The conversation about automated journal entry screening in audit methodology often gets framed as a replacement argument — will the software replace the auditor? That framing is wrong and, more importantly, it's a distraction from the actual question: what does the software do better than the auditor, what does the auditor do better than the software, and how do you design a testing program that gets the most out of both?
This is an attempt at that honest accounting. It applies specifically to the task of journal entry testing under AU-C Section 240 and AS 2110 requirements. It does not address broader audit technology questions about AI and the profession.
What Manual Testing Actually Looks Like
Manual JE testing, at most audit firms today, follows a predictable workflow: export the general ledger from the client's ERP, apply filter criteria in Excel to identify "unusual" entries, select a sample from the filtered population, obtain and review supporting documentation for each selected item, and conclude. The filters typically include: entries over materiality, entries posted after business hours, entries without textual descriptions, and entries by users not in the normal posting population.
This approach has genuine strengths. An experienced auditor reviewing an individual entry can evaluate not just the numerical characteristics but the narrative context — whether the description makes sense given the account combination, whether the timing aligns with known business events, whether the preparer's role in the organization is consistent with the amount being posted. That contextual judgment is not easily systematized.
The weakness is coverage. A population of 20,000 journal entries filtered to 3,000 items above materiality, sampled to 60 items for detailed testing, represents 0.3% coverage of the total population. The 99.7% that wasn't touched could contain misstatements that a more complete scan would have flagged.
What Automated Screening Actually Does
Automated screening tools — including AuditPulsar — apply statistical and rule-based tests to the complete journal entry population and produce a ranked list of items that deviate from expected patterns. The deviation tests typically include: Benford's Law analysis on amount fields, after-hours posting flags, round-number detection, unusual account combination detection, preparer anomaly scoring (based on the preparer's historical posting patterns), and journal entries that bypass normal workflow steps.
AuditPulsar adds supervised ML scoring to this stack. The model was trained on 14.7 million historical journal entries labeled by audit outcome — whether the entry was ultimately identified as a finding, required follow-up, or was cleared without issue. That training allows the score to reflect not just statistical deviation from the current population but deviation from what entries that turned out to be problematic typically looked like.
The strength of automated screening is coverage. Every entry in the population gets evaluated. There is no sampling decision — the question is not "which 60 entries do we review" but "which 60 entries in this population of 20,000 look most like entries that mattered in prior engagements."
The weakness is context. The scoring model doesn't know that a large unusual credit to the revenue account in December is actually a standard holiday bonus reversal that this client has processed the same way for eleven years. That context lives in the auditor's head or in the prior-year workpapers. Without it, the scoring will flag the entry, and the auditor will need to clear it.
The Actual Trade-Off Table
Breaking this down by what each approach is better at:
Population coverage: Automated wins. Manual testing of a 20,000-entry population is 0.3% coverage at realistic staffing levels. Automated screening covers 100%.
Contextual judgment: Manual wins. An auditor who has worked this engagement for three years knows which account combinations are unusual for this client and which are standard practice. The model knows statistical deviation; the auditor knows meaning.
Detection of novel fraud patterns: Roughly even, with different failure modes. Manual testing misses novel patterns because sampling creates coverage gaps. Automated scoring can miss novel patterns that don't resemble historical training data. The combination outperforms either alone.
Documentation: Automated wins on completeness. A system-generated workpaper that documents the full population, the scoring criteria, and the items reviewed is more complete than a set of Excel files with filtered populations that may or may not be reproducible.
Scalability: Automated wins strongly. Doubling the journal entry population doesn't double the effort for automated screening. For manual testing, it typically does.
The Documentation Distinction That Matters Most
When a firm moves from manual to automated screening, the most significant change isn't in what gets found — it's in how the testing procedure is documented and described in the workpapers.
Manual JE testing documentation typically looks like this: population definition, filter criteria applied, items selected, supporting documentation for each item, conclusion. The implicit assertion is that the filter criteria were sufficient to identify the high-risk items and that the sample was representative of the filtered population.
Automated screening documentation looks like this: population definition, tool used and version, parameters configured, scoring criteria and their basis, items flagged above threshold, auditor's disposition of each flagged item, conclusion about the population. The implicit assertion is that the scoring criteria are valid for this population type and that the auditor's review of flagged items was sufficient to conclude on the population.
The second documentation approach requires the auditor to make and document an explicit judgment about the reliability of the automated tool for this specific use. PCAOB AS 2110 and AU-C Section 240 both require this when using IT-assisted procedures. The judgment includes: what the tool does, what it was validated against, what its known limitations are, and why those limitations don't undermine the conclusion for this particular engagement.
Firms that switch to automated screening without updating their documentation approach end up with workpapers that describe a manual process but cite automated outputs. That's a documentation inconsistency that PCAOB inspectors notice.
Where the Combination Works Best
The testing programs that produce the best results combine automated screening for population-level coverage with manual detailed review for the flagged items. The specific workflow that works well in practice:
Step one: run the automated screen on the complete population to produce a scored and ranked list. Step two: auditor reviews the top-scored items (typically 20 to 50 items depending on population size) using the same contextual judgment that manual testing applies but targeted at the entries statistically most likely to matter. Step three: auditor clears items that have acceptable explanations, escalates items that warrant further investigation. Step four: for any items escalated, obtain supporting documentation and evaluate. Step five: document the complete process, including the scoring criteria and the auditor's basis for concluding that the screening was sufficient for this population.
This workflow typically takes 30 to 50% less time than pure manual testing on large populations while covering the full population rather than a sample. The time savings come almost entirely from elimination of the manual filtering and population reduction steps, which are now handled by the scoring algorithm.
What Not to Do
Two failure modes seen in practice when firms adopt automated screening tools:
First: using the tool's output to replace auditor judgment rather than inform it. If the platform says an entry scores 45 out of 100 and the auditor concludes it's low risk without reviewing the actual entry details, that's a documentation problem and potentially an audit quality problem. The score is a prioritization tool. The auditor's review of the item is the actual substantive procedure.
Second: running automated screening in addition to existing manual sampling without reducing the manual sampling proportionally. This is common in the first year or two of tool adoption. It results in higher total testing effort with limited incremental benefit. The value of automated screening comes partly from replacing the population reduction step, not from adding another layer on top of it.
The Honest Answer on Replacement
Automated screening reduces the time an auditor spends on mechanical population filtering and increases the time they spend on items that are actually interesting. It does not — at current capability levels — replace the judgment component of journal entry testing. It changes what the judgment is applied to: instead of evaluating which 60 entries to review from a filtered population, the auditor evaluates the 30 entries the algorithm identified as highest risk in the complete population.
That's a better use of auditor time, not a replacement of it. Whether it's also more effective depends on whether the algorithm's risk model is a better predictor of actual findings than the auditor's manual filter criteria. Based on the engagements where we've been able to compare outcomes, the algorithm's top-ranked items have a higher conversion rate to actual findings than traditional sampling approaches. The auditor's contextual review of those items remains essential to that process.