the constraint through a provider's terms and service. Those terms also reflect competition and the way the provider allocates its capacity. The path from a grid bottleneck to a price for one application is therefore indirect. A company should test the prices it would actually pay under a stated scenario.
For example, a team can ask what happens if its cost per accepted output doubles. That exercise does not require predicting why the increase occurred. It reveals whether the workflow still pays for itself and which assumptions have the largest effect on the result.
The accepted-output condition is important. A cheaper model might need more attempts before producing an answer a reviewer can use. Comparing only the price of one request can make that alternative seem more economical than it is. Include the time spent checking failed outputs in the comparison.
A task can return without a job returning
Suppose a business automates the first draft of a routine customer reply. If generation becomes expensive, a manager could ask existing staff to draft those replies again. That change restores a task to their workload. It does not automatically create a vacant position or restore the employment arrangement that existed before automation.
The manager could also simplify the service or handle fewer requests. A cost increase can reduce output. That possibility needs to remain in the scenario even when the argument is intended to reassure workers. An uncertain outcome should not be turned into a promise of rehiring.
The OECD's 2026 review of skills in the AI age describes substitution and complementarity as effects that can coexist. Their balance affects employment outcomes. That framework fits the business decision: some tasks may be automated while others continue to depend on a person, with the overall job changing around them.
A useful workforce discussion should therefore name the work being reassigned. Reviewing a generated response may require different time and expertise from writing the first draft. Moving that responsibility to a person without allowing for the review burden can erase the apparent saving.
Use observed results at the scale they support
There is evidence that assistance can improve performance in particular settings. The 2025 Generative AI at Work study examined 5,172 customer-support agents and found a 15 percent average increase in issues resolved per hour after access to an AI assistant. That is a measured result from a specific workplace setting, with effects that varied across workers.
The study gives a better starting point for a support-workflow question than a general prediction about all jobs. A different company still needs to test its own task. Its incoming issues may differ, and its review process may consume a different share of the apparent time saving.
For a price sensitivity test, keep the service standard fixed. Compare how much accepted work the current workflow completes under its actual staffing arrangement. Then calculate the cost at the higher assumed provider price. If the result becomes unattractive, identify a feasible alternative and test whether it meets the same standard.
The outcome may be a smaller automation project or a different division of work. Record which task moves to a person and how that affects their available time. That makes the staffing decision assessable before anyone turns a conditional compute scenario into an employment forecast.