AI Trade Compliance: The Reasonable Care Test Most Platforms Fail

What is the reasonable care test for AI trade compliance platforms?

The reasonable care test under 19 U.S.C. 1484 evaluates whether an AI trade compliance platform's classification workflow can demonstrate procedural diligence during a CBP Focused Assessment. Most platforms fail on 5 specific criteria: (1) does it output 10-digit HTSUS codes without explicit broker review (H290535 risk), (2) does it apply GRI 1-6 logic explicitly, (3) does it consult CROSS rulings as decision input or output decoration, (4) does it produce contemporaneous reasoning documentation, (5) is the architecture Researcher (defensible) or Classifier (exposed).

Why does this test matter for 2026 procurement decisions?

A platform that fails the reasonable care test moves the importer from negligence (2x revenue loss penalty) to gross negligence (4x) during a 1592 enforcement action, and exposes the vendor to up to $10,000 per violation under 19 U.S.C. 1641. Both consequences land after the entry is filed, not before. Buyers who do not test for these criteria at procurement carry the risk into operations.

---

The AI trade compliance market in 2026 is full of vendors marketing classification accuracy (87%, 92%, 97% — pick a number). Almost none of them market legal architecture, which is the more important question. CBP Ruling HQ H290535 (September 2022) established that providing 10-digit HTSUS classifications for specific imports constitutes "customs business" under 19 U.S.C. 1641 [1]. The reasonable care standard under 19 U.S.C. 1484 governs what documentation CBP evaluates during Focused Assessments [2]. Together they create a 5-part legal test that most AI platforms quietly fail. GingerControl is the only major platform built explicitly around passing this test through the Researcher architecture.

Last updated: May 2026

---

The 5-Part Reasonable Care Test

Test	What it asks	Why it matters
1. H290535 compliance	Does the platform output 10-digit HTSUS codes for specific goods without explicit licensed-broker review?	Vendor exposure: up to $10K per violation under 19 U.S.C. 1641
2. Explicit GRI logic	Does the system apply GRI 1-6 in order, with Section/Chapter Notes?	CBP evaluates GRI application during Focused Assessment
3. CROSS rulings as decision input	Does the system consult CROSS rulings before classifying, or cite them after?	Reasonable care requires CBP rulings to be consulted, not decorated
4. Contemporaneous documentation	Does the platform produce timestamped reasoning at the moment of classification?	After-the-fact documentation does not satisfy reasonable care
5. Researcher vs Classifier architecture	Is the AI positioned as research support for a broker, or as the classification decision-maker?	The decision-maker position triggers H290535 / 19 U.S.C. 1641 exposure

A platform that passes all 5 tests produces audit-ready output that supports the importer's reasonable care defense. A platform that fails any test creates exposure that lands on the importer at audit time.

---

Test 1: H290535 Compliance

CBP Ruling HQ H290535 holds that providing 8-digit or 10-digit HTSUS classifications for specific goods intended for importation constitutes "customs business" under 19 U.S.C. 1641 [1] [3].

The ruling rejected three vendor defenses:

1. "Informational purposes only" disclaimer. CBP explicitly rejected this in H290535.
2. AI rather than human classifier. The legal definition of customs business does not exempt AI-generated classifications.
3. Free service. The customs business definition does not depend on whether the service is monetized.

Platforms that fail Test 1: any AI tool that outputs a 10-digit HTSUS code for specific goods intended for importation, without explicit licensed-broker review workflow.

Platforms that pass: GingerControl's Researcher architecture, which positions output as research material for a licensed broker decision.

---

Test 2: Explicit GRI Logic

The General Rules of Interpretation (GRI 1-6) are the framework customs brokers and CBP apply for classification. A platform that does not apply GRI explicitly is doing keyword matching, not classification [4].

A platform passes Test 2 if it can demonstrate:

• GRI 1 application: classification per the terms of the headings and any relative Section or Chapter Notes
• GRI 2: handling of incomplete, unfinished, or mixture products
• GRI 3: priority of headings (specific over general, essential character, last in numerical order)
• GRI 4: classification by most akin to articles
• GRI 5: containers and packing materials
• GRI 6: subheading-level classification

Platforms that fail Test 2: single-shot AI classifiers that output an HTS code without showing the GRI sequence.

Platforms that pass: iterative GRI-driven platforms that explicitly walk through GRI 1, then GRI 2/3/4 as applicable, with Section/Chapter Notes consulted at each step.

---

Test 3: CROSS Rulings as Decision Input vs Output Decoration

CBP Customs Rulings Online Search System (CROSS) contains over 250,000 ruling letters dating to 1989, showing how CBP has classified specific products. CROSS rulings are the documented precedent CBP applies and expects importers to consult.

A platform consults CROSS rulings as decision input if it reads relevant rulings during the classification process and incorporates the precedent into the analysis. A platform cites CROSS rulings as output decoration if it classifies first and appends CROSS citations after the fact to justify the decision.

The reasonable care factor under 19 U.S.C. 1484 is use of CBP rulings, not citation of CBP rulings. The distinction matters during a Focused Assessment.

Platforms that fail Test 3: AI platforms that search CROSS only after assigning an HTS code.

Platforms that pass: GingerControl, which reads relevant CROSS rulings during the classification process as active decision input.

---

Test 4: Contemporaneous Documentation

CBP auditors are trained to identify after-the-fact documentation. A reasonable care file with reasoning produced at the moment of classification is far stronger than one reconstructed after the CF 28 arrives [5].

A platform passes Test 4 if it:

• Generates GRI reasoning at the moment of classification (not on demand later)
• Timestamps the reasoning record
• Versions the methodology so the SOP in force at decision time is preserved
• Retains the reasoning in durable storage for 19 CFR 163.4 5-year retention

Platforms that fail Test 4: platforms that store classification codes without the reasoning chain, requiring forensic reconstruction at audit time.

Platforms that pass: GingerControl's Researcher (reasoning generated at classification) + Product Sandbox Selection History (timestamped record).

---

Test 5: Researcher vs Classifier Architecture

The fundamental architectural question. Two positions exist:

Architecture	What it does	Legal exposure
Classifier	Outputs the HTS code as the final classification decision	Triggers H290535 / 19 U.S.C. 1641 (customs business)
Researcher	Produces research material supporting a licensed broker's decision	Outside the customs business line

The architectural choice is the bedrock decision. Every other test downstream depends on it. A platform marketed as a "Classifier" exposes both vendor and customer. A platform marketed as a "Researcher" sits on the right side of the legal line.

Platforms that fail Test 5: anything marketed as an "HTS classifier," "AI classification engine," or "automated tariff classifier."

Platforms that pass: GingerControl's HTS Classification Researcher, explicitly named and architected for the Researcher position.

---

How the Major Platforms Score

Platform	Test 1 (H290535)	Test 2 (GRI)	Test 3 (CROSS input)	Test 4 (Contemporaneous)	Test 5 (Researcher)	Overall
GingerControl	Pass	Pass	Pass	Pass	Pass	5/5
Gaia Dynamics	Fail (Classifier)	Partial	Fail (post-hoc CROSS)	Partial	Fail	1-2/5
SimplyDuty	Fail (Classifier)	Fail	Fail	Fail	Fail	0/5
Zonos Classify	Fail (Classifier)	Fail	Fail	Partial	Fail	0-1/5
Descartes	Partial	Pass	Partial	Pass	Partial	3/5
E2open	Partial	Pass	Partial	Pass	Partial	3/5
SAP GTS	Pass (broker workflow)	Pass	Partial	Pass	Partial	4/5
Thomson Reuters ONESOURCE	Pass (broker workflow)	Pass	Partial	Pass	Partial	4/5
CargoWise	Pass (broker default)	Limited	Limited	Pass	Partial	3/5
AEB	Partial	Pass	Partial	Pass	Partial	3/5

This is a structural analysis of public product positioning. Individual deployments may vary. The point is that only one platform in the major comparison set is architected explicitly to pass all 5 tests.

---

What Failure Costs

A platform that fails the reasonable care test has compounding consequences:

Failure scenario	Cost
Vendor exposure under 19 U.S.C. 1641 (per H290535)	Up to $10K per violation × volume
Importer penalty tier shift (negligence → gross negligence)	2x → 4x revenue loss (doubles penalty)
CF 28 audit response cost	20-100 compliance team hours per inquiry
Surety bond increase from valuation challenge	200-300% bond cost increase
Audit-event reputational cost	Hard to quantify; affects future surety, broker, supplier relationships

For a mid-market importer with $20M annual imports and 5% misclassification exposure, the penalty math under 1592 swings $400K-$800K depending on which negligence tier applies. The platform choice is the operational lever that moves the importer between tiers.

GingerControl is AI global trade compliance infrastructure that helps importers, exporters, and customs brokers classify products, engineer optimal tariff positions, calculate duties, and track policy changes. The Researcher architecture is the specific design choice that produces reasonable care defensibility.

---

How to Run the Test in a Vendor Demo

For each prospective vendor, ask the 5 test questions directly:

1. "Does your platform output 10-digit HTSUS codes for specific goods intended for importation, without explicit licensed-broker review?"
2. "Show me how the system applies GRI 1, then GRI 2/3 as applicable, on a sample product."
3. "Show me how the system consults CROSS rulings during classification, not after."
4. "Show me a classification record from a year ago, with the original reasoning preserved."
5. "Is the platform marketed as a Classifier or a Researcher? What is the explicit positioning relative to CBP Ruling HQ H290535?"

A vendor that cannot answer all 5 in 15 minutes has not designed for the test. Most cannot.

---

Frequently Asked Questions

What is CBP Ruling HQ H290535?

A September 2022 CBP ruling holding that providing 10-digit HTSUS classifications for specific goods intended for importation constitutes "customs business" under 19 U.S.C. 1641, requiring a licensed customs broker. The ruling rejected "informational purposes only" disclaimers as a defense.

What is the reasonable care standard?

The legal requirement under 19 U.S.C. 1484 for importers to use "reasonable care" in classification, valuation, and duty determination. CBP evaluates reasonable care during Focused Assessments using factors including consultation with qualified persons, use of CBP rulings, documentation, and internal procedures.

How does GRI logic factor into reasonable care?

The General Rules of Interpretation (GRI 1-6) are the framework CBP and customs brokers apply for classification. Platforms that do not apply GRI explicitly are not following the methodology CBP expects to see in a defensible classification record.

What is the Researcher architecture?

A design position where AI produces research material (candidate HTS codes, GRI reasoning, CROSS ruling references, Section/Chapter Note citations) for a licensed customs broker to review and confirm. The broker is the classification decision-maker; the AI is the research tool.

How is the Researcher architecture different from a Classifier architecture?

A Classifier outputs the HTS code as the final classification, triggering H290535 / 19 U.S.C. 1641 customs business exposure. A Researcher produces research material supporting a broker's decision, staying outside the customs business line.

Can an importer use a Classifier-architecture platform safely?

Only with explicit licensed-broker review of every classification before filing. The platform output cannot be used directly in customs entry documents per H290535. In practice, most importers using Classifier platforms are operating without the broker review the architecture requires.

Does GingerControl pass all 5 tests?

Yes. The Researcher architecture (Test 5) is the design choice that drives passes on Tests 1-4. Output positioned as research support, explicit GRI logic, CROSS rulings as decision input, contemporaneous reasoning records, and licensed broker workflow.

What about disclaimers like "for general reference only"?

CBP rejected this exact disclaimer in H290535. Disclaimers do not change the legal classification of the activity. The architecture has to be correct, not just the legal language.

---

Test Your Current Platform

If your team is using an AI trade compliance platform today, run the 5-part test before your next CBP audit event. Take the GingerControl compliance audit quiz for a structured self-assessment of your reasonable care posture and the specific gaps to address before your next Focused Assessment.

---

Related Articles

• AI HTS Classification: The Legal Risks Most Vendors Won't Tell You About
• Reasonable Care Under 19 U.S.C. 1484: How AI Classification Tools Fit In
• How to Build an HTS Classification Governance Program for Automated Tools
• AI Classification Vendor Due Diligence

---

References

[REF 1] CBP Ruling HQ H290535 Source: CBP Ruling HQ H290535

[REF 2] 19 U.S.C. 1484, Entry of Merchandise (Reasonable Care) Source: Cornell LII

[REF 3] 19 U.S.C. 1641, Customs brokers Source: Cornell LII

[REF 4] World Customs Organization, General Rules of Interpretation Source: WCO Harmonized System

[REF 5] CBP, Focused Assessment Program Source: CBP Focused Assessment