Pull entry level data for the period with the fields the audit needs: entry number and date, importer line detail, classification, origin, entered value, and duty by line, separating the base line from each Chapter 99 line. Most ACE reports and broker data feeds can produce this; the broker's monthly summary, with duties blended into one figure per shipment, cannot. If your data arrives blended, fixing the data feed is the audit's first deliverable.
Add two reference columns of your own: the stack you expected for that origin and heading on that entry date, built from your classification worksheets and stack sheets, and the difference in dollars. Expected stack is the heart of the exercise. An audit that only lists what was paid can find arithmetic errors; comparing against an expectation finds legal ones.
A small team cannot re examine every line, and does not need to. In most import books a minority of lines carry the great majority of Chapter 99 duty, so sort descending by additional duty paid and work down until the dollars stop justifying the hour. The common finds at the top: a product still paying a rate from a list it moved off, an exemption that took effect and was never picked up, a forced labour layer applied to an origin outside its scope or missed on one inside it.
Also read the zeros. Lines that paid no Chapter 99 duty in a family where every sibling line paid some are as interesting as the large payers, because exemptions are rarer than errors. A zero with a documented exemption behind it is a good zero. A zero because a template dropped the line is a liability accruing interest quietly, and it will be found eventually by someone whose job is finding it.
Rates attach by entry date, and stacks changed materially during 2026: layers expiring and taking effect on the same July morning is only the loudest example. Every comparison in the audit must use the stack in force on that entry's date, not today's stack and not the stack from when the product was first imported. Build the expectation by date band, and give the boundary weeks, around any known change, a line by line read regardless of dollar size.
Keep the audit workbook. Dated, with the reference stacks used and the dispositions of every flagged line, it is contemporaneous evidence of reasonable care, and it is the starting point that makes next quarter's audit a two hour job instead of a two day one. Reasonable care, documented as a habit, is the strongest general position an importer can hold.
Sort findings by liquidation status the moment they are found. Unliquidated entries with overpaid duty go to the broker for post summary correction. Liquidated entries inside 180 days go into the protest queue with dates attached. Underpayments get customs counsel and a prior disclosure conversation before anyone files anything, because the order of operations on a discovered underpayment changes its consequences materially.
Close the loop into the templates. Every finding caused by a stale broker instruction or an outdated internal stack sheet gets its source fixed and dated, or the same finding will be waiting in next quarter's export wearing a new entry number. An audit that does not change the templates is entertainment.
Keep reading: Who owns the duty when the marketplace is the importer: reading your channel's fine print, Duty paid stock that moves between warehouses: the cost follows the unit.
Entry level detail from ACE or an unblended broker feed: entry number and date, classification, origin, entered value, and duty separated by line, base Chapters 1 to 97 duty apart from each Chapter 99 line. Blended per shipment duty totals cannot support the audit, because the whole method is comparing layer by layer against the stack that should have applied on that entry date.
Start with the most recent quarter and work by dollars at stake, largest Chapter 99 lines first. Then read the boundary periods around any known rate change line by line regardless of size, because date errors concentrate there. Entries already liquidated beyond the protest window are mostly beyond remedy, so recent and unliquidated entries deserve the hours first.
Stale templates: a product still rated for a list it left, an exemption never picked up, the new forced labour layer missed or misapplied around its July 2026 start, and zero Chapter 99 lines in families where sibling lines all carry the layer. Underpayments matter as much as overpayments; route those to counsel for a prior disclosure discussion before filing anything.
The dated workbook, with expected stacks, comparisons, and dispositions, is contemporaneous evidence of reasonable care, which is the standard importers are judged against. It shows a system that checks itself. It also means the errors CBP might find were likely already found, corrected, or disclosed by you, which changes the character of that conversation entirely.