When we were scoping Tradevynt's first feature set, we made a deliberate decision: talk to working customs brokers before writing a line of code. Not "industry thought leaders." Not conference panel participants. The people who process entries every day.
We spoke with licensed brokers at a handful of mid-size freight forwarders and several independent brokerage operations — some of whom handle 80 to 150 entries a week on their own. The conversations were long and specific. What emerged was a picture quite different from the usual "automate customs filing" pitch you hear in trade tech.
The Work They Want Off Their Plate
The answer wasn't "automate my classification decisions" — it was consistently more upstream than that. The work brokers described as their biggest time sink was the data intake phase: opening a shipment packet, reading through the commercial invoice, packing list, and bill of lading, and manually keying every shipper name, consignee EIN, container number, piece count, and weight figure into their filing system before they can even start the entry.
One broker described her Monday morning as "two hours of typing before I can think." She processes about 25 entries on a heavy day. The cognitive overhead isn't the classification work, which she finds genuinely interesting — it's the transcription of information that already exists in a document but won't get into her system without her fingers on the keyboard.
This is what document extraction addresses directly. The value isn't that the machine makes classification decisions; it's that the machine reads the document so the broker doesn't have to re-enter what's already there. The classification review still happens — but it happens without the preceding hour of data entry.
What They Refuse to Automate
Classification decisions at the subheading level, in any workflow we heard about, are non-negotiable human territory. The reasons aren't purely regulatory conservatism — they're practical. When CBP sends a CF-28 or CF-29 requesting additional information or proposing a rate advance, the licensed broker is legally on record for the classification. "The software did it" is not a defense in a CF-29 response.
Beyond legal exposure, experienced brokers described classification as the core of their professional value. An 8-year broker knows that a particular electronics importer consistently misdescribes their assemblies in ways that put them in the wrong subheading, and she's developed a mental check for exactly that class of invoice. That pattern recognition is the job — and it's not something she wants to delegate.
The same thinking applies to value reconciliation. When the declared value on a commercial invoice is inconsistent with the transaction value requirements under 19 CFR Part 152, a broker needs to work through the question of whether the importer has transfer pricing arrangements, assists, or post-importation payments that should affect dutiable value. Automating the value field entry is fine. Automating the value determination is not.
The Middle Ground: Assisted Review
Where brokers showed genuine enthusiasm was for a category we'd describe as assisted review — systems that do the data intake work and surface issues, but leave the judgment call with the human.
Specifically, several brokers described wanting a system that would:
- Extract all data fields from the shipment documents and pre-populate the entry form
- Flag cross-document discrepancies automatically (piece count on the B/L doesn't match packing list, shipper on commercial invoice doesn't match shipper on the AMS filing)
- Pull the current tariff rate for the extracted HS code and show it next to the classification, so the broker can confirm or override
- Check whether the commodity triggers any PGA requirements without making the broker remember which chapter numbers map to which agencies
Notice what's not on that list: the system making a classification recommendation. Most brokers we talked to were skeptical of systems that surface a suggested HS code — not because the suggestions are always wrong, but because there's a risk the broker starts rubber-stamping suggestions rather than exercising independent judgment. The legal standard for reasonable care under 19 USC 1484 requires the importer to exercise that care, and a broker facilitating rubber-stamp classification is not serving their client well.
We're not saying AI-assisted HS suggestions are categorically bad — they can work well as a starting point when confidence is clearly labeled. But the brokers who expressed caution here had a point. The tool should serve the judgment, not replace it.
Volume vs. Complexity: Different Brokers, Different Needs
There was a clear split in what brokers wanted based on their volume and customer mix.
Brokers at forwarders handling high-volume, repetitive commodity types — garments from a single origin country, standard industrial components on regular purchase orders — were most interested in templating and pattern recognition. The invoices for their core customers are structurally identical; once you've mapped the fields for one, the next two hundred are the same extraction problem. For these brokers, throughput automation has obvious value.
Independent brokers with diversified portfolios were more interested in the discrepancy-flagging and PGA trigger functions. Their work is inherently varied — a week might include a textile shipment, a food ingredient, a piece of industrial machinery, and a medical device. For them, the machine's value is catching the things that are easy to miss when you're context-switching constantly.
Both groups were consistent about one thing: they don't want to "train" a system or curate datasets. They want something that works on their existing documents without a 90-day onboarding project. The incumbents in the space often require significant setup investment; that's a real barrier for a brokerage with 4 or 5 staff members.
What This Shaped in Tradevynt
These conversations are what drove us toward the confidence-flagged extraction model rather than a classification recommendation engine. The output we show brokers is: here are the fields we extracted, here is our confidence in each field, here are the cross-document discrepancies we found, and here is the current tariff treatment for the extracted commodity code. The broker reviews, adjusts, and files.
That design also means we're not trying to replace broker judgment in the workflow — we're trying to get the broker to the judgment step faster. The two hours of Monday morning typing becomes fifteen minutes of review. The part she finds interesting — actually working through a classification question — remains in her hands.
Building tools that practitioners actually adopt requires starting from their work as it is, not from an idealized version of what you'd like it to become. Most of what we built in the first year came directly from those early conversations, and the things brokers told us they didn't want automated — we haven't built those either.