Your AI office cannot review everything. Borrow the embedded-evaluator design.

A central AI team in a large company usually has the same problem. There are a handful of people at the center and dozens of AI projects out in the business units. The team writes the standards, but it cannot sit in every sprint review, read every prompt or watch every release. So review turns into a form the units fill in at the end, and the form says little about how the system was actually built.

Two announcements from this fortnight suggest another way to arrange it.

What Anthropic agreed

Anthropic announced a partnership with Accenture for independent evaluation of its frontier models. The work will be led by Faculty, Accenture’s specialist AI business, and covers red-teaming, alignment assessments and testing of model safeguards. Anthropic says each company expects to invest at least $1 billion in building capacity in this area over the next five years. That is a company-reported figure.

The design matters more than the money. Anthropic writes that embedded evaluators “will work inside AI companies, with access comparable to an employee’s.” They can watch models take shape in training, follow the decisions about how models are built and deployed, speak directly to staff, report incidents and give the public an account of benefits and risks. Anthropic also says the evaluators do not reduce its own accountability; they help make it more verifiable.

Anthropic is candid about what is missing. There are no standards yet for what evaluators should see or how they should report, and no settled way to fund the work. The partnership is not exclusive.

To be clear, this is about Anthropic’s own models. It is not a service Accenture sells to your business units. The useful part for a corporate AI office is the shape of the arrangement.

A second pattern: build outside, fold in later

The same day, TechCrunch reported that Vantora, formerly UP.Labs, raised $100 million from Silversmith Capital Partners. Vantora builds startups for corporate partners, which invest in the ventures and become their first customers. The change is that partners can now fold those startups into their core business and keep them. Named partners include Porsche, Alaska Airlines, J.B. Hunt, Wabash and TDG, the parent of Ashley Furniture.

Its founder, John Kuolt, put the reason plainly for industrial work: “You need to own that intelligence layer.”

Put the two stories side by side and the common thread is clear. The capacity comes from outside. Access, reporting and the right to own the result are the terms to set at the start.

An embedded-review charter for your business units

Here is how a central team of a few people can apply that design inside a company. Instead of reviewing every project at the end, place one reviewer inside a unit’s AI work for a set period, with a written charter. The charter answers seven questions:

1. Access. What can the reviewer see? The prompts, the evaluation results, the incident log and the working channels, not just the demo.
2. Independence. Who does the reviewer report to? The central team, not the unit’s budget owner.
3. Scope. Which named systems are covered? Everything the unit runs is too broad to check.
4. Findings. How are they written? Dated, specific, with a named person to respond and a due date.
5. Incidents. Can the reviewer raise one directly, without the unit’s sign-off? It should be yes.
6. Term. When does the placement end, and what does the unit keep afterwards? Usually the tests, the checklists and one trained person.
7. Report. What goes to leadership? A short account of what works, what does not and what changed as a result.

Keep, place or buy

A simple rule for deciding what the central team does itself:

• Keep at the center anything you would need in order to explain a decision to the board or a regulator: the standards, the vendor contracts, the review charter, and the final say on what goes into production.
• Place in units reviewers and platform engineers, for a set term, under a charter like the one above.
• Buy from outside surge capacity: specialist red-teaming, evaluation of a new vendor, or building a new venture. Agree ownership and fold-in terms before the work starts. Vantora’s partners now have the option to fold a startup into the core business and keep it.

If losing a capability would leave you unable to say why a system is safe to run, it stays inside. Everything else can be borrowed.

What to do this week

• List the AI systems your units run in production and mark which ones the central team has actually seen working, not just read about.
• Choose the riskiest one and draft a one-page charter for a four-to-six-week embedded review, using the seven questions.
• Check your outside AI contracts for who owns what gets built and whether the work can be brought in-house later.
• Tell unit leaders that review is moving from a form at the end to a person in the room, and why.

None of this needs a bigger team. It needs a clearer arrangement about who sees what, who they report to, and who owns the result.