Let the Model Read. Keep the Pen Somewhere Else.

A model can read an invoice, find the relevant dates, and explain what a regulation says. The authority to define the rule, preserve the evidence, and approve the outcome belongs elsewhere.

On October 10, 2026, Microsoft CEO Satya Nadella described this separation in an essay on models as insider risks: “we need to separate the supply of intelligence from the authority over it.” Nine days later, Microsoft announced security controls built on the same principle, including task-scoped Dynamic Permissions and Guardian Agents that supervise agent activity. Hayete Gallot, who leads Microsoft Security, stated the premise plainly: “Unfortunately, frontier models, regardless of their origin, cannot be trusted.” The controls sit outside the model. (Nadella’s essay, Gallot’s post, Microsoft’s announcement)

For us, that principle applies directly to trade and regulatory operations.

Three things a model should not own

In a regulated workflow, authority comes in three pieces.

The rule. Which regulation applies, which version, and what it requires.

The record. Which documents support the review, where each fact came from, and what was verified.

The signature. Who approves the action, and who is accountable for it.

A model helps with all three. None of them lives inside it.

The rule gets a named, versioned definition with a citation. The record preserves source documents and traceable facts. The signature belongs to an identifiable person, with a record of what they approved.

A prompt, a context window, and a generated answer cannot carry those responsibilities.

One invoice, split three ways

Consider an illustrative case from trade operations, the kind TraderSage handles.

A US importer receives a demurrage invoice directly from an ocean common carrier. The FMC’s billing rules in 46 CFR Part 541 set three requirements that matter here:

• The carrier must issue the invoice within 30 calendar days of the date the charge was last incurred. If it does not, the billed party is not required to pay (§541.7(a)).
• The invoice must include required information: shipment identifiers, timing and rate details, dispute information, and certifications. Missing required information eliminates the obligation to pay the applicable charge (§541.5).
• The billed party gets at least 30 calendar days from invoice issuance to request mitigation, refund, or waiver (§541.8(a)).

Now split the job.

The model reads

It extracts the container number, free-time window, charged dates, and invoice date from a scanned PDF, and matches those facts to the supporting shipment documents.

This is where a model earns its cost: messy layouts, different carriers, different languages. Every extracted value links back to the line it came from. An unclear date or a conflict between documents becomes a review item before the value enters a calculation.

The rule computes

In this example, the verified last date charged is July 2, and the verified invoice issuance date is August 6. That is 35 days, past the August 1 deadline.

Once the inputs are verified and the rule is set, the calculation is arithmetic. The same verified inputs and the same rule version return the same result, whichever model read the document.

Required-field checks run the same way. A field’s presence does not prove its contents are accurate, though. That question goes to cross-document comparison or to a reviewer.

The person decides

The timing check flags the invoice as issued outside the regulatory window. The reviewer sees the source dates, the calculation, and the cited rule.

The next step is a judgment call. Request cancellation, ask for clarification, or escalate? How hard to push a carrier the team relies on for future sailings? That decision belongs to the person who carries the commercial consequences.

The model reduces the reading burden. The rule supplies a repeatable check. A person authorizes the response.

Built for the next model

Teams change models for cost, speed, and accuracy. A durable workflow absorbs those changes without letting the model redefine the rules.

The promise is narrow on purpose: same verified inputs, same rule version, same result. A different model may read a different date or pick a different candidate document. Those differences surface as review items, because a deterministic calculation cannot fix an input that was read wrong.

That is how NURIE AI is built. Our infrastructure routes work across models based on speed, cost, and accuracy, and it runs underneath VaultSage, TraderSage, and RegSage. VaultSage connects the document record. TraderSage and RegSage bring cited rules into the review. When the model changes, teams can inspect what changed in extraction. The rule changes only when someone updates it.

The output is a reviewable chain: source fact, applicable rule, result, open exceptions.

A person holds the pen.

Three questions to ask any agent vendor

If you are evaluating an AI agent for trade, customs, or regulatory work, ask three things.

1. Where does the rule live? It needs a name, a version, and a citation outside the prompt, plus a clear process for deciding when it applies.
2. What stays consistent when the model changes? The same verified inputs and rule version produce the same calculation. Differences in extraction are visible and reviewed.
3. Who approves the action, and what can they inspect? The source evidence, the applied rule, the result, and any open exceptions.

These questions reveal where the authority sits.

If your team reviews carrier invoices or import documents, bring a representative set to a walkthrough with us. We can show how the documents, cited checks, and human review fit together.

partnership@nurie.ai

FAQ

What did Satya Nadella mean by separating “the supply of intelligence from the authority over it”?

In an October 10, 2026 essay, Nadella argued that frontier models should be treated as insider risks, with controls over access and action placed outside the model. On October 19, Microsoft announced agent security controls built on that principle, including Dynamic Permissions and Guardian Agents.

What happens if the model reads a document incorrectly?

Extracted facts stay traceable to their source. Unclear values and conflicts between documents are flagged for review before they drive a rule-based conclusion. A repeatable calculation is only as reliable as its inputs.

What happens when a regulation changes?

Each check records the rule version it used and the basis for applying it. The review record preserves what was applied at the time, so a later update does not obscure how an earlier result was reached.

How long does a carrier have to issue a demurrage invoice in the US?

Under §541.7(a), the general deadline is 30 calendar days from the date the charge was last incurred. A separate provision applies to NVOCC billing: §541.7(b) uses the issuance date of the underlying invoice the NVOCC received. The billing party’s role determines which check applies. (46 CFR §541.7)

How long does a billed party have to request mitigation, refund, or waiver?

The billing party must allow at least 30 calendar days from the invoice issuance date. Check the invoice’s stated dispute process and deadline as well. (46 CFR §541.8)

Does TraderSage decide whether to dispute an invoice?

TraderSage supports document review and cited rule checks. A person decides whether to dispute the invoice and authorizes the response.