Anthropic and Accenture plan embedded AI safety team

Anthropic and Accenture announced an embedded evaluation partnership that would bring an outside team into Anthropic to test its frontier models. Accenture’s specialist AI business, Faculty, is set to lead the work, and each company says it expects to invest at least $1 billion in related capacity over five years.

Evaluators would receive access comparable to an employee’s, including visibility into model training and decisions about model development and deployment, plus direct contact with staff. The plan turns Anthropic’s earlier commitment to embedded external evaluators into a named operating partnership.

Anthropic said standards have not yet been established for what information embedded evaluators should receive or how they should report their findings. The announcements did not specify the team’s size, members, start date, test protocols or escalation thresholds.

Faculty’s team will evaluate and red-team models, conduct alignment assessments and test safeguards, the companies said. Red-teaming means deliberately probing a system for failures or harmful behavior.

Anthropic said it will directly fund Accenture’s work because no pooled or government funding mechanism currently exists for independent evaluation. That means the model developer would pay the outside evaluator whose work is intended to make the developer’s safety practices more verifiable. The companies did not disclose contractual independence protections, conflict-of-interest controls, audit rights or remedies if evaluators are denied access.

Accenture said the work will combine technical evaluation with knowledge of how AI is used. The investment figures are expectations, not completed spending. Neither company provided a breakdown of what the money will cover or whether Anthropic’s payments to Accenture count toward either commitment.

Anthropic said responsibility for model safety will remain with the lab and described the agreement as non-exclusive. It expects to work with other evaluators, and Accenture may take similar roles with other AI developers. Anthropic said it was in dialogue with METR and other nonprofit evaluators about separate pilots and would announce other evaluators in the coming weeks; neither announcement committed to publishing Faculty’s findings.