Anthropic's first embedded evaluator is Accenture, and Anthropic is paying for it

Anthropic said on Friday that Accenture will be the first outside firm to place evaluators inside the company, with access it describes as comparable to an employee's. Each company expects to invest at least $1 billion in building capacity for this work over the next five years. It is the first concrete implementation of a commitment Anthropic's CEO made six days earlier — and the choice of partner is already the most contested part of it.

The work will be led by Faculty, the applied-AI firm Accenture acquired, and covers evaluating and red-teaming models, running alignment assessments, and testing model safeguards. Anthropic says embedded evaluators will be able to watch models take shape during training, follow the decisions that govern how those models are built and deployed, and speak directly to employees — then report incidents and give the public an account of what they find. In the companies' joint press release, Marc Warner, CTO of Accenture and CEO of Faculty, framed the premise plainly.

AI should be safe by design, not safe by accident.

Where the commitment came from

On 12 September, Dario Amodei published We Must Pace the Frontier, an essay arguing that frontier labs should deliberately slow capability gains so that safety work, evaluation and coordination can catch up. It lays out three steps. The first — embedded evaluators — is the one Anthropic said it would adopt unilaterally, and Amodei called it "the key step for verifiability of any pacing commitments."

Two details of that essay matter now. The evaluator it named as an example was METR, a nonprofit. The precedent it cited was banking supervision, where regulators are sometimes embedded alongside a firm's own staff. Friday's announcement named neither a nonprofit nor a regulator. TechCrunch reported that the pick surprised observers who had expected METR, Redwood Research or Apollo Research, and that critics read it as the industry policing itself.

The independence question, which Anthropic raises itself

Anthropic is funding Accenture's work directly. The company is unusually explicit that this is not the arrangement it thinks should endure: it says long-term funding for independent evaluation should come from pooled or government sources, as it argued in its Advanced AI Framework in June, and that neither exists today. It also says there are as yet no standards for what information embedded evaluators should have access to, or how they should report what they find.

That candour is worth crediting, and it does not dissolve the problem. The banking analogy works precisely because supervisors are paid by the state rather than by the supervised bank. An evaluator paid by the lab it evaluates, operating under rules that have not been written, is a different instrument — closer to a commissioned audit than to supervision. Anthropic's answer is that embedded evaluators "do not reduce our accountability, but help to make it more verifiable," and that the safety of its models remains its own responsibility.

There is a second entanglement, and it comes from the companies' own announcement rather than from critics. Accenture is not a neutral party arriving fresh. The press release points to an existing multi-year commercial partnership between the two, announced roughly nine months ago, aimed at expanding enterprise use of Claude. The firm that will assess Anthropic's safety practices also sells Anthropic's models to its clients.

What would make this mean something

The announcement is a structure, not yet a result. A handful of things would show whether it is an independent check or a well-funded consulting engagement:

• Whether evaluators can publish findings Anthropic would rather they did not, and under what rules.
• Whether employee-level access survives disagreement — access granted voluntarily can be narrowed voluntarily.
• Whether the pilots with METR and other nonprofits, on those organisations' own funding, actually materialise.
• Whether other frontier labs follow, or Anthropic's unilateral step stays unilateral.
• Whether funding ever moves to the pooled or public sources Anthropic says it prefers.

Anthropic says the partnership is non-exclusive, that it is in dialogue with METR and other nonprofit evaluators about piloting elements of embedded evaluation on their own funding, and that further evaluators will be announced in the coming weeks. Accenture is expected to work with other AI developers in similar capacities. Those follow-on announcements will say more than this one did. Accenture's shares rose in Friday after-hours trading on the news, which is a small comment in itself on how the market read a safety-evaluation contract.

For people building on these models

Nothing changes this week. There is no API surface here, no model update, no pricing implication. The longer-horizon thing to watch is disclosure. If embedded evaluators end up assessing training pipelines and reporting incidents rather than only grading finished models, the practical payoff for API customers would be better information about why a model shipped when it did and what its safeguards were actually tested against. That would be genuinely useful. None of it is guaranteed by what was announced on Friday.