Schellman flags critical gaps in White House AI oversight accord

The White House got the biggest names in artificial intelligence to agree on something: their safety systems should be checked by someone other than themselves.

The catch is that nobody wrote down who that someone is allowed to be. Schellman, which tracks AI governance readiness across organizations, is pointing at that blank space as the accord’s most serious weakness.

The agreement, called the White House Accord on Super Intelligence, was released on September 29, 2026. It is a voluntary framework for AI governance. Executives from Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed it alongside President Trump.

Four layers of oversight, zero definitions for the auditors

The accord sets up four layers of oversight for signatories.

First, companies need robust internal controls to monitor risks. Second, they need an internal team dedicated to monitoring. Third, they need independent external auditors or evaluators to test whether those controls actually work. Fourth, they need an independent board committee.

The problem sits in layer three. The accord asks for independent assessments without spelling out who qualifies to perform them.

It also skips over what independence actually requires. And it says nothing about how conflicts of interest should be managed.

Those calls are left to each signatory. That means two developers could technically satisfy the identical commitment in very different ways.

The confidence gap in Schellman’s data

Schellman’s own numbers suggest the definitional gap matters more than it might first appear.

The firm’s July 2026 State of AI Governance Report found that 74% of AI governance professionals believe their organizations could pass an AI compliance audit. Only 27% reported having fully mature governance programs.

The spending side looks healthy enough. According to the report, 90% of surveyed organizations have dedicated funding for AI governance.

The paperwork is thinner. Just 57% have formal governance policies in place, and only 44% have set up incident response procedures written specifically for AI.

Meanwhile, deployment is not waiting for governance to catch up. The survey found that 86% of companies have tested AI agents, and 46% have them running in production.

AI agents are systems that can take actions on their own, not just answer questions. Having nearly half of surveyed organizations already running them makes the readiness gap harder to shrug off.

A voluntary pact with no enforcement mechanism

The accord is voluntary, and that shapes everything about it.

There are no government oversight mechanisms written into the framework. There are no penalties for companies that fall short. And there is no requirement that audit results be disclosed publicly.

What this means for the AI industry and its backers

The most immediate effect is that “independent audit” is now a phrase every major AI developer will use, without a shared definition behind it. That puts the burden on investors, customers and regulators to ask follow-up questions.

Who performed the evaluation? Does that evaluator have commercial ties to the company? Were the findings published, summarized, or kept internal?

Schellman’s survey adds a warning for anyone relying on self-assessments. If 74% of organizations think they would pass an audit but only 27% have mature programs, then a company’s own statement of readiness is a weak signal on its own.