Key points

  • Anthropic and Accenture each expect to invest at least $1 billion over five years in AI-safety evaluation capacity.
  • Accenture's Faculty unit will embed evaluators at Anthropic to test models, safeguards and alignment work.
  • The arrangement is non-exclusive and its access, reporting and long-term funding standards are still being developed.

Anthropic and Accenture have announced a five-year program to build independent evaluation capacity for advanced artificial-intelligence models, with each company expecting to invest at least $1 billion. The partnership, disclosed on September 18, will place a team of evaluators alongside Anthropic's internal groups and outside safety partners. Their work is set to include model evaluation and red-teaming, alignment assessments and tests of safeguards. Accenture's specialist AI business, Faculty, will lead the effort.

A different model for outside scrutiny

The companies describe the approach as embedded evaluation. Instead of reviewing a finished model only from outside the developer, evaluators would work inside an AI laboratory with access comparable to an employee's. Anthropic said that position could let evaluators observe models during training, examine decisions governing development and deployment, speak directly with staff and identify blind spots earlier. The company also said the evaluators could report incidents and provide a more informed public account of a model's benefits and risks.

Related reporting: AI slowdown warnings send Nasdaq futures and chip stocks lower

How the $2 billion commitment is structured

The headline figure combines the two companies' expected spending: at least $1 billion from Anthropic and at least $1 billion from Accenture over five years. Neither announcement provided an annual spending schedule, staffing target or breakdown between personnel, computing, research and other costs. Anthropic said it will directly fund Accenture's work because no pooled industry or government funding system yet exists for embedded evaluation. The broader commitments may also support capacity beyond the immediate team, but the companies did not publish detailed budgets.

Faculty takes the operating role

Faculty, which Accenture acquired as its specialist AI business, will supply technical and safety expertise to the program. Accenture said the team will combine AI, security and industry knowledge, drawing on Faculty's experience evaluating models and building systems for government, defense, healthcare and infrastructure. Reuters and the Financial Times independently confirmed the investment commitments and the plan for evaluators to work closely with Anthropic's teams. The partnership is non-exclusive: Anthropic can appoint other evaluators, and Accenture can perform similar work for other AI developers.

Independence remains a live question

The arrangement creates a potential tension because Anthropic will pay an organization intended to provide independent scrutiny. Anthropic acknowledged that standards have not been settled for evaluator access, public reporting or funding. It said embedded oversight does not transfer responsibility for model safety away from the developer. The company is also discussing pilot work with nonprofit evaluator METR and other groups using their own funding, an attempt to test different structures while the field develops common rules.

What the program must prove

The practical test will be whether embedded evaluators receive enough access to identify material problems while retaining the freedom to report uncomfortable findings. Important unresolved details include how confidential information will be handled, who decides what becomes public, how disagreements are escalated and whether evaluation results will be comparable across laboratories. The initiative could influence how enterprises, regulators and investors judge the credibility of safety claims made by frontier-model developers. Anthropic said it expects the approach to evolve and plans to share more as work begins. Until those operating standards are published, the $2 billion commitment signals scale and intent, but not yet a completed framework for independent AI oversight.

Sources

AI-generated editorial image; not a photograph of the reported event. Prepared with AI assistance and source verification.