Add five questions to your AI vendor review template this quarter. Not next year. Not when your legal team asks. The questions now have real answers, which is new, and that is reason enough to add them.
Here is the block to build: five questions about outside verification, grouped in your template under a named section, with a date-stamped row for each vendor's written response. Below is what changed that makes the block worth building now.
What Changed on September Eighteenth
Anthropic, the company behind the Claude family of AI models, announced a partnership with Accenture for independent embedded evaluation of its frontier AI models. Embedded evaluation means outside evaluators work inside the AI company with access comparable to an employee's: they watch models take shape during training, follow the decisions that govern how models are built and deployed, and speak directly to employees. They can report incidents and give the public a more informed account of benefits and risks. [1]
The work will be led by Faculty, Accenture's specialist AI business. It includes red-teaming models (red-teaming means deliberately attacking a system to find weaknesses), alignment assessments, and testing model safeguards. Anthropic and Accenture each expect to invest at least one billion dollars over the next five years [1]; Reuters reported the combined figure as two billion dollars. [2]
The deal is non-exclusive, meaning neither party is locked to the other alone. Anthropic says it will work with other evaluators to be announced, and Accenture will work with other AI developers. Anthropic is also in dialogue with METR and other nonprofit evaluators using their own funding. [1]
Why This Changes What You Can Ask
Until now, most mid-size companies buying AI tools have relied on vendor self-reporting. The vendor says its model is safe, reliable, and fit for business use; the buyer has no independent way to check. A company your size cannot send engineers inside a vendor's training lab. You have been taking the vendor's word.
The pressure to change this has grown. Regulators, enterprise buyers, and researchers have pushed AI developers to show that advanced models are safe and reliable, and that push intensified after incidents where AI agents broke out of secured environments. [2] An inside-the-lab, funded, independent evaluation layer is now starting to exist at one major vendor.
That does not settle anything, but it changes what a buyer can ask. Where "does an outside party verify your models?" used to produce a shrug or a marketing brochure, it now has observable answers.
That is the decision delta: the question went from unanswerable to checkable.
The Five-Question Outside Verification Block
Add these to your template as a named section called "Outside Verification." Leave a column for each vendor's written response, and date-stamp every answer.
- Who outside the vendor's company evaluates this model before we rely on it?
- What can those evaluators publish without the vendor's approval, and where have they published before?
- Who pays the evaluators?
- When an evaluator finds a problem, do customers hear about it, and how?
- If the answer today is "nobody yet," record the gap and lean harder on your own bounded pilot test before expanding use.
The Pushback Worth Answering
"Isn't this vendor marketing?" Maybe partly. Anthropic funds Accenture's work directly, which means the payer and the evaluated are connected. Anthropic itself says there are no standards yet for what access embedded evaluators should have or how they should report. No settled system exists for how this work gets funded across the industry. [1]
That is exactly why the checklist asks who pays and what can be published, rather than asking you to accept the announcement itself as assurance. The announcement tells you the category now exists. The questions tell you whether your specific vendor's version of it holds up.
One more limit: an evaluator working inside Anthropic does not certify the specific tool your team uses. Your workflow still requires its own acceptance test before you rely on it for anything consequential.
The One Thing You Do Next
Add the outside verification block to your AI vendor review template this quarter. Date-stamp each vendor's answers. A vendor who answers all five questions clearly is not automatically safe to use; a vendor who cannot answer any of them is telling you something useful right now.
The block takes thirty minutes to add. The answers follow you into every contract renewal and every expansion decision until better industry standards exist to replace them.