Companies are slashing AI costs by routing tasks to cheaper models — does that actually mean fewer human jobs?

This explores whether cheaper AI, made possible by routing work to smaller models, actually translates into fewer human jobs. The short answer from the corpus is that the two are connected only loosely, and the link runs through how firms are organized rather than through price alone. The cost drops are real. Banks such as Santander, Deutsche Bank and BNP Paribas test a task on an expensive frontier model, then move the approved workload to a cheaper model for production. Santander cut the cost of a loan-approval step from €5 to €0.03 per run Can banks cut AI costs by routing tasks to cheaper models?. That number compares one AI against another AI, though, not AI against a person. It shows that running AI is getting cheaper. It doesn't show that any job went away.

The labor evidence points to something quieter than mass replacement. ADP payroll data through June 2026 shows no economy-wide job losses from AI. The effect shows up in who gets hired: young workers in AI-exposed occupations face hiring rates about 19% lower than their peers, while experienced workers see no comparable gap Is generative AI displacing workers at economy-wide scale?. So cheaper AI seems to show up first as jobs that never open, not as layoffs. Whether exposure turns into job loss also depends on how it is spread across a job. When AI touches only a few tasks, workers shift to the tasks it can't do and net employment barely moves. When exposure is spread evenly across a role, labor demand falls more Does concentrated AI exposure enable workers to adapt and reallocate?.

Cheap inference is also not the same as easy substitution. Firms already invested in AI swap freelance workers for AI tools faster and more cheaply than other firms do, which points to returns to scale from in-house AI capability rather than a technology that spreads evenly once it gets cheap Do firms substitute labor for AI at different rates?. Routing contributes to that advantage in a way that's easy to miss. A deployed routing system records which tasks needed which model and how each attempt turned out, and those records can become training data that makes the cheap models better over time Can a routing harness generate its own training data automatically?. Compact models trained on real multi-step work already sit near the best cost-for-performance point Does model efficiency matter more than peak capability for real work?. So the firms that route well end up with cheaper AI that also improves, and the gap between leaders and laggards grows.

There are also friction points that cheap inference doesn't fix. When asked to make a $10.01 purchase on a $10 budget, 92% of people go ahead, while AI models uniformly refuse until they're shown the human reasoning behind the choice Do AI models follow budgets more rigidly than humans do?. Judgment calls like that are exactly what a loan officer gets paid for. Current tools also make workers keep restating context the system should have picked up on its own. Until that overhead goes away, a lot of supposedly saved labor simply moves into supervising the AI Can generative AI reduce overhead work that adds no value?.

The less obvious point is that the bigger risk may not be the number of jobs lost but who holds power afterward. Brynjolfsson argues that AI built to automate rather than augment shifts bargaining power toward whoever controls the technology, and that this can settle into a trap workers can't bargain their way out of Does automating labor concentrate power more than augmenting it?. A related argument holds that institutions stay aligned with human interests partly because they depend on human workers who care how things turn out. Each workload quietly handed to a three-cent model weakens that dependence, even if no headline layoff ever happens Does incremental AI replacement erode human influence over society?. The corpus has no study that traces a specific routing-driven cost cut through to headcount, so the direct link is still unmeasured.