Anthropic has taken its proposed model of frontier-AI oversight from theory toward a live experiment, naming Accenture as its first embedded evaluator. The company and the consultancy each expect to invest at least $1 billion over five years to build evaluation capacity, with Accenture’s specialist AI business, Faculty, leading the work.¹

The arrangement is designed to go further than conventional outside audits. Anthropic says embedded evaluators will work inside AI companies with access comparable to that of employees: watching models develop during training, tracking decisions over deployment, speaking directly with staff and assessing whether safety pledges are being met. Accenture is expected to red-team models, conduct alignment assessments and test safeguards, while bringing its experience of how businesses and governments use AI.¹

Anthropic’s argument is that proximity can strengthen scrutiny rather than dilute it. “Independent embedded evaluators do not reduce our accountability, but help to make it more verifiable,” it said, while stressing that ultimate responsibility for model safety remains with Anthropic.¹

But the announcement also exposes the weak points in this emerging oversight model. There are no settled standards for an evaluator’s access, reporting or funding, and Anthropic will directly fund Accenture’s work because pooled or government-backed financing does not yet exist. The company says it is talking with METR and other nonprofit evaluators, hopes eventually to work with several groups, and calls the Accenture partnership non-exclusive.¹

The skeptical response lands in a single pointed question: “Anthropic’s first embedded evaluator is ... Accenture?”² That framing does not dispute the need for rigorous testing. It challenges whether a consultancy paid by the lab it is examining can supply the independence the experiment promises. Anthropic’s next evaluator choices—and the rules governing what Accenture can disclose—will determine whether this becomes a credible safety template or a sophisticated form of self-supervision.