Anthropic's Independent Evaluator Is Owned by Its Commercial Partner

On 18 September, Anthropic announced that Accenture will carry out independent evaluation of its frontier AI models. The work covers red-teaming, alignment assessments and testing of model safeguards. Both companies expect to invest at least a billion dollars each over five years. Anthropic is funding Accenture’s side of it directly.

The evaluators will sit inside Anthropic with access comparable to an employee’s. They will watch models take shape during training, follow the decisions that govern how those models are built and deployed, and speak to staff. The stated purpose is to verify that Anthropic keeps its safety commitments, identify blind spots, and give the public a more informed account of the risks.

Embedded evaluation is a serious idea and probably a necessary one. This particular structure would not survive contact with a pharmaceutical quality auditor. That matters more than it sounds, because the same alliance is selling AI into pharma.

The sequence

9 December 2025. Accenture and Anthropic announce the Accenture Anthropic Business Group. Approximately 30,000 Accenture professionals to be trained on Claude. Anthropic becomes one of Accenture’s select strategic partners. Accenture becomes a premier AI partner for Claude Code and deploys it to tens of thousands of its own developers. A joint offering for CIOs. Co-developed solutions for regulated industries, with life sciences named explicitly. Reported as a three-year deal.

6 January 2026. Accenture agrees to acquire Faculty, a UK applied-AI business whose service lines include AI safety.

16 March 2026. The acquisition completes. Faculty’s co-founder and chief executive, Dr Marc Warner, additionally becomes chief technology officer of Accenture and joins the company’s Global Management Committee.

18 September 2026. Anthropic announces that its independent embedded evaluation partnership will be led by Faculty.

Read in order, the independent evaluator of Anthropic’s models is a business unit of the firm Anthropic says is building one of the largest ecosystems of Claude practitioners in the world, acquired nine months into a three-year commercial alliance, and paid by Anthropic to do the evaluating.

Why the immateriality defence fails

The obvious reply is that a billion dollars over five years is a rounding error against Accenture’s revenue base, so fee dependence is negligible.

That is true at group level and beside the point at the level where the decisions actually get made.

I spent a long stretch of my career running a large P&L inside a global professional services firm. These organisations are federations. Each practice carries its own number. People are measured and paid against the number in front of them, not the consolidated result. An engagement that disappears into the group accounts can be the difference between a good year and a bad one for the team delivering it.

The sharper problem is not the evaluation team’s bonus. It is the account. Someone owns the Anthropic relationship globally, and that relationship is worth a great deal more than the evaluation fee. The evaluation lead does not control the account. Pressure inside a structure like that does not arrive as an instruction. It arrives as which findings get escalated, how they get worded, and who gets staffed next quarter.

There is a third conflict, and it is the one an auditor would flag first. Accenture’s deployment business needs these models to be certifiably safe, particularly for the regulated-industry solutions it has committed to co-develop. An adverse safety finding does not merely risk the evaluation contract. It damages the revenue line the firm has built around the product. The evaluator is long the asset it is assessing.

I should mark the line between fact and inference here. I have no evidence that Marc Warner has any role in the Anthropic alliance or in this engagement, and none is claimed anywhere. The point is structural, not personal. Personal integrity is not the variable in governance design. Reporting lines are.

What Anthropic says, and what it does not

Anthropic deserves credit for how much of this it states itself. Its announcement says plainly that no standards yet exist for what access embedded evaluators should have, or for how they should report what they find, and that there is no settled system for funding independent evaluation at all. It argues that funding should eventually come from pooled or government sources. It is in dialogue with METR and other nonprofit evaluators. The arrangement is non-exclusive on both sides.

What is not disclosed is the part that determines whether any of it works.

1. Can Accenture publish an adverse finding without Anthropic’s consent?
2. Who controls termination and renewal of the engagement?
3. Does any third party see findings that are not published?
4. What separates the evaluation team from the commercial alliance, and who enforces it?

Four questions, none of them answered. Until they are, independence is a description of intent rather than of structure.

Pharma settled this a long time ago

Anyone in life sciences reading the announcement should recognise the problem immediately, because the industry answered this exact question decades ago and then wrote the answer into law.

Quality Assurance does not report to Manufacturing. The Qualified Person who releases a batch carries personal legal liability and cannot be overruled by the plant manager. A supplier audit is not conducted by the supplier’s commercial partner.

None of that exists because the people involved were untrustworthy. It exists because the industry accepted that structure beats intention, and that a control which depends on good character is not a control.

Now hold the AI assurance you are being offered against that standard. In a growing number of cases the firm helping you deploy the model, the firm attesting that the model is safe, and the firm whose growth depends on you deploying more of it are the same firm. No segregation of duties. No independent release. No named individual carrying liability for the decision.

This is what I call the accountability gap, the failure mode sitting directly above the process layer, where governance is asserted rather than evidenced. It is the clearest public example of it I have seen.

What to do with this

Put those four questions, in writing, to whoever is selling you AI assurance. Model evaluations, safety attestations, responsible-AI certifications, validation packs, all of it.

If the assurance is produced by an organisation with a commercial interest in the deployment, you do not have assurance. You have marketing with a methodology section.

Juvenal’s line is usually quoted as a question about the state. It was not. He was asking who guards a household from the guards hired to watch it, and the joke was that the answer is nobody, because the guards answer to the man who needs watching. Pharma reached its own answer rather more expensively, and it has held for fifty years: the guard is never paid by the party being guarded.