Anthropic has taken its first concrete step toward the slower, more closely supervised frontier-AI development advocated by chief executive Dario Amodei: bringing Accenture inside the company to evaluate its most powerful models.

The partnership, led by Accenture’s specialist AI unit Faculty, will involve red-teaming models, assessing alignment and testing safeguards. Unlike conventional outside reviewers, embedded evaluators are to receive access comparable to that of an Anthropic employee—enough to observe training, trace decisions about development and deployment, and speak directly with staff. Anthropic argues that proximity is the point: it can expose blind spots and allow reviewers to report incidents from an informed vantage point. “Independent embedded evaluators do not reduce our accountability, but help to make it more verifiable,” the company said.¹

The announcement follows Amodei’s call for AI developers to submit their safety practices to third-party examination as models become more capable. Anthropic says it remains responsible for the safety of its systems and intends to work with multiple evaluators, including the nonprofit METR. Accenture’s role is explicitly non-exclusive, and the consulting firm may take on similar work for other AI developers.

But the experiment also lays bare the governance gap it is meant to address. Anthropic and Accenture each expect to invest at least $1 billion over five years to build evaluation capacity, yet Anthropic will directly fund Accenture’s work because no pooled or government-backed funding mechanism currently exists. The company says standards are likewise unsettled: there is no agreed rulebook for what access evaluators should receive or how they should disclose findings.

That leaves the pilot carrying two messages at once. It is a notable concession to deeper oversight at a frontier lab; it is also an audit model whose independence has not yet been structurally secured. As one report put it, Anthropic selected Accenture to “internally monitor the safety of its AI tools” after its CEO’s oversight pledge.²