• OpenAI is probing elevated error rates specifically affecting ChatGPT for business and enterprise customers, mirroring recurring service disruptions.
  • The incidents disrupt enterprise workflows in high-stakes sectors, with error rates varying widely across domains from 5% to 83%.
  • Short-term fixes are expected, but long-term reliability concerns persist amid growing enterprise AI adoption and regulatory pressures.

OpenAI is investigating elevated error rates specifically affecting ChatGPT conversations for business and enterprise customers, according to people familiar with the matter. The probe comes as the artificial intelligence company grapples with persistent service disruptions that have plagued its systems since ChatGPT's 2022 launch.

Recent status updates indicate ongoing investigations into similar latency and error issues across API and ChatGPT services, though this latest incident appears particularly focused on premium business clients. One enterprise customer, who requested anonymity due to confidentiality agreements, described "catastrophic failures" in critical workflows, including memory loss and hallucination issues that emerged following backend updates.

"What enterprise customers really need is regulatory stability in their AI systems," said a technology consultant working with multiple Fortune 500 companies using ChatGPT Enterprise. "When these outages hit during business hours, they disrupt everything from customer service operations to internal coding projects."

OpenAI's status page logged an elevated error and latency event on May 23, 2024, that impacted API and ChatGPT broadly before being resolved after fixes were implemented. Similar disruptions occurred March 8, 2024, lasting approximately five hours, and again on August 15, 2024, with bad gateway errors affecting thousands of users. A more recent incident in late 2024 generated over 15,000 reports via Downdetector before being fully recovered by 11:16 PM ET, caused by problems with an upstream provider.

Community forums note persistent errors like unprocessed requests and memory failures following backend updates, with business users reporting particular frustration around degraded long-conversation performance. While casual users might notice occasional hiccups, enterprise stakeholders are demanding better service level agreements and more transparent communication about ongoing issues.

Efforts to improve reliability have hit snags as OpenAI scales its infrastructure to meet surging demand. The company's annualized revenue surpassed $4 billion in 2025, driven largely by enterprise subscriptions and API usage, but profitability remains challenged by high compute costs. With a valuation exceeding $150 billion as of late 2025, the pressure to maintain service quality while expanding capabilities creates constant tension.

Industry analysts point to broader market trends affecting all major AI providers. "You're seeing similar latency complaints at Anthropic with their Claude API and Google with Gemini spikes," noted an AI infrastructure specialist. "Global data center shortages and the complexity of maintaining these large language models mean occasional disruptions are almost inevitable."

Regulatory pressures add another layer of complexity. The EU AI Act enforces reliability standards for high-risk systems, potentially putting additional compliance burdens on OpenAI's enterprise offerings. U.S. executive orders on AI safety similarly emphasize the need for dependable systems, particularly in business and economic applications where error rates have improved to 15-20% from previous highs around 50%.

OpenAI representatives did not respond to multiple requests for comment on the current investigation timeline or specific remediation efforts. The company's standard communication has been through status page updates, which typically note when issues are "identified" and "monitored" before resolution.

Short-term, experts expect quick fixes similar to past incidents, with monitoring and backend adjustments to stabilize service. Long-term, however, the variability in error rates across different domains—from coding (10-50%) to healthcare (up to 83%)—suggests fundamental challenges in creating consistently reliable AI systems for business applications.

Correction: An earlier version of this article stated error rates had improved to 5-20% across all domains. The correct range is 5-83% depending on specific applications, with business and economics showing 15-20% error rates in recent testing.