San Francisco, September 18 — Anthropic and Accenture will put independent evaluators inside one of the world’s leading artificial intelligence developers, under a non-exclusive partnership the two companies announced this week. The deal commits a combined $2 billion — $1 billion each — over the next five years to build capacity for independent AI safety testing.
The structure, not the sum, is the story: Accenture’s evaluators will work inside Anthropic with what the companies call “employee-like access.” That means monitoring model training in real time, inspecting internal development decisions, and communicating directly with Anthropic staff. Typical external audits happen after the fact, from the outside.
This one happens during.
Red-teaming from the inside
The evaluation team’s core focus will be red-teaming models — deliberately trying to trigger flaws or harmful outputs — along with alignment assessments and testing of safety safeguards. The initiative will be directed by Faculty, a specialized AI business Accenture acquired earlier in the year. Anthropic is a US company.
Accenture is based in Ireland. The pairing brings enterprise scale to a job that has mostly belonged to nonprofit monitoring groups.
On funding, the companies said no standardized, government-pooled fund for independent AI evaluation exists yet, so Anthropic is paying for Accenture’s work directly in the short term. Both companies say they want public funding of independent evaluation over the long term.
The Amodei pledge behind the deal
The partnership stems from a pledge by Anthropic CEO Dario Amodei, who recently called on the tech industry to slow down development and allow third-party observers to scrutinize frontier models. Both companies emphasized the agreement is non-exclusive. Anthropic plans to bring on additional independent evaluators, including pilots with nonprofits such as METR, and Accenture intends to provide similar model verification services to other AI developers.
The timing matters. Independent safety testing has become a central policy question, and this deal places an enterprise-scale evaluator inside a leading frontier developer at that exact moment.
Accenture, through Faculty, gains a formal role in AI model verification it can market elsewhere. Anthropic gains an outside check on its own training processes — funded, in the near term, by Anthropic itself. Watch whether other frontier labs follow with embedded evaluators of their own, and whether the public funding both companies advocate for ever materializes.
