HQ Ruling H350722: What AI Can and Can't Do in Customs Work

What the ruling actually says

HQ H350722 addresses the application of 19 CFR to AI-driven tools. The ruling restates what counts as "customs business" under existing federal regulation, then applies that definition to AI systems.

Customs business, per 19 CFR, includes any activity involving transactions with CBP concerning the entry and admissibility of merchandise. Specifically:

• Classification at the 10-digit level
• Valuation of merchandise for entry
• Payment of duties, taxes, or other charges
• Refunds, rebates, and duty drawback
• Preparation of documents or forms in any format, and activities relating to such preparation

The ruling is explicit. AI tools cannot perform customs business in the entry lifecycle. The plain reading is that automated systems are not permitted to make the legally binding determinations that licensed customs brokers are authorized to make.

What this means in practice

For any AI tool that touches the entry process, the ruling creates a hard stop on several specific activities:

• Classification at the 10-digit level. An AI system cannot make the final HTS determination on an entry. The classification must be made by a licensed broker or an importer with a self-filer license.
• Valuation determinations. AI cannot determine the value of goods for entry. Valuation involves legal interpretation that the ruling reserves for licensed personnel.
• Duty calculation and payment. AI cannot compute and submit duty payments as the system of record. Calculation can be assisted, but the binding determination must come from a human.
• Refunds and drawback. AI cannot file refund or drawback claims on behalf of an importer. This includes the IEEPA refund process and any future refund programs.
• Document preparation tied to entry. The "activities relating to" language in 19 CFR is broad. Any preparation directly tied to an upcoming filing is captured.

The permissible path: product database screening

The ruling does more than restrict. It clarifies where AI can legitimately operate, and the clarification is significant.

The permissible path is a product database model. AI is permitted to assist with classification, valuation, country of origin, and compliance screening when the work is happening in advance of an entry. The work lives in the importer's product database, ERP, OMS, or other internal system.

In this model, AI does not perform customs business. It maintains a clean, current, defensible product master. When an entry is later filed, a licensed broker references that database to make the final determination.

This is the path that delivers the operational benefit of AI without crossing the statutory line. The work shifts from "the AI files the entry" to "the AI keeps your product master compliant, current, and audit-ready."

Why this matters for vendor selection

The ruling makes vendor evaluation easier in one specific way. Any vendor whose product diagram includes "the AI files the entry" or "the AI computes the duty" is offering a product that violates federal regulation. That should end the conversation.

Useful AI in trade does several things, none of which involve filing the entry itself:

• Maintains product database accuracy through continuous monitoring
• Surfaces classification gaps, missing PGA data, and 99-code sequencing errors before filing
• Audits entries after the fact to catch overpayment and underpayment
• Screens vendors and counterparties against sanctions and restricted party lists
• Models duty exposure under different sourcing scenarios

Each of these activities supports the entry lifecycle without performing customs business as defined by 19 CFR.

The auditability requirement

The ruling reinforces a requirement that already existed in practice. Every determination needs to be defensible. When CBP follows up on an entry, the importer or broker has to produce the reasoning, the data, and the citations behind a given classification, valuation, or origin call.

AI tools that produce determinations without source citations, without confidence scoring, and without an audit trail create exposure even when used in the permissible path. A product database populated by an AI that cannot explain its own decisions is not defensible.

For any AI tool entering a trade compliance tech stack, the minimum bar is:

• Source citations. Every determination cites the specific regulation, ruling, GRI, or internal policy that supports it.
• Confidence scoring. Every output includes a measure of certainty. Low-confidence items trigger human review.
• Audit trail. Every decision is reconstructable. CBP must be able to walk the chain from the final entry back to the underlying data and reasoning.

Without these three properties, the AI is producing plausible answers that cannot be defended under scrutiny.

What changes for current AI implementations

For importers and brokers already using AI tools in trade work, the ruling requires a quick audit.

• Is the AI making final determinations that get filed with CBP? If yes, that workflow is non-compliant and needs to be restructured. The AI moves upstream into the product database. The licensed personnel make the binding call at filing.
• Is the AI assisted by a human-in-the-loop process? If the human is reviewing every determination, the workflow is likely compliant. If the human is rubber-stamping AI output without independent review, that is functionally automated customs business and creates exposure.
• Can the AI produce a defensible audit trail for every determination it has supported? If not, the gap needs to be closed before the next audit cycle.

The forward-looking read

HQ H350722 did not break new ground. It applied existing 19 CFR definitions to a new class of tools. The ruling is consistent with the broader trajectory of customs enforcement: more scrutiny, better detection, and a higher bar for defensibility.

The ruling is also clarifying in a constructive way. It tells the market exactly where AI belongs in the entry lifecycle and where it does not. For vendors who built their products around legitimate use cases (product database accuracy, screening, scenario modeling, audit support), the ruling validates the model. For vendors who marketed end-to-end automation of the entry filing process, the ruling forces a redesign.

Importers and brokers should be reading the ruling not as a restriction but as a clarification. The permissible path is broad, the operational benefit is real, and the boundary is now drawn in writing.