Anthropic Hired Its Own Sales Partner to Grade Its AI Safety

Anthropic just named its first "embedded evaluator" for AI safety. It's not METR. It's not a safety nonprofit. It's Accenture — the same firm Anthropic struck an undisclosed-value partnership with nine months ago to train 30,000 of its consultants on Claude.

What actually got announced

On September 18, Anthropic and Accenture said each side expects to invest at least $1 billion over five years so Accenture's AI unit, Faculty, can operate as an "embedded evaluator" inside Anthropic [1][2][5]. Embedded evaluators get access close to an employee's — reviewing training decisions, sitting in on deployment calls, red-teaming models, running alignment assessments, testing safeguards [2]. The idea traces back to CEO Dario Amodei's essay "We Must Pace the Frontier," where he committed to giving outside evaluators office-badge-level access and the right to publish findings without Anthropic's editorial veto — redactions limited to legal, security, commercially sensitive, and third-party-confidential material [3].

That's a genuinely good idea. Labs grading their own homework has been the industry's standing problem since the first safety card. Bringing in outsiders with real access and a public paper trail is the right shape of fix.

Accenture, though, is a strange first pick to prove it.

The part every outlet buried in paragraph six

Anthropic's own announcement sells Accenture on experience, not distance: "Accenture helps businesses and governments deploy AI across many industries. Their understanding of how enterprises use AI in practice informs their safety approach" [2]. What it doesn't mention: in December 2025, Anthropic and Accenture launched the Accenture Anthropic Business Group — a multi-year partnership training roughly 30,000 Accenture staff on Claude, with dedicated engineers embedding Claude inside client environments, plus a premier deal making Accenture developers heavy users of Claude Code [4]. Announcing Accenture as Anthropic's flagship coding partner is the same company now grading the safety of the models it sells.

Anthropic's own blog post concedes the funding problem directly: no settled model exists yet for who should pay evaluators, and long-term it wants pooled or government funding instead [2]. Until that exists, it's paying for the evaluation itself — from the same relationship that's already worth nine figures a year in enterprise deployment.

TechCrunch's reporting on the deal notes that discussions about embedded evaluators had centered on AI-safety research groups — METR, Redwood Research, Apollo Research — nonprofits with no deployment business riding on the verdict [1]. Anthropic says it's "in dialogue" with those groups too, and that this arrangement is non-exclusive [2]. Accenture just got there first, with the biggest check.

Anthropic's safety evaluator and Anthropic's sales partner are the same company, on the same five-year contract.

Who else was in the room

The nonprofits in that table have a structural advantage Accenture doesn't: nothing to lose if the verdict is bad. Accenture has a Claude reseller business to protect.

The loop, drawn out

The relationship isn't one contract — it's two, running in the same direction, with the same two names on both ends.

flowchart TD
  A[Anthropic] -->|"Dec 2025: enterprise partnership,<br/>30,000 staff trained on Claude"| B[Accenture / Faculty]
  A -->|"Sept 2026: $1B over 5 years<br/>to fund embedded evaluation"| B
  B -->|"Red-teaming, alignment tests,<br/>safeguard audits"| M[Anthropic's frontier models]
  M -->|"Deployed to enterprise clients,<br/>many via Accenture's own delivery teams"| C[Enterprise customers]
  B -->|"Findings published;<br/>Anthropic can redact only<br/>legal/security/commercial/<br/>third-party material"| P[Public disclosure]

Every arrow in that diagram is real. None of them individually looks like a conflict. Stacked together, they describe a company paying its own reseller to tell the public how safe its product is.

What this means for your own vendor checklist

Nobody reading this is signing a nine-figure evaluator contract. But the pattern underneath it shows up at SMB scale constantly: a vendor's "independent audit," "third-party certification," or "security-reviewed by [name]" badge is worth exactly as much as the relationship behind it — and that relationship is rarely disclosed on the badge itself.

Before you take a vendor's safety, security, or compliance claim at face value, ask one question: who's paying the auditor, and do they have any other deal running with the company they're auditing? If the answer is "yes, a bigger one," the audit is marketing with extra steps. This is the same discipline behind how I structure agentic automation engagements — every claim a vendor makes about what its AI can or can't touch gets independently checked before it goes into a client's stack, the same way every AI-touched record in Palmer's $1.25M migration got logged and validated instead of taken on faith. I wrote about a version of this exact failure mode a few weeks back — a vendor's own claims about its tool's behavior turning out to need outside verification, not trust.

Source: Dario Amodei — We Must Pace the Frontier

Sources

[1] TechCrunch — Anthropic's first embedded evaluator is … Accenture? — techcrunch.com

[2] Anthropic — Partnering with Accenture on embedded evaluation — anthropic.com

[3] Dario Amodei — We Must Pace the Frontier — darioamodei.com

[4] Accenture Newsroom — Accenture and Anthropic Launch Multi-Year Partnership to Drive Enterprise AI Innovation and Value Across Industries — newsroom.accenture.com

[5] Accenture Newsroom — Accenture and Anthropic Partner to Build Team of Embedded Evaluators at Anthropic — newsroom.accenture.com

The short version

• Anthropic named Accenture (via its Faculty unit) as its first "embedded evaluator" for AI safety — $1B+ from each side over five years, employee-level access, red-teaming and alignment testing included.
• Accenture isn't a safety nonprofit. Since December 2025 it's also Anthropic's flagship enterprise partner: 30,000 staff trained on Claude, dedicated deployment engineers, premier Claude Code access.
• METR, Redwood Research, and Apollo Research — the groups usually named for this role — have no commercial stake in the verdict. Anthropic says talks with them are ongoing but unresolved.
• Anthropic's own announcement admits there's no settled funding model for evaluators yet, and that it wants pooled or government funding eventually. Today, it's paying the auditor directly.
• The operator lesson travels past AI labs: before trusting a vendor's "independent" audit or certification badge, check who pays for it and what other contract runs between the same two names.