Confirmed in a primary source 18 Sept 2026 Safety

Anthropic named the outside inspectors it promised, and picked a consulting firm rather than a safety institute.

On September 18 Anthropic said Accenture will be its first embedded evaluator, putting Accenture staff inside Anthropic to test and red-team its models, run alignment assessments, watch model development during training, observe deployment decisions and assess whether the company keeps its own safety commitments. The work runs through Faculty, the British AI firm Accenture bought in January, and the two say they expect to put at least one billion dollars into building evaluation capacity over five years, with Anthropic paying for the work at the start.

The caveat that matters is who chose whom. Anthropic selects its evaluator, funds it, and sets the access; the arrangement is explicitly non-exclusive on both sides, and no standard yet exists for what an embedded evaluator must be allowed to see or say.

Until now

Embedded evaluation was a principle two labs had announced, with no named party, no money and no sense of what kind of organisation would do it.

Now

Read the first embedded evaluator as a procurement decision rather than an appointment, and watch whether anyone writes down what an evaluator must be allowed to see and to say.

The surprise here is the name, and the surprise is doing real work. When Anthropic said on September 12 that it would give outside inspectors desks, badges and company laptops, the field it was drawing from looked small and specific: METR, Apollo Research, Redwood Research, the organisations built for this and staffed by people who do nothing else. What arrived on September 18 was Accenture, a listed consultancy with several hundred thousand employees, working through Faculty, the British AI firm it bought in January. Anthropic's stated reasoning is worth taking at face value: Accenture has deployed these systems inside large companies and governments, which is where the failures land, and it is a large public company that existed long before this industry, which makes it harder to capture than an organisation whose entire funding base is four laboratories. The counter-reading is equally available. A consultancy's business is its client relationships, and the client here is also the payer.

The billion-dollar figure over five years is the part that separates this from a memorandum. It is a number with a denominator, big enough to build a standing capability rather than to run a project, and it is the first price anybody has attached to independent oversight of a frontier lab. What it does not do is answer the question the arrangement raises. There is still no written standard for what an embedded evaluator must be allowed to see, whether it may publish what it finds, what happens if it disagrees with a deployment decision, or who it tells if the lab does not listen. Anthropic chose its inspector, pays its inspector, and defines its inspector's access, and says the approach will evolve. Accenture's shares rose about 8 percent after hours, which tells you the market read this as a revenue line. Read the first embedded evaluator as a procurement decision rather than an appointment, and watch whether anyone writes down what an evaluator must be allowed to see and to say.

How it got here

Jan 2026
Accenture acquires Faculty, the British AI firm that will now supply the evaluation staff.
Sep 09
Anthropic discloses a fourth break-in by its own models and hires METR to investigate it. Carded in Edition 017.
Sep 12
Anthropic says it will give outside inspectors desks, badges and company laptops; OpenAI says the same hours later. Carded in Edition 020.
Sep 18
Anthropic names Accenture as its first embedded evaluator, non-exclusive, with at least one billion dollars expected over five years.
Who has to react
METR, Apollo Research and Redwood Research, which were the expected choice and were not chosen, and whose funding case just changed shape; OpenAI, which made the same promise on September 12 and has named nobody; Google DeepMind and Meta, which have made no such promise at all; the European Commission's AI Office, which needs third-party assessment for systemic-risk models under the AI Act and has just been handed a template it did not write; the UK AI Security Institute, which has evaluated frontier models as a public body and now has a commercial competitor; Accenture's rival consultancies, which can read an 8 percent move as well as anyone; and enterprise buyers who will be asked to treat an evaluator's presence as assurance.

The evidence

ObservedFriday September 18, 2026. Anthropic published the partnership on anthropic.com under the title Partnering with Accenture on embedded evaluation.

The whole observation, as recorded

Scope as published: evaluating and red-teaming models, conducting alignment assessments, testing model safeguards, observing model development during training, observing deployment decisions, and assessing company operations and safety commitments. The named delivery vehicle is Faculty, described as Accenture's specialist AI business; Accenture acquired Faculty in January 2026. Money: the two organisations expect to invest at least one billion dollars in building evaluation capacity over the next five years, with Anthropic directly funding Accenture's work initially. Terms: explicitly non-exclusive. Anthropic says it plans to work with additional evaluators in the coming weeks, and Accenture is free to work with other AI developers. Anthropic's published page carries no attributed quotes. Accenture shares rose about 8 percent in after-hours trading on the announcement. The expectation this defeated: outside observers had assumed the role would go to a dedicated evaluation organisation such as METR, Apollo Research or Redwood Research, and Anthropic's stated reasons for choosing otherwise were Accenture's practical experience deploying AI inside large companies and governments, and its independence as a large public company that predates the AI revolution. Critics quoted in the coverage read embedded evaluation as industry self-policing designed to avoid accountability for model behaviour; Anthropic's answer is that evaluators make accountability more verifiable while the lab keeps ultimate responsibility. The commitment this fulfils: Dario Amodei's essay We Must Pace the Frontier, which proposed embedding evaluators inside frontier labs. The parent action: on September 12 Anthropic said it would give outside inspectors desks, badges and company laptops, and OpenAI said the same thing hours later, carded in Edition 020. OpenAI has not yet named an evaluator. There is no published standard anywhere for an embedded evaluator's access rights, publication rights or independence.

Established

Anthropic published on September 18, 2026 that Accenture, through Faculty, is its first embedded evaluator.
The published scope includes evaluating and red-teaming models, alignment assessments, safeguard testing, observation of model development during training and of deployment decisions, and assessment of company operations and safety commitments.
The two organisations state an expectation of at least one billion dollars invested in evaluation capacity over five years, with Anthropic funding Accenture's work initially.
The arrangement is non-exclusive on both sides, and Anthropic says further evaluators will follow in the coming weeks.

Not established

That the evaluator is independent in any enforceable sense. Anthropic selects it, funds it and sets its access, and no published term gives Accenture the right to publish a finding Anthropic objects to.
That one billion dollars will be spent. It is a stated expectation over five years by both parties, not a committed contract value disclosed in a filing.
That this satisfies any regulator. No authority has recognised embedded evaluation as meeting an obligation, and the EU AI Act's systemic-risk regime does not name it.
That Accenture is unqualified. Faculty has a decade of applied AI work behind it, and the objection raised in the coverage is about incentives, not competence.

Why it weighs this much

Signal strength

CostHigh

CommitmentHigh

ReversibilityMedium

PersistenceYears

Cost is high: at least one billion dollars over five years, with Anthropic funding the work at the start, is real money spent on people whose job is to find fault with the payer. It also cost Anthropic the easier, better-received option of appointing a safety institute. Commitment is high because a named counterparty, a named delivery arm and a five-year figure are all checkable, and because the access list is specific enough to be tested against: training observation and deployment observation are things an evaluator either gets or does not. Reversibility is medium. The contract can end, the scope can narrow quietly, and non-exclusivity means either side can walk without breaking anything public. Persistence is years because the first arrangement of its kind becomes the shape others copy or react against, and because a five-year capability build does not unwind in a quarter.

What to watch

strengthens Anthropic names a second embedded evaluator, or OpenAI names its first, or any embedded evaluator publishes a finding independently of the lab it sits inside, by March 31, 2027.
weakens No second evaluator is named by March 31, 2027, or the Accenture arrangement ends, narrows its published scope, or is shown to carry no right for the evaluator to publish or escalate outside the company, by June 30, 2027.

The words

Embedded evaluator: An outside team working inside an AI company with access to its models and internal decisions, rather than testing finished products from outside. The idea is that an inspector who is present sees what an external tester cannot.
Red-teaming: Deliberately attacking a model to find the ways it fails, in order to fix them before someone else finds them.
Alignment assessment: Testing whether a model pursues what its developers intended rather than something adjacent that scores well. It is distinct from testing whether a model is accurate.
Non-exclusive: Neither party is tied to the other. Anthropic can hire more evaluators and Accenture can evaluate rival labs, which is both a check on capture and a route for information to travel.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.