Agent Horizon

Anthropic has announced a partnership with Accenture in which Accenture’s specialist AI unit, Faculty, will place staff inside Anthropic to evaluate models, run red-teaming exercises, and assess whether systems behave as intended. Unlike typical external audits, these evaluators will get access comparable to employees, letting them observe training in progress and talk directly with Anthropic’s own teams. Both companies say they expect to invest at least $1 billion each over five years to build out this capability. The full announcement is available on Anthropic’s newsroom.

Why this matters

This is the first concrete action taken on the “embedded evaluator” concept that Anthropic’s CEO floated in a policy essay just days earlier, and it’s a genuinely unusual move: most AI safety oversight today comes from either internal teams or arm’s-length external audits, not evaluators with near-employee access sitting inside the building. Whether that access actually produces more honest, catching-things-early oversight, or just a closer relationship that softens scrutiny, is the real open question, and it’s one that outside observers will be watching closely.

It’s worth being clear-eyed about the limits here. Anthropic is directly funding Accenture’s work itself, which is a real conflict-of-interest risk that Anthropic openly acknowledges, saying the industry still lacks agreed standards for how such evaluations should be funded or disclosed. The partnership is also non-exclusive, and Anthropic says it’s separately talking to the nonprofit research group METR about piloting similar work with independent funding, which would be a stronger signal of genuine arm’s-length oversight.

Compared to competitors, this puts Anthropic further out in front on institutionalizing safety commitments into concrete organizational structure rather than just publishing more policy essays. Whether this becomes an industry norm, or Anthropic’s competitors respond with something comparable, will be a good marker of whether “embedded evaluation” is a genuine shift or a one-lab experiment.