Companies curb AI expenses and address reliability as costs, hallucinations and regulations tighten

Tech firms and other enterprises have recognized that indiscriminate AI usage – dubbed "tokenmaxxing" – generated unexpectedly high expenses. Reports cite internal competitions at firms such as Meta that produced tens of billions of tokens, prompting costly over‑spending and, in some cases, workforce cuts. In response, a new practice called “modelmaxxing” is gaining traction: firms deliberately match tasks to the most cost‑effective AI model rather than defaulting to the most powerful one. Start‑ups offering model‑routing software are emerging to automate this selection.

At the same time, businesses are confronting the reliability problem of generative AI. Studies show that up to 35 % of AI‑generated answers contain false statements, with hallucinated facts appearing in legal citations and corporate reports, leading to financial losses and legal risks. German firms report that three‑quarters of data leaders have faced operational issues from AI hallucinations.

Regulatory pressure is building as the EU AI Act moves toward full enforcement in 2026, requiring high‑risk AI systems to undergo safety certification by notified bodies before market entry. Companies must now plan for either self‑assessment or external conformity evaluation, adding another compliance layer.

The evolving landscape also includes privacy‑focused offerings such as Proton’s Lumo chatbot, which emphasizes zero‑access encryption and user‑controlled data retention, reflecting growing demand for trustworthy AI tools.