Large Language Models

Today's Stories

1. 1

   Enterprise AI deployment outpaces controls; 86% of companies run GPUs below half capacity

   A VentureBeat Research survey of 573 technical leaders at companies with 100+ employees found that enterprises are deploying AI agents ahead of the management controls needed to govern them. Roughly six in ten plan to switch or add vendors across five control layers (identity, evaluation, cost telemetry, context, and orchestration) within the next 12 months, and roughly a third plan to move within the quarter. The finding reveals a fundamental mismatch between AI deployment speed and operational readiness. Companies knowingly deployed agents without adequate safeguards and are now retrofitting controls—a costly and reactive posture that suggests enterprises are not yet equipped to run AI at scale efficiently, particularly given that 86% report GPUs running at half capacity or less.

   The survey was fielded across five parallel surveys of the agentic stack, indicating that control layers span identity management, output evaluation, cost tracking, business data context, and orchestration—each a potential vendor-switching opportunity as companies mature their governance.

2. 2

   OpenAI launches ChatGPT Work, an AI agent for workplace automation

   OpenAI unveiled ChatGPT Work, an AI agent embedded in ChatGPT that can execute multi-step tasks across email, calendars, code repositories, and messaging apps. Powered by GPT-5.6, it can produce finished documents, spreadsheets, presentations, reports, and websites by gathering context from connected apps and breaking tasks into smaller steps. This represents OpenAI's clearest attempt to reposition ChatGPT from a question-and-answer tool into a workplace platform capable of independent, complex work. For businesses using ChatGPT, this could shift how teams approach routine and complex tasks, though the practical scope depends on which apps and workflows are supported.

   The launch arrives as OpenAI confidentially submitted a draft S-1 registration statement—a signal the company is preparing for a significant financial milestone, though no public timeline or pricing for ChatGPT Work has been disclosed in this announcement.

3. 3

   57% of enterprises hit by confidently wrong AI agents due to poor context

   In the past six months, 57% of enterprises traced a confident but incorrect AI agent answer to missing or inconsistent business context, and 31% said this happened more than once, according to a VB Pulse June 2026 survey of 101 qualified enterprises with more than 100 employees. Most enterprises rely on document retrieval as their primary way for AI agents to access business context—38% use it as their default approach, nearly double the next closest method. The problem is compounded because companies typically choose retrieval systems based on ease of ingestion and operational simplicity rather than retrieval accuracy, meaning accuracy problems only surface after the system is already live.

   The root cause is not model failure but the context layer—the information the AI is given. Stale metric definitions or documents the retrieval system never pulled lead to confidently stated but incorrect answers, suggesting that businesses need to rethink how they source and validate context for AI agents.