How IT Leaders Can Audit AI Vendor Promises and Protect Sovereign Data

Auditing AI Vendor Promises to Protect Sovereign Data
Auditing AI Vendor Promises to Protect Sovereign Data

Written by Robby Jakovljevic, Director of IT, Paskenta Band of Nomlaki Indians in General Articles

Date: November 17, 2026

Every year in London, a nation's new coins are put on trial. It is called the Trial of the Pyx, and it is a real court. Throughout the year, the Royal Mint sets aside sample coins from its production in fixed proportions: one coin from every batch of each denomination. 

In the early months of the year, those coins are brought before a jury of assayers from the Goldsmiths' Company, with the King's Remembrancer, a senior judicial officer, presiding. The jury has weeks to weigh and test them. Then the court reconvenes, and the verdict is read: whether the coins the nation carries in its pockets are what the Mint says they are.

The trial goes back to the 12th century. The earliest surviving record of a public trial of the coins is dated Feb. 24, 1248. The name comes from the pyx, the wooden box that held the sample coins. In the old days, a Master of the Mint whose coins failed could be fined, removed from office, or imprisoned.

The Royal Mint Museum describes the trial in one phrase that I think every IT leader should keep on a card: an examination by "a jury entirely independent of the Royal Mint”. The Mint does not grade its own coins. It hasn't for nearly eight hundred years.

Our Coins

I'm the Director of IT for the Paskenta Band of Nomlaki Indians in Northern California. Most weeks, someone hands me a coin and tells me what it's made of.

"We don't train our AI on your data."

"Your data stays in the United States."

"Everything is encrypted, at rest and in transit."

"Your security score is excellent."

Each of those is a promise about something I can't see. And most of the time, the only party telling me the promise is true is the party that made it. There are numbers in my own reports that I no longer believe. One of them is a vendor's own score for how secure my environment is, which somehow rises when I buy more of that vendor's features. Nobody is lying when that number goes up. But that is the Mint grading its own coins.

Promises Change

The coin I worry about most right now is the first one. "We don't train on your data." For a tribal government, that promise matters more than it does for most customers. The data we hold includes enrollment records, health records, government records, and in some cases, cultural materials that belong to the nation and nobody else. If that data goes into the training of someone's model, there is no getting it back out.

And a promise is only as good as the terms it's written in, on the day you read them. In February 2024, the Federal Trade Commission published a warning to technology companies titled "AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive". 

The agency's staff wrote that a company may be acting unfairly or deceptively if it starts using customer data in new ways, including for AI training, and tells people only through a quiet, retroactive change to its terms or privacy policy. They ended with a line I wish every vendor would read: "there's nothing intelligent about obtaining artificial consent." That warning tells you two things. The regulators saw it happening often enough to write about it. And it happens in the terms of service, the document almost nobody reads at renewal.

The Independent Jury

So who tests the vendor's coin? Our industry does have juries. The best known is the SOC 2 report, an examination of a vendor's controls by an independent CPA firm, covering some combination of security, availability, processing integrity, confidentiality, and privacy. A Type II report covers how those controls actually worked over a period of time, not just how they were designed. For AI specifically, an international standard, ISO/IEC 42001, was published in December 2023 that sets requirements for how an organization manages its AI systems, and vendors can be certified against it.

Those are real juries, and I ask for them. But a jury only rules on the charge before it. A SOC 2 report tells you whether the controls in its scope were operating. If nobody put "customer data is never used for model training" in that scope, the report says nothing about it, however thick it is. 

An AI management certification tells you a vendor has a system for governing its AI. It does not, by itself, tell you what that system decided to do with your data. Read the scope and the date before you read the conclusion. The most important thing to look for in an attestation is not a failure. It is a silence about the one promise you actually care about.

The Trial Plate

There is a story from the Trial of the Pyx that I like even more than the trial itself. In 1710, Isaac Newton was Master of the Mint. The jury tested his gold coins and found them below standard. Newton did not accept the verdict. He went after the measuring stick. He showed that the reference the coins had been tested against, a gold trial plate made in 1707, itself contained too much gold. The coins were fine. The standard was wrong. The older trial plate, from 1688, went back into use. So independent testing has its own limits. It is worth nothing if the standard you test against is wrong for what you're measuring.

That is the tribal version of this problem. Most of the frameworks vendors are certified against were written for companies. They assume customers, not citizens. They treat stored data as a liability to be minimized, not as the records of a nation that intends to exist permanently. 

A vendor can pass every test in the book and still fail the one that matters to a sovereign, because the book was never written with a sovereign in mind. Sometimes, like Newton, you have to check the trial plate.

Writing My Own Charge

What I've done about it is write the charge myself. I've drafted what I call a Tribal Sovereignty Vendor Addendum, a working baseline for contracts that touch tribal data. It isn't my organization's actual contract, and I'm not a lawyer. It is a practitioner's starting point, and any tribe using language like it should work through it with its own counsel.

Two of its clauses do the work of the Trial of the Pyx. The first is the promise itself, in the contract instead of the brochure. The vendor, and anyone the vendor hires to process the tribe's data, may not use tribal data to train or modify any AI model, under any circumstances, whether or not the data has been de- identified, anonymized, or aggregated.

The second is the jury. The vendor provides its current third-party attestations on request. The tribe, or an auditor it chooses, can audit the vendor's controls once a year, and again after any security incident. Material findings have to be fixed, and failing to fix them is a breach. When the relationship ends, the vendor certifies in writing that the tribe's data has been destroyed.

It is meant to be attached to a vendor's agreement, not used to rewrite it clause by clause. In my experience, most vendors can process a well-drafted addendum faster than a line-by-line redline of their own paper.

Where I work, there is one more jury. Every AI tool that wants into our environment has to be approved by our Tribal AI Committee first, and the committee, not IT, decides whether to keep it. A vendor's promise about AI gets read by people across the organization, not just the person who wanted to buy the tool.

For The Next Renewal On Your Desk

Here is how I'd test a vendor's coin, in the order I'd do it. Get the promise in the contract. If it only appears in a sales deck, a blog post, or a trust page, it can change without you. Ask for the attestation, and read its scope and dates before it concludes. Look for the promise you care about. If it isn't in scope, ask why not.
Keep the right to audit, and say when you will use it: once a year, and after any incident. Read the terms at renewal, not just the price. On our team, someone reads what changed in the terms every time a contract renews. The FTC's warning reminds us where the quiet changes happen.

Check the trial plate. Ask whether the standard a vendor passed was written for an organization like yours. If it wasn't, write your own charge. None of this assumes vendors are dishonest. Most of the people I work with on the vendor side are doing their jobs well, and the best of them welcome this kind of question. The Royal Mint still sends its coins to trial every year, nearly eight centuries on, and nobody thinks it is because the Mint is crooked. It is because trust that gets tested is the only kind that holds its value. The coins go on trial so that the people carrying them don't have to wonder. We owe the people whose data we carry the same thing.