When I joined Deloitte in Seattle, I expected AI to be something the partners discussed in strategy meetings. Instead it was already in the daily work. Algorithms were flagging anomalies in transaction data. Automated tools were handling parts of documentation testing. Patterns that would have taken an analyst hours to surface were showing up in minutes. Nobody announced it. It was just there.
That experience changed how I think about AI in accounting—and it is why I have spent the years since studying the transition more formally. My 2026
The most visible application right now is document processing. Computer vision models combined with natural language processing can read invoices, contracts and receipts, pull the relevant fields, and generate accounting entries automatically. The accuracy is not perfect—it depends heavily on how clean and standardized the incoming documents are. But for organizations processing large transaction volumes, it eliminates a real amount of manual data entry. The practical shift is subtle: instead of catching data entry errors after the fact, the job becomes making sure the input data is clean enough for the algorithm to read correctly. Same problem, different place in the process.
Transaction classification is changing too. Machine learning models trained on a firm's historical data learn how that specific organization records its operations over time, then apply that pattern to new transactions. A survey of 454 accounting professionals found measurable improvements in processing speed and data accuracy when these systems were in use. What the research tends to understate is what happens when the input data gets inconsistent or when the business changes significantly. The model keeps applying the old pattern until a human notices something is off. This is no minor caveat. It is, in fact, what requires the most professional attention.
Continuous audit monitoring is the change that matters most for internal auditors. Traditional audit is built on sampling. You review a portion of transactions and draw conclusions about the rest. It works, but you are always working with incomplete information. Continuous monitoring systems analyze transactions as they occur, flagging unusual patterns in real time rather than surfacing them in a quarterly review. The algorithms do not need a predefined definition of fraud. They identify transactions that look different from everything else, and they do it across the full population, not a sample. The window between when something goes wrong and when someone notices it gets significantly shorter.
None of this replaces the accountant. The systems cannot evaluate whether a flagged anomaly represents a real problem or a legitimate exception. They cannot assess whether the pattern they learned from last year's data still makes sense for a business that has since changed its pricing model or acquired a new entity. They cannot explain their reasoning to a regulator in a way that satisfies a legal standard. A study published in the Journal of Accounting, Ethics and Public Policy found that as AI handles routine processing, accountants are shifting toward interpretation, judgment and strategic analysis—which is, frankly, a better use of the profession's skills. But it requires different preparation than most practitioners received.
There are real failure modes worth understanding. Algorithms amplify data quality problems rather than fixing them, so if the source data has errors or inconsistencies, the AI output will reproduce those problems at scale, and it will not be obvious until you look closely. Neural networks can produce highly accurate outputs while remaining difficult to interpret, which creates compliance friction in financial reporting contexts where every entry needs a logical explanation. Models trained on historical data can drift when business conditions change, quietly producing wrong outputs until someone catches it. And the infrastructure costs of full AI integration are significant enough that the technology is not equally accessible across firm sizes.
The skill that matters most right now is not programming or data science. It is the ability to look at what an algorithm is doing and ask the right questions about whether it is doing it correctly. CPAs who develop that habit—who treat AI outputs the way they would treat any other work product that needs review—will be in a strong position. The ones who either ignore the technology or trust it uncritically are both going to have problems, just different ones.
I grew up in Kazakhstan and built my accounting career across three countries. The profession looked essentially the same in all of them: careful, methodical, built on human judgment applied to manual processes. That foundation is being rebuilt right now. The manual process is moving to machines. The judgment is staying with us. The practitioners who engage with that change seriously, who understand what these systems can do and where they fail, will define what the profession looks like in 10 years.
