Anthropic Is Paying Accenture $1 Billion to Watch Over Its AI From the Inside

Key points

• Anthropic and Accenture each plan to invest at least $1 billion over five years in a programme that places independent safety reviewers inside Anthropic itself.
• The work will be led by Faculty, Accenture's specialist AI unit, and will cover red-teaming, alignment assessment, and safeguard testing.
• METR, one of the best-known nonprofit evaluators, employs fewer than 50 people and raised roughly $71 million in the six months to August 2026, up from $13.6 million in all of 2024.
• OpenAI fired three employees last week amid disputes over staff communication with outside evaluators.

For most of the past decade, the people paid to poke holes in AI systems operated quietly and cheaply on the fringes of the industry. That is over.

Anthropic, the San Francisco AI company behind the Claude family of models, announced a formal partnership with consulting giant Accenture committing both companies to at least $1 billion each over five years. The goal is something the industry hasn't tried at this scale: placing independent safety reviewers inside an AI lab, with access close to what a full employee would have.

What does "embedded evaluation" actually mean?

Unlike a traditional audit, where an outside firm examines a finished product, an embedded evaluator watches the product being made. Accenture's Faculty team will observe Anthropic's models during training, follow the decisions that shape how those models are built and deployed, and speak directly to staff. That's a level of access current external evaluators simply don't have.

It matters because AI safety problems often develop during training, long before a model reaches the public. Catching a problem then is far cheaper than recalling a deployed system.

Anthropic is funding Accenture's side of the work directly, at least for now. The company acknowledged in its announcement that there are "no standards" yet for what embedded evaluators should see or how they should report findings, and said it believes long-term funding should come from pooled or government sources. We first covered Accenture's move into this kind of embedded engineering role on 8 September, when the firm paired with Google to place AI engineers inside large enterprises.

Why does independent oversight need a billion dollars?

Because the current ecosystem is tiny relative to the labs it's supposed to check. METR, probably the best-known nonprofit evaluator, employs fewer than 50 people. It raised about $71 million over six months in 2026, which sounds substantial until you set it against the tens of billions the leading labs have raised in the same period.

For-profit evaluators are growing fast too. Vals AI, which builds benchmarks to measure how models perform on industry-specific tasks, grew from 8 to roughly 30 employees this year and closed a $40 million funding round in August 2026.

The funding gap creates a conflict researchers have been raising for months. If a small evaluator depends on an AI lab for revenue, it has a financial reason to write favourable reports. Suresh Venkatasubramanian, a computer science professor at Brown University, told CNBC: "If I am an auditor and I put out a report that looks unfavorable to this company, is my business going to dry up?"

The Anthropic-Accenture deal sidesteps that problem by having Anthropic pay upfront, but it creates a different question: can an evaluator funded by the company it watches ever be truly independent?

What does this mean for ordinary people?

AI models now help write medical letters and screen job applications for millions of people daily. If those models behave badly and there's no credible watchdog to catch it, most people wouldn't know until the harm had already happened. Our earlier story on Anthropic's models being exploited outside companies four times this year shows the threat isn't hypothetical.

Embedded evaluation is an attempt to build that watchdog function before it's urgently needed. Whether it works depends on details Anthropic itself admits are still being worked out.

Meanwhile, OpenAI fired three employees last week for what the company called policy violations around accessing sensitive information. Two of them, Mikita Balesni and Tomek Korbak, say the dismissals were a warning to staff not to communicate freely with third parties. OpenAI denied that reading and said it's finalising contracts with safety assessors.

The speed of all this is striking. Twelve months ago, AI evaluation was a niche academic exercise. Today it's attracting billion-dollar commitments and firing controversies in the same week. Watch whether those "no standards" Anthropic admits to become a loophole that lets embedded evaluators be managed rather than independent.