Accenture Joins Anthropic as First Embedded AI Safety Evaluator

Anthropic has named Accenture as its first embedded AI safety evaluator, placing staff from Accenture’s specialist AI unit, Faculty, directly inside the company to red team models, run alignment assessments, and test safeguards. Both companies say they’ll each invest at least a billion dollars over five years, a two billion dollar commitment total.

This is the first concrete move in a plan CEO Dario Amodei laid out on September 12 to slow the pace of frontier AI development. His three step proposal called for AI labs to give outside evaluators employee level access, the kind that lets someone actually watch a model get trained rather than just audit it after the fact.

And that distinction matters more than it sounds. Traditional audits happen after decisions are made. Embedded evaluators, by contrast, observe training in real time, follow deployment calls as they happen, and talk directly to the people building the thing. Anthropic says this setup helps “assess how a company operates, verify that it is keeping its safety commitments, and identify blind spots.”

Here’s the part that raised eyebrows. Accenture isn’t a frontier AI research shop. It’s a consulting giant best known for helping enterprises and governments deploy technology at scale. Anthropic’s own explanation leans into that gap rather than around it, arguing Accenture’s practical experience with how companies actually use AI in production gives it a perspective a pure research lab wouldn’t have.

Faculty, the unit doing the actual work, has experience evaluating models for other major AI labs already, so this isn’t Accenture’s first exposure to safety testing specifically.

Anthropic has been upfront that nobody’s figured out the rules yet. There’s no established standard for what evaluators should be allowed to access, how they report findings, or how independent funding for this kind of work should even be structured long term. The company said its approach will likely change as it goes.

One thing Anthropic keeps repeating: this doesn’t let it off the hook. “Embedded evaluators do not reduce our accountability, but help to make it more verifiable,” the company said, framing the whole arrangement as making its safety claims checkable rather than outsourcing responsibility for them.

Not everyone’s convinced that’s how it’ll actually play out. More than 100 researchers, including AI pioneer Geoffrey Hinton, signed an open letter in September arguing that access without genuine independence risks becoming a performance rather than real oversight. That criticism lands harder given Anthropic is reportedly targeting a two trillion dollar valuation around the same time it’s making these safety commitments public.

Anthropic says Accenture won’t be the only evaluator. The company is in active talks with METR, a nonprofit research group focused on AI evaluations, along with other third parties, about piloting similar embedded arrangements using separate funding sources.

For more on how Anthropic is handling security and safety concerns elsewhere, see our coverage of its insider risk investigator tool and the broader story on AI safety tests and security risk. If you want to see what Anthropic’s models actually do day to day, our Claude AI tool page covers the assistant this evaluation work is meant to keep in check.