AI Was Already in the Room
When I joined Deloitte in Seattle, I expected AI to be something we discussed in strategy meetings. Instead, it was already part of daily work. Algorithms flagged anomalies in transaction data. Automated tools handled portions of documentation testing. Patterns that would have taken a human analyst hours to surface were appearing in minutes.
None of it felt dramaticโthat was what surprised me most. The shift was not a big announcement. It was incremental, practical, already underway, and it quietly changed what I actually spent my time on.
Some articles on AI in accounting automation[1] examine how intelligent systems are being integrated into accounting infrastructure across organizations. The conclusion these studies found was that the technology is not replacing accountants, but it is changing what accounting work actually involves.
This article aims to give CPAs a realistic picture of that change.
Before getting into specific applications, it helps to draw a distinction that does not always get made clearly. AI tools in this space operate across two different domains.
What AI Is Actually Doing in Accounting Right Now
Before getting into specific applications, it helps to draw a distinction that does not always get made clearly. AI tools in this space operate across two different domains. Some are deployed by management to automate internal accounting and reporting processesโEnterprise Resource Planning (ERP)-based document processing, automated journal entries, real-time reconciliations. Others are used directly within audit engagements: tools that support evidence gathering, transaction testing, and anomaly detection during the audit itself. These two domains often overlap in practice, and auditors increasingly need to understand bothโnot just what their own tools do, but how management’s automated systems work and where they might break down.
Processing documents
The most widespread application of AI tools is document processing. Computer vision models combined with natural language processing can read invoices, contracts, and receipts, extract the relevant fields, and generate accounting entries automatically. This is primarily a management-side tool, implemented within ERP systems like SAP or Oracle rather than by auditors directly.
The technology is not flawless. Accuracy depends heavily on data quality and how standardized the incoming documents are. For organizations with large, structured transaction volumes, it removes a significant amount of manual data entry. More interestingly, it shifts where errors tend to occur. Instead of catching data entry mistakes after the fact, the challenge becomes ensuring the input data is clean enough for the algorithm to work with in the first placeโa different kind of quality control problem, and one that auditors increasingly need to evaluate when assessing management’s control environment.
Classifying transactions
Machine learning models trained on historical transaction data are increasingly used to classify accounting operations and generate journal entries. Two techniques appear most often. Gradient boosting works by building a sequence of decision trees, where each tree corrects the errors of the one before it; the result tends to be highly accurate on structured financial data. Neural networks take a different approach, learning complex patterns from large datasets by passing information through interconnected layers loosely modeled on how neurons work in the brainโthey are especially useful for detecting non-linear relationships that simpler models miss.
Both approaches work by learning how a specific organization records its transactions over time, then applying that pattern to new ones. Research surveying 454 accounting professionals found measurable improvements in processing speed and data quality when these systems were in use,[2] results that are consistent with what the broader literature shows. What the research tends to underreport is how quickly performance degrades when input data becomes inconsistent or when the business model changes. The model keeps applying the old pattern until someone catches it.
Continuous audit monitoring
Traditional internal audit is built around sampling. A representative portion of transactions gets selected, reviewed, and used to draw conclusions about everything else. The approach works, but its limitation is obvious: most of what happened never gets examined.
Continuous monitoring systems work differently. They process every transaction as it occurs, flagging anomalies in real time rather than catching them during a quarterly review. Among the methods used here, unsupervised learning techniques such as clustering and autoencoders are particularly well-suited to this task. Autoencoders are algorithms trained to compress transaction data into a compact representation and then reconstruct it; transactions the model struggles to reconstruct accurately stand out as potential anomalies. Crucially, these methods do not require predefined fraud rulesโthe system identifies what looks unusual relative to established patterns, not what matches a checklist.[3]
From an audit standpoint, continuous monitoring systems implemented by management represent a meaningful internal control that auditors need to understand and test. When they work well, they move control closer to the moment a transaction occurs and shrink the window between a problem and its detection.
Three Ways Organizations Are Implementing AI
Across the organizations covered in my research, three implementation models appear consistently, each suited to a different level of organizational readiness:
1. Modular ERP integration is the most common entry point. AI components are added to existing ERP systems to handle specific tasksโtransaction classification, data validation, anomaly alertsโwithout disrupting the broader infrastructure. Most organizations start here because the investment and risk are manageable.
2. Centralized analytics platforms pull data from multiple accounting systems and run cross-functional analysis, generating consolidated reports automatically. These are better suited to larger organizations with complex, multi-entity structures where a single ERP module is not sufficient.
3. Distributed continuous audit systems are the most advanced approach. Algorithms run in the background at all times, monitoring every transaction and blocking entries when deviations are detected. The control value is high, but so is the infrastructure requirement. Organizations with fragmented or inconsistent data typically struggle here.
What This Means for CPAs
AI is not replacing accountants. It is, however, taking over a meaningful share of what many accountants spend their days onโmanual transaction processing, routine reconciliations, repetitive data entry, sampling-based transaction testing. Intelligent systems handle these tasks efficiently and at scale.
Where they fall short is interpretation. A flagged anomaly still needs a human to determine whether it represents a real problem or a legitimate exception. A pattern learned from last year’s data may not apply to a business that has reorganized. An algorithm cannot explain its reasoning to a regulator in terms that satisfy a legal or professional standard. These are judgment calls, and judgment has not been automated.
Research consistently shows that as AI absorbs routine processing, accountants are spending more time on interpretation, strategic analysis, and communication of findings (Greenman et al., 2024; Sampaio & Silva, 2025).[4] Whether that feels like a welcome shift or an unsettling one probably depends on the individual. Either way, it means the skills that matter most going forward are less about executing manual procedures and more about interrogating automated ones.
The most practical thing CPAs can invest in right now is data literacyโnot programming, not a computer science degree, but the ability to look at what an algorithm is producing and ask the right questions about whether it is working correctly.
Where This Goes Wrong
These systems have real limitations. Practitioners need to go in with eyes open.
Data quality problems do not disappear when you add AIโthey scale. An algorithm trained on inconsistent or error-prone source data will reproduce those errors across every output it generates, often faster and more systematically than a human would. I have seen this happen in practice. It is not always obvious until you look closely at the underlying data.
Explainability is a genuine problem for financial reporting. Neural networks can achieve high accuracy while remaining nearly impossible to interpret, even for the engineers who built them. In accounting and audit contexts, every entry needs a logical explanation. Regulators and auditors expect to understand why a decision was made, and algorithmic opacity does not satisfy that expectation in most compliance environments.
Model drift is easy to overlook. A model trained on last year’s transaction history may quietly start producing wrong outputs when the business changes its pricing structure, accounting policies, or operational model. This does not announce itself. It requires ongoing monitoring, which organizations often underinvest in after the initial implementation.
Finally, the cost of full AI integration is not trivial. Large firms can absorb it. Many small and mid-size practices cannot, at least not yet. The technology is advancing faster than it is becoming accessible, and that gap is worth naming.
A Final Thought
When I was building my accounting career across Kazakhstan, the United Kingdom, and the United States, the core of the work looked the same everywhere: careful, methodical, and dependent on human judgment exercised over painstaking manual processes.
The manual layer is moving to machines. What stays is the judgmentโknowing when to trust an output, when to question it, and how to explain it to someone who does not understand the underlying system. That is harder than it sounds, especially when the algorithm is usually right.
The practitioners who navigate this well will be the ones who engage with these tools seriously enough to understand where they break down. That has always been accounting. The tools are just different now.
Cite as “AI in the Audit Room: What It Actually Looks Like in Practice” DOI Reference: https://doi.org/10.67283/7B80AB19
References
Bou Reslan, F., & Jabbour Al Maalouf, N. (2024). Assessing the transformative impact of AI adoption on efficiency, fraud detection, and skill dynamics in accounting practices. Journal of Risk and Financial Management, 17(12), Art. 577. https://www.mdpi.com/1911-8074/17/12/577.
Greenman, C., Esplin, D., Johnston, R., & Richards, J. (2024). An analysis of the impact of artificial intelligence on the accounting profession. Journal of Accounting, Ethics & Public Policy, 25(2), 188. doi.org/10.60154/jaepp.2024.v25n2p188.
Kokina, J., Blanchette, S., Davenport, T. H., & Pachamanova, D. (2025). Challenges and opportunities for artificial intelligence in auditing: Evidence from the field. International Journal of Accounting Information Systems, 56, Art. 100734. DOI:10.1016/j.accinf.2025.100734.
Primzharova, L. (2026). Application of AI technologies in accounting automation and internal audit. Putevoditel Predprinimatelya, Vol. 19, No. 2. DOI: 10.24182/2073-9885-2026-19-2-29-41.
Sampaio, C., & Silva, R. (2025). Digital transformation in accounting: An assessment of automation and AI integration. International Journal of Financial Studies, 13(4), Art. 206. https://www.mdpi.com/2227-7072/13/4/206.
Al Najjar, M., Gaber Ghanem, M., Mahboub, R., & Nakhal, B. (2024). The role of artificial intelligence in eliminating accounting errors. Journal of Risk and Financial Management, 17(8), Art. 353. DOI:10.3390/jrfm17080353.
Alruwaili, T. F., & Mgammal, M. H. (2025). The impact of artificial intelligence on accounting practices: An academic perspective. Humanities and Social Sciences Communications, 12, Art. 1197. DOI:10.1057/s41599-025-05004-6.
[1] Primzharova, โApplication of AI technologies.โ
[2] Bou Reslan and Jabbour Al Maalouf, โAssessing the transformative impact,โ art. 577.
[3] Kokina et al., โChallenges and opportunities for artificial intelligence in auditing,โ art. 100734.
[4] Greenman et al., โAn analysis of the impact of artificial intelligence,โ 188; Sampaio and Silva, โDigital transformation in accounting,โ art. 206.


