Earlier this year, U.S. Customs and Border Protection issued Headquarters Ruling H350722, which addresses something I have been thinking about quite a bit lately: what happens when AI starts doing more of the work traditionally performed by customs brokers?
The easy summary of the ruling is that CBP said AI cannot replace a licensed customs broker. The actual ruling is a little more nuanced than that.
H350722 involved an unlicensed foreign company operating an online platform for importers. Among other things, the platform connected importers with brokers, used OCR to pull entry information from shipping documents, and used AI to generate HTSUS classification suggestions.
CBP concluded that merely connecting an importer with a broker could be permissible. But the company crossed into regulated “customs business” when its software decided what information should appear on an entry, provided entry-related classifications beyond the six-digit level, or completed and submitted CBP Form 5106 on behalf of importers.
CBP also made a fairly consequential point: when an automated tool is conducting customs business, the actual decision about classification or other information needed for the entry ultimately needs to be made by a properly licensed person.
That is stronger than simply saying, “AI can do the work as long as a broker looks at it afterward.”
It also raises a question that I suspect is going to become much more important over the next few years.
First, what exactly is a customs broker?
If you run a business that imports products into the United States, you have probably dealt with a customs broker even if you don’t spend much time thinking about what the broker actually does.
A U.S. customs broker is a person or company licensed by CBP to conduct customs business on behalf of importers. Among other things, brokers help prepare customs entries, classify merchandise under the Harmonized Tariff Schedule, deal with valuation and admissibility questions, calculate duties, and communicate with CBP.
The classification part is especially important because the HTS code assigned to a product can determine how much duty you pay and whether other requirements apply.
Importers don’t legally have to use a customs broker. A company can conduct customs business on its own behalf. In practice, however, many small and midsized importers use brokers because customs rules are complicated and the consequences of getting them wrong can be expensive.
Becoming an individually licensed broker is not trivial. The Customs Broker License Examination is an 80-question, 4.5-hour open-book test, and you need 75 percent to pass. Passing the exam is then followed by the license application and CBP’s vetting process.
I have some scar tissue here. As I mentioned in my BrokerBench post, I took the broker exam years ago and missed passing it by a few points.
“Open book” sounds reassuring until you see the books.
Why keep the human in the loop?
The legal reasoning in H350722 is fairly straightforward. Congress and CBP regulations define certain activities as customs business, and customs business performed on behalf of somebody else generally requires a licensed broker. Software itself isn’t a licensed person.
There is also an obvious practical reason for keeping a human involved.
Customs decisions can have significant financial consequences. Classification affects duty. Valuation affects duty. Country of origin can affect duty and admissibility. Mistakes can result in penalties, and intentional misconduct can result in much more serious consequences.
Once you remove the human from the decision, accountability gets murkier.
With a licensed broker, there is somebody CBP can regulate, discipline, and hold responsible. The importer also retains its own legal obligations. From that perspective, keeping a human in the loop makes sense.
But there is an assumption underneath this model: the human is the expert and the AI is the imperfect assistant.
I’m not sure we can assume that will always be true.
What happens when the AI is better?
When I built BrokerBench, I wanted to see how well frontier AI models could perform on customs brokerage tasks. In my current runs, models such as Claude Fable and GPT-5.6 Sol have reached scores around 89 percent on what I deliberately designed to be a difficult collection of brokerage questions.
BrokerBench is still a design-review benchmark, so I don’t pretend that an 89 percent score means an AI is “89 percent as good as a customs broker.” It doesn’t. But the results are already good enough that I think the larger question is worth asking.
Modern models also have an advantage that is easy to underestimate. They don’t have to rely entirely on whatever knowledge was encoded during training. They can search regulations, CBP rulings, tariff material, government websites, product specifications, and other sources while they work.
A research problem that might send a human digging through references for an hour can sometimes be assembled by an AI agent in minutes.
The models are also improving quickly.
We are already seeing frontier models perform extremely well on difficult programming, mathematics, and professional reasoning benchmarks. Benchmarks are not the same thing as doing real work safely, but they do show how rapidly these systems are becoming capable of solving complicated problems.
It doesn’t seem far-fetched to imagine models eventually becoming faster and more accurate than most humans at routine entry-writing and classification work.
I don’t know whether that means outperforming 80 percent of entry writers, 95 percent, or 99 percent. The exact percentage isn’t really important.
The important point is that at some level of performance we are going to face an odd situation: the law may require a human to approve work produced by a machine that is consistently better at the task than the human doing the approving.
At that point, we may need to rethink what “human supervision” is actually accomplishing.
One path: keep the broker accountable
The most likely near-term outcome is probably the least dramatic one.
AI gets embedded deeper and deeper into customs workflows while the licensed broker remains responsible for the final result.
That is already happening throughout knowledge work. Accountants use AI. Programmers use AI. Attorneys use AI. The professional reviews the output and remains accountable for the work.
Customs brokerage will probably follow a similar path.
H350722 complicates this somewhat because CBP appears to be drawing a line before simple human approval. If an unlicensed third-party system is actually originating entry information or making entry-specific classification decisions, putting a broker at the end of the process may not automatically make the underlying activity permissible.
That distinction matters.
Suppose an AI system researches the HTS, reads the chapter notes, finds relevant CBP rulings, examines the product specifications, and recommends a classification. A broker carefully evaluates that work and makes the final decision. That is fairly easy to understand as AI-assisted brokerage.
But imagine the system becomes good enough that the broker almost never changes its answer.
At some point, requiring the broker to click Approve starts to look less like expert supervision and more like regulatory theater.
Another path: regulate the outcome instead
A more aggressive approach would be for CBP and Congress to gradually loosen the restrictions around who—or what—can perform certain customs tasks.
Keep the accountability.
Keep the penalties.
Keep the importer’s obligation to exercise reasonable care.
But allow companies more freedom to decide what combination of people, AI models, and software they want to use to meet those obligations.
If a company decides to classify millions of dollars of merchandise using a cheap AI model and it gets everything wrong, the importer still owns the consequences.
That isn’t completely different from choosing a bad customs broker today. There are excellent brokers and mediocre brokers. The same thing will be true of AI systems.
Over time, the market could sort out which activities are safe to automate and which still require substantial human expertise.
Getting there legally would not be simple. Some changes might be possible through agency rulemaking, while larger changes could run into the customs broker statute itself, 19 U.S.C. § 1641, and require Congress.
CBP gets AI too
There is another side of this discussion that may be even more important.
If customs brokers and importers can use AI, so can CBP.
Suppose private-sector systems become extremely good at HTS classification. CBP could use similar technology to look for classifications that appear unusual or incorrect.
That could turn into something resembling what has happened in cybersecurity, where increasingly capable technology gets used on both sides.
CBP also has an enormous informational advantage. It has years of historical entry data, enforcement history, rulings, audit results, and knowledge about how products have traditionally been classified.
It isn’t difficult to imagine CBP eventually having specialized models that examine classifications, valuations, country-of-origin claims, and other entry information and identify the small percentage of transactions that deserve closer human review.
That doesn’t require eliminating import specialists. It could make each specialist much more effective by directing attention toward the entries that are most likely to contain a problem.
CBP is already experimenting with AI and machine learning in other areas, so this would not be starting from scratch.
What if CBP offered the AI directly?
There is another possibility I find particularly interesting.
Suppose CBP eventually develops or commissions a customs model that performs significantly better than the general-purpose AI systems available to the public.
Why not let importers use it?
Imagine an official CBP classification service.
An importer describes a product, uploads specifications or photographs, and answers whatever questions the system needs. The AI works through the tariff schedule, chapter notes, CBP rulings, and other relevant material and produces a recommended classification.
You could even imagine some type of safe harbor. If the importer supplied accurate information and relied on CBP’s system, CBP would generally honor the result.
For a small importer, that could remove a substantial amount of uncertainty and expense from routine classifications.
Brokers and customs attorneys would still have plenty to do. Difficult merchandise, ambiguous rules, valuation questions, country of origin, antidumping and countervailing duties, admissibility, protests, and enforcement issues are not going away.
But a lot of routine classification work could become much easier.
CBP could even use the same underlying system internally when reviewing entries, giving the importer and the government a common starting point instead of having two completely separate processes trying to reach the same answer.
There would obviously need to be appeals and human review for difficult cases. I wouldn’t want an AI system to become judge, jury, and customs officer.
Still, if a system like this removed even a meaningful portion of the friction involved in importing goods, it could make international trade more efficient for U.S. companies.
There is also a competitive angle. If another country makes importing substantially easier through better automation while the United States retains a much more labor-intensive process, companies will notice.
The law will move more slowly than the technology
The biggest obstacle to most of these ideas is probably not technological.
It is legal and political.
The customs broker regulatory system has existed in one form or another for generations. Changing it involves statutes, regulations, administrative procedure, industry interests, liability, and politics.
None of those things move particularly quickly.
AI does.
That’s what makes H350722 interesting to me.
The ruling makes sense within today’s legal framework, and probably within today’s technology. We still want an identifiable human professional making important regulatory decisions.
But I don’t think we can assume the human will always be the most capable participant in that process.
If AI systems continue improving at anything close to their current pace, customs brokerage is going to become increasingly automated whether the regulatory framework changes or not. Brokers themselves will use these tools because they make them faster and more productive.
The larger question is what we do when the technology eventually gets good enough that “human in the loop” no longer necessarily means “expert in the loop.”
At that point, we may have to decide whether the purpose of the human is to improve the answer, provide accountability, or simply satisfy a rule written for a very different technological world.