Governance has become the sales wedge

The pitch to big companies has shifted in one very specific way: Anthropic spent April 9 talking about “organization-wide controls” for Claude Cowork, while OpenAI cut the price of its Pro plan for Codex to $100 a month from a higher premium tier. One company pushed controls, the other pushed cheaper access, and both moves were aimed at getting cautious enterprise buyers over the line. (siliconangle.com) That change tells you what buyers are worried about now. A large company is no longer asking only, “Is this model smart,” but also, “Who can use it, what can it touch, and what record do we keep if something goes wrong.” (siliconangle.com) Anthropic’s new controls were built around Claude Cowork, which it introduced in January 2026 as an autonomous agent that can handle multistep tasks on an employee’s computer. That is a much riskier product than a plain chatbot, because it can organize files, run browser tasks, and produce documents instead of just answering a prompt. (siliconangle.com) So Anthropic added role-based access controls for Enterprise customers, which means an administrator can decide which employee gets which capability. It also added group spend limits, so a legal team or marketing team can hit a budget ceiling before an experiment turns into a surprise invoice. (siliconangle.com) Anthropic also expanded OpenTelemetry support for Claude Cowork, which lets companies pipe events like tool calls and file modifications into their security systems. That is the software equivalent of putting security cameras in the hallway before you let a new contractor walk through the building. (siliconangle.com) The connector problem is part of the same story. Anthropic announced a new Zoom connector and more precise controls for other Model Context Protocol connectors, because once an assistant can reach meetings, documents, and internal tools, the question stops being raw intelligence and becomes controlled access to company memory. (siliconangle.com) OpenAI has been building the same kind of control layer into ChatGPT Enterprise and its application programming interface platform. Its help documentation says apps are disabled by default for ChatGPT Enterprise and Education workspaces, and administrators can assign app access through role-based access control instead of giving everyone every connector. (help.openai.com) OpenAI’s platform docs also spell out the mechanics: custom roles, group syncing through System for Cross-domain Identity Management, project-level permissions, internet protocol allowlists, and immutable audit logs for events like key creation, user changes, and login failures. Those are not flashy features, but they are exactly the features a security team asks for before approving a rollout. (developers.openai.com) (help.openai.com) The reason this sales message is landing is that many enterprise projects are not producing clean returns. Computerworld reported this week that a Gartner survey found only 28 percent of artificial intelligence use cases in infrastructure and operations met return-on-investment expectations, while 20 percent failed completely. (computerworld.com) Gartner’s finding was even more blunt than that: returns did not hinge on model sophistication as much as integration, governance, and fit with daily operations. If the software cannot be monitored, budgeted, and limited to the right people, a chief information officer has a harder time defending it to finance, security, and legal teams. (computerworld.com) That is why “governable” has become a better wedge than “smarter.” In 2026, the enterprise sale is increasingly won by the vendor that can show who approved the connector, which group had access, what the agent touched, and when a human had to step in. (siliconangle.com) (computerworld.com)