Anthropic taps Accenture for first embedded AI safety evaluator
Accenture’s Faculty unit will test Anthropic models under a nonexclusive partnership backed by planned $1 billion commitments from each company.
By Maya Okafor · Markets Writer
· 3 min read
Anthropic has chosen Accenture as its first embedded evaluator, placing a team from Accenture’s Faculty business alongside Anthropic’s internal teams and safety partners to examine AI models. For investors tracking the business of AI safety, the Anthropic Accenture embedded evaluator partnership is an early test of whether outside specialists can be built into a model developer’s safety process while the developer retains responsibility for its products.
Accenture and Anthropic announced the arrangement Sept. 18. The planned team will evaluate and red-team models, conduct alignment assessments, and test model safeguards, according to Accenture’s release.
The companies each expect to invest at least $1 billion over five years in AI safety. That is a forward-looking commitment, rather than a completed investment. Anthropic said it will fund Accenture’s work directly for now because pooled and government funding sources it would prefer do not yet exist, CNBC reported.
What will Accenture’s embedded evaluators do at Anthropic?
Anthropic said it will initially embed Faculty employees to test safeguards, red-team models and assess whether models behave in line with human values, according to CNBC. Faculty is Accenture’s applied-AI business, and Accenture said it will draw on that unit’s technical and safety expertise.
An embedded evaluator works within the company being assessed rather than reviewing it solely from outside. The announcement does not describe Accenture as a regulator or give it enforcement powers. It also does not specify the size of the team, when work will begin, the exact access it will receive, how incidents would be reported, or whether findings will be made public.
How does the deal fit Dario Amodei’s proposal?
CNBC described the partnership as Anthropic’s first concrete move under Chief Executive Dario Amodei’s three-step proposal to slow the development pace of the most advanced AI systems. Under the first step described by Amodei, third-party evaluators would receive employee-level access to verify safety practices and report incidents.
The Accenture arrangement does not put the full proposal into operation. TechCrunch reported that standards for evaluator access and communications have yet to be established, and Anthropic expects its approach to develop over time.
Anthropic has also said the partnership is nonexclusive. The company is discussing possible embedded-evaluation work with the research nonprofit METR and other third parties, CNBC reported. TechCrunch said some of those nonprofit discussions concerned pilots funded by the organizations themselves.
Who remains accountable for Anthropic’s models?
Anthropic said working with embedded evaluators does not reduce its responsibility for its models’ safety. That distinction is central to the structure: Accenture’s team is being brought in to perform specified assessments, but the announcement does not transfer operational or safety accountability away from Anthropic.
For Accenture, the agreement creates a sizable planned AI-safety commitment tied to its Faculty unit. For Anthropic, it provides a defined outside-testing partnership while key details about access, reporting and disclosure remain unsettled.
This story draws on original reporting from CNBC.
