Axios: Artificial intelligence already costs more than staff at some companies

Companies that were waiting for artificial intelligence to deliver quick savings are getting the opposite effect: in many cases, AI turns out to be not a replacement for expensive labor, but a new costly budget line item. Instead of the promised cost reduction, businesses face growing bills for models, cloud infrastructure, integrations, and quality control of results. For some teams, this is no longer an experiment, but a financial surprise that changes calculations for hiring, automation, and product development.

As Axios notes, in a number of companies, the costs of maintaining AI solutions are becoming more noticeable than the payroll of employees performing similar tasks manually. In practice, the problem rarely comes down to just subscribing to a model. Money goes to API calls, processing large volumes of data, storage, fine-tuning, protecting corporate information, as well as developers and analysts who must integrate AI into real processes.

If the system operates around the clock or serves a large flow of requests, infrastructure costs begin to grow faster than expected at the project's start. That is why financial directors increasingly demand from AI teams not beautiful demos, but proven returns on each scenario. Part of the overspend is explained by the fact that pilots are almost always calculated too optimistically.

At the beginning of a project, a company tests one scenario on a small group of users and gets acceptable unit costs. But after scaling, request queues appear, capacity reservation, additional integrations with CRM and document management, licenses for multiple teams, and monitoring costs. If sales, support, marketing, and internal operations connect to the service simultaneously, the total bill grows non-linearly.

What looked like an affordable experiment quickly becomes a permanent operating expense in industrial deployment. This is especially painful for companies that bought AI as a tool for instant optimization. In presentations, automation looks like a way to reduce manual labor, speed up responses to customers, and lighten the load on teams.

But in real operation, it turns out that the model must be constantly tested, constrained, retrained on internal scenarios, and backed up by humans at critical stages. In other words, AI does not always remove an employee from the chain: often it adds another layer of work, where a person checks, corrects, and takes responsibility for the result. When you add requirements for security, legal review, and data quality, the economics of implementation become even more complex.

For regulated industries, large corporations, and services with high error costs, it is not enough to simply connect a model and wait for an effect. You need to audit responses, monitor for leaks, control access to data, compare model versions, and measure where automation really saves money and where it only creates an appearance of progress. Therefore, the most expensive element of a project is often hidden not in the model itself, but in the layer of support around it.

This does not mean that AI is overvalued or that business will massively abandon such systems. Rather, the market is moving out of the phase of naive expectations into a phase of sober calculation. Companies are beginning to look not at general promises, but at the cost of one useful action: a processed document, a prepared response, a closed ticket, or a successfully completed operation.

Where AI truly reduces cycle time and errors, it will stay. Where automation is more expensive than a person or requires constant manual backup, projects will be reconsidered, simplified, or stopped. The main conclusion for business is simple: implementing AI can no longer be considered a universal cost-saving measure.

Without precise unit economics and strict expense control, even strong technology quickly becomes one of the heaviest cost centers.