How a Three-Person Customs Brokerage Reclaimed 78.6% of Their Reconciliation Hours

By the time Arnoldo reached out to GingerControl in early 2026, his three-person customs brokerage, Quantum Customs Brokers, had become a victim of its own success. New client growth had outpaced operational capacity, and the team was drowning in inboxes. Each broker was processing between 50 and 100 emails a day, and every email carried a different shipment, a different document set, and a different question.

The work itself was familiar to anyone in the trade, pull the Commercial Invoice, cross-check it against the Packing List, reconcile both against the Bill of Lading, then key the numbers into ABI. Familiar, but unforgiving. A single mismatch between weights, quantities, or consignees, and the entry would bounce. A misfiled HTS code, and the client would call.

Arnoldo's team had already invested in an AI-assisted ABI platform. On paper, the tool should have absorbed most of the manual load. In practice, the data lived in silos. Identical clients shipping identical products would appear as fresh records each time. Brokers found themselves clicking between modules, copying and pasting between tabs, and rebuilding context that the system had captured only hours earlier.

We began with what we call a Workflow Audit. Over two weeks, a GingerControl consultant sat alongside each broker for full operating days, mapping every keystroke from inbox to ABI submission. We instrumented their workflow with lightweight timers and tagged every action by category, reading email, extracting numbers, comparing documents, re-keying data, chasing missing information, filing.

The numbers were sobering. Across the team, 42% of working hours were spent on what we classified as reconciliation labor, the act of moving the same data from one place to another and checking that it agreed with itself. Another 19% was spent on email triage. Only 27% of working time was spent on the work that customs brokers are actually licensed to do, classification judgement, compliance review, and exception handling.

Equally telling was where errors emerged. The team's CBP rejection rate was low, a credit to their experience, but the near-miss rate was high. When we reviewed three months of internal QA logs, we found that 1 in 6 entries had been caught and corrected internally before submission, and most of those corrections involved a number that appeared correctly in one document and incorrectly in another.

We presented Arnoldo with two intervention points. The first was the inbox itself. The second was the layer beneath it, the customer and shipment record that every email and every document ultimately referenced.

Rather than build two disconnected tools, we proposed a single architecture, an email agent that reads, classifies, and acts on incoming correspondence, sitting on top of a unified data pipeline that every member of the team would share. The agent would not replace the broker's judgement. It would absorb the mechanical work, identify cases the broker still needed to look at, and surface the right context the moment a broker opened a thread.

The design review took four sessions over three weeks. We sketched intent classifications with Arnoldo's team, what kinds of emails do you actually receive, and what is the right action for each? We mapped document fields across CI, PL, and BOL to a canonical schema, so that "gross weight" meant the same thing whether it arrived from a freight forwarder in Shenzhen or a shipper in Hamburg. We agreed on the threshold for autonomy, when should the agent act on its own, and when should it stop and ask?

Development ran for six weeks. The Email Agent was built around three capabilities, all sitting on top of the new unified record.

The Unified Customer and Case System was the structural piece. Every client, every shipment, every document, and every email now references a single record. When a broker opens a case, they see the full timeline, prior shipments for the same client, recurring product lines with their historical classifications, every document attached, every email exchanged. The records that previously lived across modules now exist in one view, accessible with one click.

We ran the system in shadow mode for the first ten days, comparing the agent's proposed actions against what the brokers actually did. We tightened the autonomy thresholds based on the disagreements, then flipped to live mode.

By the end of the third month after deployment, the team's reconciliation labor had fallen from 42% to 9% of working hours, a 78.6% reduction in the share of the day spent moving numbers between documents. Email triage time dropped by 79%. The broker workday, once a chase of inboxes and tabs, became a series of decisions on cases the agent had already prepared.

Throughput moved accordingly. The brokerage processed 79% more entries per broker per week, with no change in headcount. Client communication latency, the time between a client sending a question and receiving a substantive reply, dropped from a median of 4 hours 12 minutes to 54 minutes, almost entirely because the agent surfaces resolved cases the moment the resolving document arrives.

What Arnoldo cared about most was the part that does not show up cleanly in a metric.

Three things, in retrospect, distinguished this engagement from a generic automation rollout.

1. 01

   The diagnosis preceded the design.

   We did not arrive with a product. We arrived with a stopwatch and a notebook, and we let the workflow tell us where the friction lived. The intervention that resulted, an email agent fused to a unified data layer, was not the intervention either party would have proposed at the outset.

2. 02

   The data architecture preceded the agent.

   An email agent built on top of fragmented records would have inherited the fragmentation. We resisted the temptation to ship the agent first and rebuild the data later. The order was the design decision.

3. 03

   The autonomy thresholds were negotiated, not assumed.

   Arnoldo's team did not want an agent that acted on everything, and they did not want an inbox assistant that asked about everything. The shadow period was where the boundary was drawn, case by case, with the brokers in the room.

Small customs brokerages sit in an unusual position. The volume of work that crosses their desks is enterprise-scale, but their operational headcount is artisanal. Off-the-shelf brokerage software is built for either end of that spectrum, the lone practitioner or the multinational. The middle, where Arnoldo lives, is where most of the trade actually happens, and where most of the manual reconciliation tax is paid.

If your team is spending more time moving numbers between documents than judging the numbers themselves, the intervention is not faster typing. It is connected data and a thin layer of agent intelligence that knows what a resolved case looks like.

That is the work GingerControl was built for.