AUDIT FIRM FINDS $1.7 MILLION IN AI OVERBILLING ERRORS ACROSS 60 COMPANIES — CHARGED FOR WORK THAT NEVER HAPPENED A startup called Vaudit has published findings from an audit of $34 million in AI invoices covering 60 companies from March through June 2026, and the results paint a picture of systematic billing errors that enterprise customers had no way to detect on their own. The audit found $1.7 million in disputed charges, and Panasonic, HP, and Honda are among the corporations confirmed to have been affected. The billing errors were not random. Vaudit identified five recurring patterns: customers billed at premium rates for older or cheaper model versions; prompts that returned no output but were charged anyway; retry storms in which autonomous agents repeated failed requests and stacked duplicate charges; orchestration errors in which a cloud provider sent the same request to two models simultaneously; and general token miscounting that inflated usage figures. Most of the errors were traced to Claude Code invoices, though OpenAI billing was also flagged. Both Anthropic and OpenAI disputed the findings but ultimately refunded roughly 80 percent of the challenged charges across the accounts involved. That response itself is revealing. Companies billing at scale for token usage have an inherent information advantage over customers who lack the technical tools to audit their own consumption. Every enterprise that has ever trusted an AI invoice without a way to verify it is in the same position Panasonic and Honda were in. Most of them just don’t know it yet. Keywords: AI overbilling, Vaudit AI audit, Anthropic billing errors, OpenAI overcharges, token billing