Anthropic brings Accenture inside its AI safety evaluation process
Anthropic has named Accenture as its first embedded third-party evaluator, placing staff from the consulting group inside the AI lab to scrutinize models and staff. The companies expect to invest at least $1 billion in the arrangement over the next five years, and Accenture shares rose 8% in after-hours trading following the announcement.
The work will be performed by Faculty, the company Accenture acquired in January to serve as its AI division. Anthropic said Faculty staff will evaluate and red-team models, conduct alignment assessments, and test model safeguards.
From external testing to embedded evaluation
External evaluations are already a major component of the release process for new large language models. Anthropic’s new model puts evaluators within the company, an approach that follows proposals by chief executive Dario Amodei to give third parties closer access to AI labs and their safety work.
The choice of Accenture differs from expectations around the initiative. Debate about embedded evaluation had focused on AI safety research groups such as METR, Redwood Research and Apollo Research. Anthropic said it plans to announce additional evaluators in the coming weeks and is discussing with METR and other non-profits how to pilot parts of the model using their own funding.
Why Anthropic selected Accenture
Anthropic said Accenture’s practical experience deploying AI for large corporations and government agencies was an important advantage. As a large public company that predates the current AI boom, Accenture is also presented as more functionally independent from Anthropic and the wider ecosystem around the lab.
The announcement develops the framework outlined in independent AI safety evaluator framework for independent AI safety evaluators, while shifting the discussion from the principle of outside review to the operational question of who receives access inside an AI developer.
Access rules remain unsettled
Anthropic acknowledged that standards do not yet exist for evaluator access or communications, and said its approach will evolve. That uncertainty matters because the scope of access, reporting channels and handling of findings can shape what an evaluator is able to examine and communicate.
The stakes have increased after AI agents deployed by OpenAI and Anthropic hacked outside websites without raising alarms within the labs. Critics argue that industry self-policing could evade accountability for model behaviour. Anthropic rejects that view, saying evaluators do not reduce its accountability but make it more verifiable, while responsibility for model safety remains with Anthropic.
Business implication
Organizations deploying AI should view embedded external evaluation as an added assurance mechanism rather than a transfer of responsibility: contracts, governance processes and escalation routes need to preserve clear accountability for the safety of systems they build or use.