Accenture and Anthropic Expand AI Safety Evaluation
Anthropic and Accenture have announced a partnership to establish a team of embedded evaluators inside Anthropic. The initiative will independently assess frontier AI models and examine how Anthropic develops and deploys them.
Accenture and Anthropic Build Embedded AI Evaluation Team
Anthropic and Accenture are partnering to establish a team of embedded evaluators within Anthropic. The team will work alongside Anthropic’s internal teams and safety partners.
The evaluators will assess frontier AI models through red-teaming, alignment assessments, and safeguard testing. They will also examine how Anthropic develops and deploys its models.
Both companies expect to invest at least $1 billion each over the next five years. Together, their commitment could reach at least $2 billion for AI safety and evaluation capacity.
Embedded Evaluators Will Work Inside Anthropic
The partnership introduces what Anthropic calls “embedded evaluation”. Unlike conventional external assessments, evaluators will work inside an AI company with access comparable to employees.
As a result, evaluators can observe models during training and follow decisions affecting their development and deployment. They can also speak directly with employees involved in those processes.
This access also lets evaluators examine how Anthropic applies its safety commitments in practice. They can identify potential blind spots, report incidents, and provide additional information about model risks and benefits.
Anthropic said the approach remains in an early stage. In particular, no established standards currently govern evaluator access, reporting procedures, or long-term funding.
Therefore, Anthropic expects the model to evolve as more evaluators participate. The company also plans to work with several evaluation organisations rather than relying on one external partner.
Accenture’s Faculty business will lead the partnership. Faculty joined Accenture through its acquisition of the applied AI company in 2025 and will provide technical and AI safety expertise.
Faculty to Lead AI Safety Assessments
Faculty will conduct model evaluations and red-team exercises as part of the programme. It will also carry out alignment assessments and test safeguards built into Anthropic’s systems.
Meanwhile, Accenture brings experience working with organisations across sectors such as government, healthcare, defence, and infrastructure. Consequently, the partnership will apply that broader experience to assessing frontier AI systems.
Accenture said its dedicated team will combine AI, security, and industry expertise. The company also views embedded evaluation as an emerging part of the wider AI safety landscape.
Moreover, Anthropic said the partnership is non-exclusive. Accenture can therefore work with other AI developers, while Anthropic can bring additional independent evaluators into its own programme.
Anthropic is already discussing pilot projects with METR and other nonprofit evaluation groups. These efforts could test parts of the embedded evaluation model under different funding arrangements.
The company also expects additional evaluator partnerships to emerge in the coming weeks. Thus, the initiative aims to build an ecosystem rather than a single evaluation arrangement.
AI Evaluation Moves Closer to Model Development
The partnership comes as AI developers face increasing pressure to demonstrate that advanced models remain reliable and controllable. Recent incidents involving AI agents have also increased attention on model safeguards and unexpected behaviour.
For Anthropic, embedded evaluators provide a way to examine safety practices from within the development process. However, the company has stressed that the evaluators will not transfer responsibility for model safety away from Anthropic.
Instead, Anthropic says independent evaluators should make its safety commitments more verifiable. The company also wants multiple organisations to conduct assessments under shared standards as the field develops.
Meanwhile, Anthropic has argued that long-term evaluation should eventually receive funding from pooled or government sources. For now, however, Anthropic will directly fund Accenture’s work while discussions continue with nonprofit evaluators.
The companies’ five-year investment commitment therefore combines evaluation capacity with a broader effort to develop new safety practices. At the same time, the programme could provide evaluators with closer visibility into how frontier AI systems are trained, tested, and deployed.
As AI development continues, the partnership brings independent assessment closer to the development process. Consequently, Anthropic’s approach could provide a working model for how AI companies involve external evaluators while retaining responsibility for their systems.
The key partnership details come primarily from Anthropic’s announcement and Accenture’s newsroom release, with Reuters providing independent reporting on the investment commitment and broader context.
