Anthropic is embedding Accenture evaluators inside its own company, and paying for them
Anthropic said on September 18 that it is partnering with Accenture on independent evaluation of frontier AI, with evaluators working inside the company rather than testing finished models from outside. The two say they each expect to invest at least $1 billion in building capacity for this work over the next five years.

Anthropic's own diagram of the safeguards it puts around agents, published on August 31 with the changes it made after Claude models reached real computer systems. Image: Anthropic
The partnership is led by Faculty, Accenture's specialist AI business, and covers evaluating and red-teaming models, alignment assessments and testing model safeguards. Anthropic frames it as a step towards a commitment in its CEO's essay "We Must Pace the Frontier" to embed evaluators within Anthropic.
Embedded evaluators get an employee's level of access
This is the detail that makes the deal different from the external audits the industry has run so far. Anthropic says embedded evaluators will work inside AI companies "with access comparable to an employee's", which lets them "watch models take shape in training, follow the decisions that govern how those models are built and deployed, and speak directly to employees".
Note where they sit. Accenture is a consultancy that deploys other companies' AI for its clients, but this is not an evaluation of customer deployments: the evaluators are placed inside Anthropic, looking at Claude while it is still being trained and at how the company decides what to ship.
From there, Anthropic says, they can assess how the company operates, check that it is keeping its safety commitments, find blind spots, report incidents and "give the public a more informed account of benefits and risks". The company adds a line that reads as a pre-emptive answer to the obvious objection: "independent embedded evaluators do not reduce our accountability, but help to make it more verifiable".
Anthropic is funding the people who check its work
The money moves in one direction. "Given the importance and urgency of this work, Anthropic will fund Accenture's work directly," the post says. The $1 billion figures are what each side expects to spend on building capacity in this area, not a contract signed between them, and Anthropic has published no fee, no headcount and no schedule.
Anthropic does not dodge the tension. It says long-term funding "should come from pooled or government sources", as it argued in its Advanced AI Framework in June, and that since no such system exists it will work with different evaluators under different funding arrangements. It also says it is in dialogue with METR and other nonprofit evaluators about piloting elements of embedded evaluation on those evaluators' own funding.
The other admission is blunter: "There are, as yet, no standards for what information embedded evaluators should have access to, or how they should report what they find." A company is paying an outside firm to inspect it, under rules neither of them has written yet.
What has not been set
There is no start date, no named evaluation staff and no published schedule for what gets reported or to whom. Anthropic says the partnership is non-exclusive, that it will name other evaluators "in the coming weeks", and that Accenture will work with other AI developers in similar capacities.
The context for all this is on Anthropic's own site. On August 31 it set out what it had changed after three incidents in which Claude models gained unauthorised access to real computer systems, which it first reported on July 30, and said it was planning to work with METR for an independent review. Outside eyes inside the building are a response to that kind of month.
For readers watching what Claude is allowed to do rather than who watches Claude, the same fortnight brought looser biology safeguards for vetted research teams and Anthropic's claim that Claude led 26% of its own AI research work in August. Both are decisions an embedded evaluator would, in principle, now be in the room for.
None of that starts until the first evaluator turns up. Anthropic says it will share more as the work begins.