Anthropic’s appointment of Accenture as its first embedded evaluator partner is best read as an effort to make external scrutiny part of how frontier AI models are developed, rather than a review conducted only after systems have been built. The companies said on September 18, 2026, that Accenture, working through its specialist AI business Faculty, will place staff inside Anthropic to evaluate and red-team its models, assess alignment and test safeguards.
Anthropic and Accenture each expect to invest at least $1 billion over the next five years to build capacity for the work. Faculty staff will operate within Anthropic with access comparable to that of an employee. That arrangement is intended to allow evaluators to observe models during training, track decisions on their construction and deployment, and speak directly with employees.
The partnership gives Anthropic access to Accenture’s scale while drawing on Faculty’s experience assessing complex AI systems. Accenture acquired Faculty in January. Faculty’s previous work has covered government, defense, healthcare and infrastructure, including the development of the UK National Health Service’s Early Warning System during the COVID-19 pandemic.
Accenture generated approximately $70 billion in FY25 revenue and has approximately 9,000 clients and 799,000 employees. Its shares shot up 8% after hours following the announcement.
Anthropic said the arrangement fulfils a commitment set out in its chief executive’s essay, ‘We Must Pace the Frontier’, to bring evaluators into the company. Dario Amodei had also proposed placing third-party safety evaluators within AI laboratories.
The model seeks to give evaluators a closer view of a company’s development process than an outside assessment might provide. Faculty will be able to examine models as they take shape, rather than relying solely on systems after their release or on material supplied for a specific review.
Anthropic said responsibility for the safety of its models will remain with Anthropic. The company will fund Accenture’s work directly, an arrangement that highlights a broader unresolved question over how independent evaluation should be financed.
There are currently no standards defining the information that embedded evaluators should receive or the way they should report their findings. There is also no settled funding model for independent evaluation. Anthropic called for pooled or government funding for embedded evaluation in its Advanced AI Framework, released in June.
The companies did not set out a timetable for Faculty staff to begin work at Anthropic, the size of the team, or the particular methods it will use for alignment assessments and red-teaming. The announcement also left open how findings could be shared publicly or with regulators, and how the arrangement will address any disagreement between Faculty’s recommendations and Anthropic’s business interests.
Anthropic said it is in dialogue with METR and other nonprofit evaluators about embedded evaluation, and plans to announce further evaluator partners in the weeks ahead. The Accenture arrangement is non-exclusive. Anthropic said it will work with other evaluators, while Accenture will be free to perform comparable work with other AI developers.
The partnership therefore establishes a first test of an embedded-evaluator approach at Anthropic, while leaving several practical questions unsettled. Its effectiveness will depend on the access evaluators receive, the independence they can exercise, and the arrangements that emerge for reporting and funding their work.
Comments (0)