Developing Explainable Artificial-Intelligence Audit Tools to Identify Revenue-Recognition, Payroll, Expense-Management, and Financial-Reporting Risks While Supporting Professional Audit Judgment

Full text

Text extracted from the article PDF for reading and search. Tables, figures and equations are in the PDF, which remains the authoritative version.

1. Introduction

Auditing has traditionally relied on sampling: an auditor ex-amines a few dozen payroll records or card transactions and ex-trapolates. Today, auditors can obtain entire populations of pay-roll, expense, and financial-reporting data, and audit standards and professional guidance encourage data analytics that exam-ine every record [1, 2]. Payroll and expense-reimbursement schemes remain among the most common forms of occupa-tional fraud [3], and improper revenue recognition is a frequent subject of SEC enforcement [4].

Machine learning can rank millions of records for review, but its outputs create a new problem for the audit profession. An auditor who acts on a flag must document why the item was selected and what evidence was obtained [5, 6]. A black-box anomaly score does not provide that reason. It also invites two failure modes documented in research on human–automation interaction: over-reliance, in which the auditor accepts the tool’s output without scrutiny, and under-reliance, in which a correct flag is dismissed because it cannot be explained [7, 8]. Post-hoc explanation methods help, but they explain an opaque model rather than making the model itself transparent [9, 10].

This paper takes a different approach. We develop XAT (Ex-plainable Audit Toolkit), in which every component is inher-ently interpretable: transparent audit rules that encode known fraud schemes, peer-group baselines that compare each record with similar records (for example, the same job title in the same agency), and, where labels exist, a sparse linear model whose coefficients can be read directly. Every flagged item comes with an explanation card: which rules fired, how far the item devi-ates from its peers, and what follow-up procedure is suggested. The auditor, not the tool, decides what the evidence means.

We evaluate XAT on the four areas named in its purpose: • Payroll: 544,338 New York City payroll records for fiscal year 2024. • Expense management: 109,011 Oklahoma state purchase-card transactions posted in April–June 2024, the last quarter of the state’s fiscal year.

Revenue Recognition and Financial Reporting: 35,166 U.S. public-company firm-years (2003–2008) labeled with SEC Accounting and Auditing Enforcement Releases (AAERs) [11].

Because payroll and card data carry no fraud labels, we follow the established practice of seeding realistic anomalies into real data and measuring how many the tool surfaces within a fixed review budget [12]. For revenue recognition, we use real en-forcement labels.

The contributions are: (i) an explainable toolkit design that combines rules, peer baselines, and glass-box models, with explanation cards for audit documentation; (ii) a reproducible evaluation on three real U.S. public datasets covering payroll, expenses, and revenue; (iii) evidence that the transparent com-bination outperforms both rules alone and a black-box isolation forest at realistic review budgets, and that a glass-box revenue model gives up little accuracy relative to a black-box model; and (iv) real-data findings, including the scale of rule exceptions an auditor would face, a Benford’s-law test, and a check for fiscal year-end spending.

2. Related Work

Audit data analytics. Appelbaum et al. [2] set out research needs for analytics in the modern audit, and the AICPA’s guide describes how analytics fit into risk assessment and substantive procedures [1]. Nigrini and Mittermaier [13] introduced Ben-ford’s law as an analytical procedure, and Nigrini [14] codified conformity thresholds that remain standard in forensic account-ing.

Anomaly detection in accounting data. Schreyer et al. [12] used deep autoencoders to detect anomalous journal entries and evaluated them partly with seeded anomalies. Isolation forests [15] are widely used as general-purpose detectors. These methods are powerful but opaque, and their scores do not trans-late directly into audit evidence. Fraud prediction and explainability. Beneish [16] and De-chow et al. [4] built interpretable ratio models of earnings ma-nipulation, including revenue-related signals such as receiv-ables growth. Bao et al. [11] showed that ensemble learning on raw financial data improves detection. Rudin [9] argues that in high-stakes settings, inherently interpretable models should be preferred when they perform comparably. Commerford et al. [8] found that auditors’ reliance on AI-based evidence differs from their reliance on human specialists, underscoring the need for tools whose reasoning auditors can evaluate. XAT builds on these strands by making interpretability a design constraint rather than an add-on.

3. The Xat Toolkit

Fig. 1 shows the design. A shared engine applies two kinds of transparent tests to each module’s data, and a human-in-the-loop cycle turns flags into audit evidence.

3.1. Transparent Rules

Rules encode well-known schemes, each with a stated thresh-old that an audit team can adjust (Table I). For payroll they target excessive overtime, pay above salary, large unsupported payments, and the same person employed full-time in two agen-cies. For purchase cards they target split purchases (several same-day charges from one cardholder at one merchant that together exceed a single-transaction limit), possible duplicate payments, weekend or holiday use, and merchant categories Fig 1: XAT design. A shared explainable engine (rules plus peer baselines) serves three modules. Each flag produces an explanation card that feeds a human-in-the-loop cycle: the auditor judges the evidence, performs follow-up procedures, and tunes rules and thresholds over time.

that are rare for the agency. We use $5,000 as an illustrative single-transaction limit.

3.2. Peer-Group Baselines

Rules miss schemes that stay below thresholds. XAT therefore compares each record with its peers using robust z-scores. For feature j of record i in peer group g, (1)

where med and MAD are the peer-group median and median absolute deviation, which resist distortion by the very outliers being sought [17]. Only upward deviations, which correspond to overpayment or overspending, are scored. Peer groups are as specific as possible: agency × job title × pay basis for pay-roll, and agency × merchant category for cards, falling back to broader groups when a group has fewer than 20 members. Payroll features are overtime hours and pay, overtime pay per hour, other pay, regular gross pay, and the ratio of gross pay to base salary. Card features are the amount, the amount relative to the cardholder’s usual spending, and the rarity of the merchant category in the agency.

3.3. Combined Score and Explanation Card

The review score is si = log(1 +max j zi j) + w· ri, where ri is the number of rules triggered and w = 3, so that a rule hit counts as much as a roughly twentyfold larger peer deviation. The score is a ranking device only. The explanation card lists the triggered rules, the top deviating features with their values, peer medi-ans, and peer-group sizes, and a suggested follow-up procedure (Fig. 4).

Table 1: Transparent Rules and Real-Data Exceptions Rule Description Exceptions Payroll rule (NYC FY2024, 534,338 real records) R1 Overtime > 1,500 hours in the base year 443 R2 Regular gross pay > 120% of base salary (annual basis) 17,905 R3 Other pay > regular pay and > $20,000 1,549 R4 Same name and start date, full-time in 2+ agencies 285 Any payroll rule 20,146 (3.7%) Purchase-card rule (Oklahoma, 107,913 real transactions) E1 Possible split purchase (711 groups, $3.53M) 3,001 E2 Possible duplicate (≤3 days apart, $2.93M) 7,367 E3 Weekend or state holiday 13,005 E4 Merchant category <0.1% of agency’s transactions 2,816 Any card rule 23,888 (22.1%)

3.4. Glass-Box Revenue Model

For revenue recognition, where enforcement labels exist, XAT uses an ℓ1-regularized (sparse) logistic regression [18] on eight standardized revenue red flags: the Beneish days-sales-in-receivables index (DSRI), sales growth index, and gross-margin index [16]; receivables and inventory relative to sales; and the change in receivables, change in cash sales, and share of soft assets [4]. Each coefficient states directly how a red flag moves the risk. Following [11], each test year t (2003– 2008) is scored by a model trained on 1991 to t − 2, with serial frauds that span both periods recoded in training.

4. Data and Evaluation

4.1. Data

Payroll. The NYC Citywide Payroll dataset, published by the Office of Payroll Administration [19], reports each employee’s agency, title, pay basis, base salary, regular hours and gross pay, overtime hours and pay, and other pay. We keep FY2024 records with positive regular or overtime pay: 543,338 records across 159 agencies and 1,543 titles. Purchase cards. Oklahoma’s Office of Management and Enterprise Services publishes all state and higher-education purchase-card transactions [20]. We use the files posted in April, May, and June 2024: 111,988 rows, of which 3,875 re-funds are excluded, leaving 107,913 transactions worth $49.5 million by 4,925 cardholders in 115 agencies. Revenue and reporting. We use the public firm-year dataset of Bao et al. [11], with AAER fraud labels; test years 2003– 2008 contain 35,166 firm-years, 261 of them fraudulent. Although both transaction datasets are public and name in-dividuals, this study reports only aggregates and anonymized examples.

4.2. Seeded Anomalies

payments (a charge repeated within 0–2 days), personal-use purchases ($100–$1,500 on a Saturday in a merchant category the agency does not use), and inflated amounts (5– 15 times the original). Altered records replaced their originals; cloned and added records were appended. In total, 1,000 payroll records and 1,098 card transactions were seeded.

4.3. Baselines and Metrics

We compare XAT with rules alone, peer z-scores alone, a global z-score that ignores peer groups, and an isolation forest [15, 21]. The main metric is the share of seeded anomalies found in the top 1% and 2% of the review list; random review would find 1% and 2%. We also report AUC and average precision (AP). For revenue, we report mean out-of-sample AUC and the share of frauds captured in the riskiest 5% of firms.

5. Results

5.1. Payroll and Expense Modules

Table II summarizes detection of seeded anomalies. Within a 1% review budget, XAT surfaces 31.1% of seeded payroll anomalies and 26.0% of seeded card anomalies, more than any baseline and about three to four times the share found by the isolation forest. XAT also has the highest AUC and AP in both modules.

Fig. 2 breaks detection down by scheme and shows why the combination matters. Rules and peer baselines catch different schemes. In payroll, the peer z-score is the best detector of padded overtime (47.6%), because padded hours rarely cross the 1,500-hour rule but stand out sharply against colleagues with the same title. Rules are best for ghost employees (30.0%), because a cloned employee looks normal relative to peers. XAT keeps most of the strength of each. Two schemes remain hard: unsupported other pay (11.6%), because real other-pay amounts vary widely, and inflated card amounts, which no method de-tected well (at most 11.0%), because legitimate card amounts span several orders of magnitude even within a merchant cate-gory.

The rule weight w has a modest effect. Varying it from 1 to 5 changes XAT’s top-1% payroll detection between 24.2% and 32.2%, and card detection between 25.6% and 26.9%.

5.2. What Auditors Would Face in the Real Data

Table I reports how many real records trigger each rule. These are exceptions for follow-up, not findings of fraud. The volumes are instructive: 22.1% of card transactions trigger at least one Fig 2: Share of each seeded scheme surfaced within a 1% review budget. Each row compares five methods; the gray bar spans their range. XAT (navy diamond) combines the scheme-specific strengths of rules and peer baselines. Fig 3: First-digit test of 101,184 Oklahoma purchase-card amounts of $10 or more against Benford’s law.

card rule, mostly weekend use and possible duplicates, far more than any audit team can review. This is the practical case for ranking: XAT orders exceptions by combined evidence, so the team reviews the most unusual first. Two population-level tests found no red flags. The first-digit distribution of card amounts conforms closely to Ben-ford’s law (Fig. 3; mean absolute deviation 0.0076, within Ni-grini’s acceptable-conformity range [14]). Average weekday card spending in the last two weeks of the fiscal year ($635,000) was slightly below the April–May average ($692,000), so there was no year-end spending surge.

5.3. Explanation Cards

Fig. 4 shows two explanation cards for seeded anomalies, illus-trating what an auditor would see. The payroll case triggered no rule, but its overtime pay was extreme relative to its 24 peers, which placed it 102nd of 544,338 records. The card case trig-gered two rules and was ranked 93rd of 109,011 transactions. In both cases the card states the evidence, so the auditor can de-cide whether the explanation is plausible and which procedure to perform.

5.4. Revenue-Recognition Module

Table III reports the revenue module. The glass-box sparse lo-gistic model achieves a mean AUC of 0.647 and outperforms the Beneish-style revenue index (0.553) in every one of the six test years (Fig. 5b). DSRI alone is no better than chance (0.505). The black-box gradient-boosting model is moderately better (AUC 0.671; 18.9% versus 11.2% of frauds captured in the riskiest 5% of firms). That gap is the price of transparency on this task: auditors can choose the glass-box model when the reason for each score must be documented, or use the black-box model as a secondary screen.

Because the model is linear, its reasoning can be read directly from Fig. 5a, and it is stable across years. The share of soft as-sets (assets other than cash and property, plant, and equipment) is the strongest red flag, followed by the change in receivables and the change in cash sales, consistent with [4]. Once these are included, DSRI receives a small negative weight, a reminder that individual ratios should not be read in isolation. The sales growth index is nearly zero and was dropped entirely in one year.

6. Discussion

6.1. Supporting, Not Replacing, Judgment

XAT is designed so that the auditor remains the decision-maker. Each exception carries evidence the auditor can ver-ify against source documents, which supports documentation requirements [5, 6] and reduces both blind reliance and unwar-ranted dismissal of flags [7, 8]. The results show that this trans-parency need not cost detection power: in payroll and card data, the transparent combination outperformed a black-box isolation Fig 4: Explanation cards for two seeded anomalies, as presented to an auditor. Each card states the rules triggered, the deviation from peers, and a suggested follow-up procedure; the auditor judges the evidence. Personal identifiers are omitted.

Fig 5: Revenue module. (a) Standardized coefficients of the glass-box model; dots show the mean and lines the range across the six yearly models, so an auditor can see both the direction and the stability of each red flag. (b) Out-of-sample AUC by test year.

forest by a wide margin. In the revenue module, transparency carried a moderate cost, which auditors can weigh explicitly.

6.2. Practical Implications

For internal auditors and inspectors general, the toolkit can run over complete payroll and card populations each period at negligible cost, using thresholds the audit team sets and doc-uments. Rule exception volumes in real data, 3.7% of pay-roll records and 22.1% of card transactions, show that rules alone overwhelm review capacity; peer-based ranking is what makes population-level testing feasible. The same design ap-plies to private-sector payroll and expense systems, which con-tain equivalent fields.

6.3. Limitations

Seeded anomalies are a controlled test, not a substitute for con-firmed fraud, and some schemes were designed with the same red flags the rules encode; detection rates for those schemes therefore measure how well XAT surfaces known patterns amid real-world noise, not its ability to discover new schemes. Real exceptions were not investigated, so their fraud rate is unknown. Public payroll data report annual totals rather than pay-period detail, which limits tests such as duplicate payments within a period. Rule thresholds are illustrative and should be set from each organization’s policies. Finally, the revenue module cov-ers 2003–2008, and its performance on recent filings should be verified.

7. Conclusion

This paper developed XAT, an explainable AI audit toolkit that combines transparent rules, peer-group baselines, and glass-box models, and evaluated it on real U.S. payroll, purchase-card, and financial-reporting data. Within a 1% review budget, XAT surfaced 31.1% of seeded payroll anomalies and 26.0% of seeded card anomalies, outperforming rules alone and a black-box isolation forest, and its glass-box revenue model came close to a black-box model on SEC-identified fraud. By attaching verifiable evidence to every flag, the toolkit helps auditors use complete-population data while keeping professional judgment at the center of the audit. Future work will test the toolkit prospectively with audit teams, using investigated exceptions as labels, and extend it to pay-period and journal-entry data.