Anthropic Picks Accenture as Its First Embedded AI Safety Evaluator

What happened

On September 18, 2026, Anthropic announced it is partnering with Accenture to establish a team of embedded evaluators who will work inside Anthropic alongside its internal teams and safety partners. The partnership will be led by Faculty, Accenture's specialist AI business, which Accenture acquired and which has experience testing and evaluating models for leading AI labs.

The embedded evaluators will evaluate and red-team models, conduct alignment assessments, and test model safeguards. Unlike traditional external evaluators who work from outside the companies they assess, embedded evaluators will have access comparable to an Anthropic employee's—allowing them to watch models take shape during training, follow the decisions that govern how models are built and deployed, and speak directly to employees.

Anthropic and Accenture each expect to invest at least $1 billion in building capacity in this area over the next five years. Accenture CEO Julie Sweet said the firm is bringing together a dedicated team with deep AI, security, and industry expertise to work alongside Anthropic.

The move fulfills a commitment Anthropic CEO Dario Amodei made in his September 2026 essay "We Must Pace the Frontier," in which he pledged to embed independent evaluators within the company.

Why it matters

This is the first concrete implementation of the "embedded evaluator" concept—a model where independent assessors sit inside an AI lab rather than reviewing it from the outside. If it works, it could become a template for how frontier AI companies are held accountable as models grow more powerful.

The choice of Accenture is notable and, to some, surprising. Anthropic is a company built around AI safety research; Accenture is the world's largest consulting firm, better known for enterprise deployments than safety scholarship. Anthropic's rationale is that Accenture's deep experience helping businesses and governments deploy AI across industries gives it a practical understanding of how models are used in the real world—a perspective that informs safety evaluation.

Faculty, the unit leading the effort, brings more targeted credentials. It has built complex AI systems for government, defense, healthcare, and infrastructure clients, including the UK National Health Service's Early Warning System during the COVID-19 pandemic. It is described as a recognized expert in testing and evaluating models for some of the world's leading AI labs.

The scale of the investment—at least $2 billion combined over five years—signals that both companies see embedded evaluation as a long-term institutional capability, not a one-off audit.

What to watch

Anthropic itself acknowledges that embedded evaluation is new and that many operational details are still being worked out. Key open questions include: how will Accenture's evaluators maintain independence while holding employee-level access inside Anthropic? What will they be allowed to report publicly, and what will remain confidential? How will conflicts of interest be managed given Accenture's broad commercial AI work?

Watch for the first public outputs from the embedded team—incident reports, alignment assessments, or safety findings—as those will set the precedent for what transparency looks like under this model. Also watch whether other frontier labs (OpenAI, Google DeepMind, xAI) adopt similar embedded-evaluator arrangements or push back on the concept.