Machine learning substantially improves IRS audit efficiency and revenue forecasting—cutting no-change audit rates by about 15–20 percentage points and forecasting error by up to 30%—but ROI-optimizing classifiers threaten to raise audit pressure on low-income filers unless equity-by-design safeguards, explainability, federated infrastructure, and statutory modernization are enforced.

Main Finding

AI and predictive analytics can substantially improve U.S. tax administration performance (better risk scoring, audit targeting, forecasting, and fraud detection) but, if deployed naively, will amplify distributional harms. The review reports robust performance gains (e.g., AUC in compliance scoring ≈ +14 percentage points; audit no‑change rates down ~15–20 pp; revenue‑forecast MAPE down ~15–30%; fraud detection lag from ~47 days to <4 hours), while highlighting structural problems (selective‑labels bias, legacy data infrastructure, §6103 privacy limits, and LLM dual‑use risks) that require governance, equity‑by‑design, XAI, MLOps, and possible statutory modernization.

Key Points

• Quantified benefits (from Tier‑1 and Tier‑2 evidence):
  • Compliance scoring: gradient boosting models AUC ≈ 0.85 vs logistic ≈ 0.71 (+~14 pp).
  • Audit selection: ML ensembles reduce audit no‑change rates by ~15–20 pp and increase net revenue per audit.
  • Revenue forecasting: deep LSTM models reduce corporate MAPE by ~22–30% vs ARIMA; state forecasts see ~15–25% MAPE reductions in volatile periods.
  • Fraud/anomaly detection: autoencoder + LSTM AUC‑PR 0.89–0.93 vs 0.72–0.78 for rule systems; detection lag drops to hours.
  • Secondary effect: algorithmic deterrence may boost voluntary compliance by ~10–15%.
• Core methodological challenge — selective‑labels problem:
  • Historical audit outcomes exist only for audited filers, biasing models toward historically audited populations and under‑flagging under‑audited, often higher‑income groups.
  • Proposed remedies: randomized exploration (5–10% randomized audits), NRP‑style reference audits, inverse propensity scoring/off‑policy evaluation, and reformulating targets from binary classification to regression (expected adjustment amounts).
• Methods and architectures commonly recommended:
  • Models: gradient boosting (with monotonicity/sparsity constraints), ensembles, neural nets, LSTM for time series, autoencoders, graph neural networks for networked fraud, RAG/transformers for document processing.
  • Privacy/architecture: federated learning + differential privacy (recommended ε ≤ 1.0 as a starting guidance), but legal questions remain whether gradient/parameter exchange violates 26 U.S.C. § 6103.
• Governance, accountability, and explainability:
  • Map to NIST AI RMF, EO 14110, and OMB M‑24‑10; require model cards (disaggregated), algorithmic impact assessments, adversarial/red‑team testing, continuous MLOps monitoring, and TIGTA/Oversight review.
  • Equity‑by‑design prescriptions: vertical equity monotonicity (audit rate should decrease across AGI deciles controlling for complexity), demographic disparity ratio cap (≤1.5:1 without justification), burden proxies (audit prep costs), and disaggregated appeals/resolution metrics.
• Agentic AI taxonomy for IRS deployment:
  • Advisor (recommendations, HITL), Conditional (low‑stakes automated actions with human contestability), Agentic (autonomous enforcement actions — legally problematic and likely out of bounds without Congressional authorization).
• Practical limits/risks:
  • Legacy IT (COBOL) constrains ML pipelines; model drift and calibration require MLOps; opaque deep models raise administrative‑law explanation concerns; LLMs pose dual‑use risks for avoidance strategies.

Data & Methods

• Review design: expert‑led narrative (depth over exhaustive systematic review).
• Search strategy: structured queries in SSRN, Google Scholar, Web of Science, Scopus, and government/institutional repositories for 2020–2026.
• Screening: 312 records identified → de‑duplication, full‑text screening → 37 included sources.
• Evidence grading:
  • Tier 1: peer‑reviewed empirical journals (14 sources, 38%).
  • Tier 2: institutional/working papers (13 sources, 35%).
  • Tier 3: preprints/grey literature (10 sources, 27%) — hedged.
• Inclusion criteria: empirical AI/ML applied to tax or closely related public‑finance problems with clear U.S. applicability; excluded purely theoretical economic models, unattributed web content, and redundant reports.
• Common evaluation metrics reported across studies: AUC/AUC‑PR, MAPE, no‑change audit rate, detection lag, yield accuracy, Brier/ECE calibration scores, and fairness/disparity metrics.
• Recommended experimental/causal designs left as future work: randomized exploration tranches, NRP‑style reference audits, longitudinal/difference‑in‑differences or instrumental designs to estimate equity effects and deterrence.

Implications for AI Economics

• Revenue and efficiency impacts:
  • Improved audit targeting and forecasting can materially increase recovered revenue per audit and reduce forecast error, improving short‑term cash management and longer‑term fiscal planning. Even modest accuracy gains scale to large dollar impacts given a >$600B gross tax gap.
  • Algorithmic deterrence could amplify revenue gains beyond direct detection.
• Distributional trade‑offs:
  • Classifier objectives optimized for ROI can concentrate audit burden on lower‑income filers (e.g., EITC audit rates 5–7× higher in reported studies) unless re‑specified as regression targets or constrained by fairness limits. Choice of target function (classification vs regression) therefore has direct redistributive consequences in enforcement.
• Policy and institutional consequences:
  • Legal and statutory constraints (26 U.S.C. § 6103, administrative law doctrines, due process/non‑delegation) shape feasible system architectures (e.g., limits on gradient sharing, need for Congressional authorization for agentic actions).
  • Regulators and procurement must embed XAI, fairness auditing, MLOps, and workforce upskilling requirements to capture benefits while controlling harms.
• Research and modeling priorities for AI economics:
  • Causal studies to quantify distributional impacts of algorithmic selection and deterrence, including randomized exploration experiments.
  • Multi‑objective optimization frameworks that explicitly trade off revenue, equity, cost, and taxpayer burden.
  • Cost–benefit analyses incorporating privacy mitigation costs (federated learning + DP), operational modernization costs, and legal compliance/oversight costs.
  • Economic modeling of second‑order effects: taxpayer behavior (chilling/gaming), market responses to automated enforcement, and welfare impacts of improved forecasting on fiscal policy.
• Practical recommendation for economists advising policy:
  • Evaluate ML deployments not only on predictive accuracy and ROI but on distributional metrics, operationalizability under current statutes, and the full fiscal/ecosystem costs of safe deployment (MLOps, audits, appeals, privacy compliance).
  • Advocate experimental pilots (e.g., 5–10% randomized audit tranche; IRS–SSA federated pilot under strict review) with pre‑registered fairness and welfare outcomes to inform scale decisions.