Enterprise AI Wants Controls

Big companies are no longer asking artificial intelligence vendors for the smartest model in a demo. They are asking who can decide which employee gets which tool, who approved an action, and what got changed in a live system. (siliconangle.com) Anthropic said on April 9 that its Claude Cowork product is getting organization-wide controls for enterprise customers, including role-based access so administrators can choose exactly which capabilities each employee can use. Anthropic also added group spend limits so one team’s experiments do not blow through another team’s budget. (siliconangle.com) Claude Cowork is not a chat box that only answers questions. Anthropic describes it as an autonomous agent that can handle multistep tasks on an employee’s computer, including organizing files, creating reports, and running browser tasks. (siliconangle.com) That changes the buying problem for a bank, a law firm, or a manufacturer. When software can click around inside real systems, the useful feature is no longer just raw intelligence; it is the permission slip attached to every action. (siliconangle.com, oracle.com) Anthropic also expanded OpenTelemetry support so companies can monitor events such as tool calls and file modifications inside their existing security information and event management systems. In plain English, that means the artificial intelligence leaves a trail that the security team can inspect later. (siliconangle.com) OpenAI moved on a different lever the same day. It cut the price of its Pro plan for Codex to $100 per month, down from $200, while saying the new tier gives five times as much usage as the existing $20 plan for heavier coding work. (siliconangle.com) OpenAI is also reportedly preparing a cybersecurity product with tighter distribution instead of a wide public launch. Axios reported on April 9 that OpenAI was finishing a security-focused offering for a select set of partners, which mirrors Anthropic’s decision to keep its own cyber model behind a narrow gate. (axios.com, gizmodo.com) Anthropic’s Claude Mythos Preview shows why these gates are showing up. Anthropic said on April 7 that the model is being limited to a select group of companies for defensive security work through Project Glasswing because the same system that finds software weaknesses could also help bad actors exploit them. (cnbc.com, axios.com) The first Project Glasswing partners include Apple, Google, Microsoft, Nvidia, Amazon Web Services, CrowdStrike, and Palo Alto Networks, plus roughly 40 other companies. A model with that kind of access pattern is being sold less like a public app store product and more like a controlled lab instrument. (cnbc.com) Oracle’s launch on April 9 makes the same point from the other direction. Its new Fusion Agentic Applications for finance and supply chain are built to work inside existing approval hierarchies, permissions, policies, and transactional context, and Oracle says the software surfaces exceptions when human judgment would materially change the outcome. (oracle.com) Oracle listed 12 applications available now across enterprise resource planning and supply chain software, including tools for claims settlement, collections, and cost accounting close. Those are jobs where a wrong move can delay cash, distort inventory, or break an audit, so the selling point is controlled execution inside the rules, not a free-ranging bot. (oracle.com) Even the rivalry language is shifting. In a memo reported by CNBC on April 9, OpenAI argued that more computing power will lower the cost per unit of intelligence, but the products landing in the market this week were wrapped in budgets, permissions, logs, and partner-only access. (cnbc.com, siliconangle.com, oracle.com)