QAD, in conjunction with AWS, recently commissioned a third-party firm, Dimensional Research, to survey compliance leaders across the global trade community to gauge the true impact of trade volatility and the use of artificial intelligence, in a formal report released in July, 2026. Two findings frame everything that follows. First, 79% of respondents are already using AI to manage foreign trade zone operations, and 99% are either using it or planning to. Second, only 20% have software that updates automatically when tariff changes are issued. The rest rely on manual updates or are frustrated by the lack of automation.
The industry is pouring money into AI, and still losing ground on tariffs. That’s neither a contradiction nor an indictment of AI. The investment instinct is right, but there’s a gap created by AI that’s layered on antiquated systems that can’t store or apply today's tariff structures, which consequently will never deliver, no matter how good the intelligence is on top.
The core difficulty is determining which tariffs apply to which product, country and time period. Over the past 16-18 months, that determination has become increasingly difficult to make.
Consider what's been layered on in a short span: imposition of the International Emergency Economic Protection Act (IEEPA) tariffs, expanded Section 232 and Section 301 actions, the Supreme Court's striking down of IEEPA tariffs in February, the Section 122 tariffs that replaced them four days later, the Brazil 301s, the forced labor 301s, and now the new polysilicon derivative 232s.
Each addition is a new set of codes under Chapter 99 of the U.S. Harmonized Tariff Schedule that stack on top of a product's primary classification, differing by country of origin, applying only during specific effective periods, and coming with annexes of carve-outs, some written to a single HTS code, others sweeping in an entire chapter or subheading.
What trade teams are really maintaining is a living map, incorporating every product and its primary classification, cross-referenced against an ever-expanding list of countries, punitive tariff programs and exceptions. The cells in that grid have multiplied many times over, but the people maintaining it have not.
The Clock Has Changed Too
Historically, U.S. Customs and Border Protection gave the trade advance notice. A modification was announced, and teams had time to assess the impact, update systems and brief their brokers. That cushion is gone.
The forced labor 301s are the clearest example. The final action, with its actual rates, country tiers and exemption annexes, was released on July 23 and took effect at 12:01 a.m. the next morning, the very minute the temporary Section 122 tariffs expired. It touched 60 trading-partner economies, each with its own tariff treatment and Chapter 99 reporting requirements, and the annexes of exceptions had to be parsed, interpreted and mapped against a company's own product catalog with essentially no time to respond.
That is an interpretation-and-systems problem, executed under a deadline measured in hours.
Much of the tariff determination problem is deterministic. If you know the primary classification, country of origin and entry period, the applicable stacked tariffs can be derived, provided your systems can store and apply that logic. AI can genuinely help teams understand what a change means and decide how to respond, but the core calculation is rules, not generation.
Which brings us to the real bottleneck: legacy systems of record. If your system can’t even store the stacked tariff structure that Customs now requires, no amount of AI layered on top will save you. The platform underneath has to absorb a change the same week, sometimes the same day, that Customs issues it.
The pairing is the point. AI working against a capable, current system of record is a genuine force multiplier; working against an antiquated one, it just produces wrong answers faster.
The Consumer-Grade AI Trap
Go back to those two survey findings: near-universal AI adoption, but only one in five organizations with software that automatically absorbs tariff changes. The most plausible explanation is that much of the AI in use today is consumer-grade: general-purpose chat tools pressed into service for classification research, document review and regulatory interpretation. Adoption is racing ahead of readiness.
General-purpose tools lack two things compliance work requires: proper context and trusted sources of truth. The failure mode is subtle. An analyst works through one product's classification in a chat session, then asks about an unrelated product in the same session. The prior context quietly skews the output. The answer looks confident and is wrong for reasons the user can’t see, while the tool may be filling gaps from web sources of unknown reliability.
Using tools like that for official compliance work increases risk invisibly.
The alternative is call compliance decision memory — AI workflows grounded in the company's own compliance decision history. Every classification, ruling and valuation position taken becomes trusted context the agent reasons against. It turns tribal knowledge, the kind that lives in the heads of two or three veteran analysts, into an embedded, queryable part of the workflow.
It also enables what AI genuinely excels at: pattern recognition. With decision memory in the loop, an agent can flag a value discrepancy on an incoming product, notice that a supplier has changed from the historical vendor, or surface a country of origin the company has never sourced from before. Those catches used to depend on the right person happening to look at the right entry.
From Co-Pilots to Agentic Workflows
That’s why investment is shifting from chatbots and co-pilots toward agentic workflow capability — an operator paired with an AI agent that carries the company's compliance decision memory into every step.
Agentic, however, doesn’t mean autonomous. Regulators on both sides of the Atlantic have drawn the same line. In the U.S., a January, 2026 CBP headquarters ruling found that official Customs business can’t be handled start to finish by automated systems, whether AI-driven classification beyond the six-digit level or OCR-based document import. In Europe, the EU AI Act (Regulation 2024/1689) requires that high-risk AI systems, a category that reaches border and customs applications, be designed for effective human oversight. The obligation stays with the importer of record, and the decisions stay with accountable humans.
The design requirements follow: human-in-the-loop approvals at critical stages, configurable thresholds that determine when an agent proceeds and when it escalates, and full auditability so that any decision can be reconstructed and defended. Instead of being removed from the workflow, operations move up a level, managing exceptions instead of processing every line.
Given the current volume and pace of change, it’s no longer feasible for trade organizations to operate unassisted. That is simply arithmetic. And well-designed automation delivers two outcomes at once: Operational efficiency goes up and compliance risk goes down. Those are usually traded off against one another. Here, they move together, because consistency is exactly what both reward.
That’s the real case for AI in global trade compliance: empowering, not replacing compliance teams, with the institution's own compliance knowledge embedded in every decision the workflow makes. The formula is AI paired with a system of record built to absorb change as fast as Customs issues it. Get that pairing right, and the technology gap our survey uncovered stops being a liability, resulting in the biggest efficiency and compliance opportunity trade organizations have seen in years.
Joshua Guy is vice president of global trade compliance at QAD.
