Anthropic announced on September 18 that it picked Accenture as its first "embedded evaluator," and every outlet filed it under AI safety. But read the access terms and it's a staffing document. Faculty, the independent UK AI firm Accenture agreed to acquire back in January, will place people inside Anthropic with access "comparable to an employee's": they can observe training runs, examine deployment decisions, and sit down with engineering staff to ask questions. Anthropic and Accenture are each committing at least $1 billion over five years. It's the first concrete move on CEO Dario Amodei's proposal to slow AI development down enough to check it's safe, and it's not exclusive. Anthropic says it's in talks with the nonprofit METR and others too. In the essay laying out the commitment, Amodei spelled out the kit: "Desks in our offices, access badges, and company laptops. Access to workspaces, tools, and permissions mostly comparable to what internal risk assessment teams have."
And it is a safety story. Dario Amodei has been public about wanting outside eyes on frontier model development before something goes wrong that a lab grading its own homework wouldn't catch. His case for it: "Embedded evaluators can check at the level of nuts and bolts whether an AI company is actually following the training, deployment, operational, and safeguards practices they claim to be following."
Fine.
But strip out the AI framing and look at what actually got created here: a new class of person, employed by a different company, sitting inside your buildings with the same access as your own engineers, empowered to walk up to your staff and ask about decisions they made. That's not a safety program. That's a hire — just one your HR department didn't run and your badge system didn't vet on its own terms.
I've spent years staffing roles where "who gets access to what" is the entire job description. Compliance seats, embedded QA, on-site vendor engineers who need production credentials to do their work. The pattern is always the same: the access decision gets made by whoever's excited about the capability, and the vetting process gets built after the first incident, not before. Anthropic and Accenture clearly built the governance layer for this deal. $2 billion and a multi-year commitment doesn't happen without contracts, scope limits, and probably a very long NDA. Accenture chair and CEO Julie Sweet pitched it as a capability her firm is built for: "Safety requires both deep technical expertise and a clear understanding of how AI is used in the real world." Most companies granting a vendor this kind of access have none of that. They have a Slack DM and a shared drive folder.
Worth noting that not everyone thinks the paid version solves it either. The same week, more than 100 AI researchers (Geoffrey Hinton among them) signed a public letter arguing evaluators shouldn't be commercially tied to the labs they evaluate. "We're just trying to really demonstrate a shared common ground on basic principles and ensure that independent oversight can be a meaningful tool for managing AI risk broadly," Conrad Stosz, chair of the AI Evaluator Forum consortium that organized the letter, told CNBC. The letter targets Anthropic and OpenAI both, and it isn't about Accenture specifically — but it lands on the same question I care about: who decides the terms of the access, and who they answer to.
If you're a CTO or VP Eng thinking "we don't have frontier models, this doesn't apply to us" — you're missing the transferable part. You almost certainly have vendors, contractors, or consultants with access to your codebase, customer data, or infrastructure broader than the job justifies, and no one owns that decision. AI agent access is the same failure mode I wrote about here in mid-September. 88 percent of companies had an agent security incident in the past year because nobody's job was to own the agents already running. Embedded evaluators are the human version of the same gap: capability arrives faster than the org chart entry that's supposed to govern it.
Here's the staffing question this should actually raise for you: when you bring in an embedded outside person (an auditor, a specialist consultant, a vendor engineer), who in your organization owns writing the access scope, and who owns re-checking it in six months? If the honest answer is "whoever asked for the help, and nobody after that," you have the exact structural gap Anthropic just spent $2 billion trying to close before it became a headline. You don't need $2 billion. You need one named owner and a review date on the calendar. That's a staffing decision, not a security tool purchase, and it's the kind of role I'd rather help you define now than help you backfill after an incident report.
The interesting hire coming out of this, if you're paying attention past the safety framing: "embedded evaluator" is about to become a real job title outside of AI labs too — someone whose entire function is sitting inside another company's systems with scoped, time-boxed, employee-level access and reporting out. If your compliance or vendor-management function doesn't own that lifecycle, that's worth fixing before you're the case study instead of the reader.