Anthropic and Accenture (NYSE: ACN) have announced a partnership to conduct independent safety evaluations of Anthropic's frontier artificial intelligence models. Both companies have committed to investing at least $1 billion each in the initiative over the next five years.

Partnership Structure and Scope

The collaboration will be led by Faculty, Accenture's AI business unit. The scope of work includes evaluating and red-teaming Anthropic's models, conducting alignment assessments, and testing model safeguards.

The arrangement is structured as an "embedded evaluation." This model allows evaluators to work inside the AI company with access comparable to that of an employee. Evaluators will observe models during training, monitor decisions governing model building and deployment, and interact directly with staff. Anthropic stated it will fund Accenture's work directly under this arrangement.

Accenture will leverage its decades of investment in responsible AI, deep technical and safety expertise through its acquisition of Faculty, one of the world's leading applied AI companies. Faculty is a recognized expert in testing and evaluating models for some of the world’s leading AI labs and has extensive experience building complex AI systems that are safe and ethical by design. Faculty's work spans government, defense, healthcare and infrastructure, including development of the UK National Health Service's Early Warning System during the COVID-19 pandemic.

Leadership Commentary

Julie Sweet, chair and CEO of Accenture, stated that Accenture is bringing together a dedicated team with deep AI, security and industry expertise to work alongside Anthropic. She noted that safety requires both deep technical expertise and a clear understanding of how AI is used in the real world. Sweet described embedded evaluation as an emerging area and expressed anticipation for partnering with Anthropic to help accelerate the development of embedded evaluators, which she sees as an important part of the safety landscape going forward.

Dr. Marc Warner, chief technology officer of Accenture and CEO of Faculty, added that Faculty was founded on the belief that AI should be safe by design, not safe by accident. He stated that joining forces with Anthropic as embedded evaluators is exactly the kind of work Faculty was built to do and that as part of Accenture, they have the platform to do it at a scale that can genuinely move the industry.

Industry Context and Future Plans

Anthropic described embedded evaluation as distinct from existing external evaluation practices. The company noted there are currently no established standards for what information embedded evaluators should access or how findings should be reported. Anthropic said it views long-term funding for such evaluations as ideally coming from pooled or government sources.

The partnership is non-exclusive. Anthropic plans to announce additional evaluators in the coming weeks, and Accenture may work with other AI developers in similar capacities. Anthropic is also in discussions with METR and other nonprofit evaluators to pilot elements of embedded evaluation using separate funding.

Anthropic stated that the use of independent evaluators does not shift responsibility for model safety away from the company. The announcement was framed as an early step toward a broader commitment to embed evaluators within Anthropic, as outlined previously by the company's chief executive.

Anthropic reported that its Claude model led 26% of the company’s measured AI research and development work in August, a sharp increase from just 1% in March.

The disclosure highlights the growing integration of AI agents into Anthropic’s core development processes, alongside details on safety research allocation and internal agent oversight.

Claude’s Role In AI Research Surges

According to a blog post by the AI startup, Claude was responsible for leading more than one-quarter of its measured R&D work in August. This figure is based on a measurement framework developed by independent nonprofit Epoch AI.

Anthropic clarified that Claude did not work fully autonomously in any covered area. Instead, the metric reflects scenarios where AI played the leading role while collaborating with human researchers. The company noted that more than 90% of its AI R&D work in August was performed at the "AI collaborates" level or higher.

Internal Agent Activity

Anthropic stated that approximately 30,000 AI agents were conducting research and engineering work simultaneously on its most-used internal platform in August.

To ensure safety, the company screens every action taken by these agents before execution. Of the more than a billion decisions made during the month, roughly one in 47,000 was blocked.

Safety Compute Allocation

The company disclosed that about 6% of the computing power used for AI research during a sample week in July went toward safety research. For research conducted by AI itself, this share rose to approximately 12%.

Anthropic described these figures as conservative, noting that computing used for work advancing both AI capabilities and safety was classified as capability-related.

What the Numbers Show

The divergence between the 6% general safety compute share and the 12% share for AI-conducted research suggests that autonomous or semi-autonomous AI workflows may require proportionally higher safety validation resources compared to human-led research efforts.

Industry Context

Last week, Anthropic CEO Dario Amodei called on AI companies to slow the pace of development, citing growing warnings from researchers about potentially catastrophic risks. This follows the resignation of researcher Jacob Coxon, who cited concerns over the company’s approach.

In funding developments, Nvidia Corp. (NASDAQ: NVDA) is reportedly considering investing as much as $10 billion in Anthropic’s potential IPO. The Claude developer seeks to raise up to $100 billion at a valuation of roughly $2 trillion. Rival OpenAI is reportedly holding preliminary talks with major investors about a new funding round that could value the ChatGPT maker at around $1.2 trillion.