How Much to Re-Check: A Risk-Based Sample Plan for AI-Drafted Entries

An entry list drafted by an AI tool can run to dozens of rows, each with an amount, a label and a page reference. A Reviewer with a full stack of files has to decide where the reading time goes. The Firm can choose which rows get a full read before following the list's order.

A risk-based sample plan for AI-drafted entries settles which rows get a full read before anyone reads one. The Firm sorts the rows into tiers, checks every row in the top tier, checks a stated sample in the rest, and widens to a full check when a sampled row fails. The tool drafts the rows. A person sets the plan, reads the rows, and decides whether the file can be signed.

Sampling here is the Firm's chosen method for spending review time. It isn't a statement of what any standard asks for, and the Firm can choose to read every row on any file it likes.

Rows, not files

Some Firms sort whole files before the season and decide which ones get an AI pass at all. That sorts files. This plan works one level down: it sorts the rows inside a single file that has already been drafted.

How one row gets read is a separate question. The Reviewer's readback sets the order and the send-back reasons for a single row, and this plan only decides which rows get that readback.

Sort the rows into tiers

The Firm picks the inputs and writes them down so Reviewers can use the same stated sorting rule. Four inputs fit an AI-drafted entry list.

Dollar size. Rows at or above a dollar line the Firm sets for the file, plus a band just below it.

Entry type. Types the Firm marks as carrying judgment, such as a row that splits one printed figure into two, or a row whose label was chosen rather than copied from the page.

First-year item. A row with no match on the prior file: a new account, a new payer, a new kind of income.

Prior send-backs. Rows from a document or source that drew a send-back on this client's file before.

Under one approach the Firm can choose, any row at or above the dollar line, or flagged as first-year or prior send-back, goes to the top tier. Among rows not assigned to the top tier, rows in the band below the line and judgment-type rows go to the middle tier. Everything else goes to the bottom tier.

The Firm decides who sorts. This piece assumes the Reviewer does it, not the Preparer who drafted and checked the rows, so the person choosing what to read isn't grading their own work. Each row keeps its source page beside it in the plan, so a sampled row can be opened against its page without a search.

Write the sample plan before reading

The Reviewer writes the sample plan on the file before reading a row. The Firm can record the plan first to distinguish its initial choices from later changes.

The plan states five things: the tier rule used, how many rows sit in each tier, how many rows will be read in each tier, how the sampled rows are picked, and the widening rule.

How many to read in each tier is the Firm's call. This piece doesn't propose a number, and the Doe counts below are illustrative only. The Firm can set a higher share for the middle tier than the bottom, and can raise both for a new client or a new kind of document.

How the rows are picked matters as much as how many. The Firm can pick by a fixed method stated in advance, such as every fourth row by row number from a starting row chosen before any amount is read. The Firm can use that method to select sampled rows while leaving the Reviewer responsible for judgment and sign-off.

When a sampled row fails

A sampled row fails when the readback sends it back to the Preparer. Under the Firm's rule, one failed row in a tier widens that whole tier to a full check. A widened tier isn't sampled again; every row in it gets read. The Firm can also set a second line: two failures anywhere on the file widen every tier.

The Firm can adopt a rule that the plan may widen but cannot shrink once reading starts, and that a failed row stays in the results rather than being replaced by a fresh sample. Replacing a failed row does not resolve the failure the readback identified.

The Preparer corrects a failed row and notes what changed. The row doesn't count as cleared until the Reviewer reads it again against its page.

A row can also be pulled into a full check for another reason. If a document's rows didn't tie out to its printed totals, the Firm can treat every row from that document as top tier.

Record the plan and the results

The Firm can keep one results line per tier: rows in the tier, rows read, rows failed, whether the tier widened, who read the rows, and the date. The results line records who checked which rows by row number, alongside the tier counts.

The line also shows the rows not read, by count. A clean sample says the rows that were read cleared. It doesn't prove the unread rows are right, and the record says so plainly rather than leaving a later reader to assume it.

The sample plan and its results stay on the file. The Reviewer signs off only when every row read has cleared and every widened tier has been read in full.

A Doe example

The counts below are made up for illustration.

Pat Doe, a Reviewer at Doe & Co. CPA, opens the Doe Holdings file. Jane Doe prepared it from an AI-drafted entry list of 40 rows.

Pat sorts first. The top tier gets 9 rows: 5 at or above the Firm's dollar line, 2 first-year accounts, and 2 from a statement that drew a send-back on last season's file. The middle tier gets 16 rows and the bottom tier 15.

Pat writes the plan: all 9 top-tier rows, every fourth middle-tier row for 4 of 16, and every fifth bottom-tier row for 3 of 15. That's 16 rows to read out of 40.

All 9 top-tier rows clear. Pat completes the four-row middle-tier sample; the readback sends the third sampled row back to Jane. Under the Firm's rule, Pat widens the middle tier to all 16 rows. Reading the remaining 12 rows produces one more send-back. Jane corrects both rows and notes what changed. Pat reads both again against their pages, and they clear. The 3 sampled bottom-tier rows clear.

The results lines show 28 rows read out of 40, 2 failures, the middle tier widened, and 12 bottom-tier rows not read. Pat signs off.

Where TracePrep fits

TracePrep's one shipped workflow today is the Reasonable Compensation Study for an S-corporation shareholder-employee. It's a different job from this sample plan, and it keeps the same kinds of records:

We've built a TracePrep plugin that runs these steps in your Firm's workflow.

Want the TracePrep plugin in your Firm? We'll tell you when access opens.

Join the waitlist

This is general information, not tax or legal advice. Your licensed professional and your Firm's Reviewer are responsible for every return and every conclusion. Keep client data inside the systems your Firm has approved. As general guidance only, not advice: if you use an AI tool with client data, firms generally look for a business or commercial plan whose terms cover how that data is handled; read the vendor's current terms before you do.