AI in the Accounting Workflow: Use Cases, Limits, and Judgment

The most useful question about AI tools for accountants isn’t whether to use them. It’s where in the workflow they actually deliver, and where the professional’s judgment is still doing the work that matters.

The gap between what AI tools are marketed to do and what they reliably do inside a real accounting or tax practice is narrowing. Some tasks have reached a point where AI assistance is genuinely time-saving and accurate enough to trust with appropriate review. Others remain friction-prone, output-unreliable, or professionally risky without robust human oversight. Knowing the difference is where AI competency actually lives.

Where AI Is Delivering in Practice

Tax research has emerged as one of the clearest productivity wins. Large language model tools can surface relevant code sections, regulations, and rulings quickly, summarize them in plain language, and flag conflicting authority. The caveat that applies everywhere applies here too: the output requires review, and hallucinated citations are a documented failure mode. Used as a starting point rather than a conclusion, AI-assisted tax research compresses the time from question to informed answer.

Document work is another high-adoption area. Engagement letters, client-facing summaries, meeting notes, and first-draft management commentary are all tasks where AI can produce a workable draft quickly. A practitioner who reviews and edits a draft is generally faster than one starting from a blank page, as long as the review step is taken seriously and not skipped because the draft looks polished.

In audit and assurance, data analytics applications are seeing genuine adoption. Anomaly detection in large transaction populations, journal entry testing, and continuous monitoring workflows are areas where AI tools can process volume that would take staff teams far longer to work through manually. The AICPA’s work on AI integration in audit practice, including collaboration with vendors on the Dynamic Audit Solution (DAS) platform, reflects the direction the profession is moving at an institutional level, per Journal of Accountancy reporting from February 2026.

What Does Agentic AI Mean for a CPA’s Daily Work?

Most of the AI tools CPAs encounter today are generative: the practitioner provides a prompt, the tool produces an output, the practitioner reviews it. The loop is short, the human is present at every step, and the scope of what the tool does without oversight is limited.

Agentic AI changes that model. An AI agent can be given a goal and then execute a sequence of tasks to achieve it without requiring input at each step. In accounting contexts, that might mean an agent that queries the general ledger, reconciles to subledger data, flags discrepancies, and generates a draft report — all before a human reviews the results. The Journal of Accountancy (February 2026) describes agentic AI performing real-time reconciliation and continuous monitoring tasks in audit workflows in exactly this way.

More autonomy means more to review, not less.

The human-in-the-loop principle is the professional standard response to that dynamic. Agentic systems can process more volume and execute more steps than a practitioner reviewing each output manually. But the practitioner’s professional responsibility for the work product doesn’t diminish with the level of automation. It makes the review of what the agent did more consequential, not less.

The Client Data Problem That Catches Firms Off Guard

When a practitioner submits client information to an AI tool, that submission is a disclosure under the AICPA’s Confidential Client Information Rule (Section 1.700.001 of the AICPA Code of Professional Conduct). The rule requires client consent before confidential information is disclosed. It applies regardless of whether the recipient is a person or a software platform.

The friction point in practice is that many AI tools used casually in accounting workflows — general-purpose large language models accessed through consumer interfaces — operate under data terms that allow submitted content to be used for model training or retained on external servers. Practitioners who haven’t read those terms are making a consent decision on their client’s behalf that the client hasn’t authorized.

Enterprise-tier products exist specifically to address this. Many AI vendors offer enterprise configurations with contractual data handling commitments that differ from consumer defaults: data isn’t used for training, it’s processed in isolated environments, retention periods are defined. Evaluating which tier a tool operates on, and whether that tier is consistent with client confidentiality obligations, is a threshold question before any client data goes in.

AI Tools for Accountants: Where Professional Judgment Still Lives

AI tools are reliable at surface-level tasks: pattern recognition in data, text generation from prompts, retrieval of information from large corpora. They’re less reliable at tasks that require contextual judgment, professional skepticism, or an understanding of what a client’s situation actually means beyond the numbers.

A tax research tool can surface the relevant code section. Determining whether that section applies given the client’s specific facts, what the risk of a position is, and how to document it is practitioner judgment. An AI that summarizes financial statements can produce a coherent management commentary. Deciding whether the narrative accurately reflects the business reality, and whether it’s the right story to tell, is still a human call.

The AICPA’s framing of AI adoption as a competency question is useful here. Competent use of AI means knowing what the tool can and can’t do reliably, applying appropriate skepticism to its outputs, and maintaining professional responsibility for the final work product. The AICPA published non-authoritative guidelines for AI use in forensic and valuation services engagements in 2025, which reference the Code of Professional Conduct’s Compliance With Standards Rule as the applicable framework. The principle extends beyond forensic and valuation work: the practitioner’s signature remains an attestation regardless of which tools were used.

Before You Adopt: Evaluation Questions That Reduce Exposure

Firms and practitioners building AI into their workflows are applying a consistent set of evaluation criteria before tools reach client work. Data handling comes first: where does client data go, under what contractual terms, and is that consistent with confidentiality obligations? Domain accuracy is the second filter: a tool that performs well on general text tasks may produce unreliable output on technical tax or audit questions.

Documentation is the third consideration. Recording what tools were used, what data was submitted, and how outputs were reviewed creates an engagement file trail that supports both quality management and professional liability defense. The AICPA’s AI guidelines for forensic and valuation work specifically call out documentation of prompts and outputs as part of the engagement record. Applying that discipline more broadly is reasonable professional practice.

AI is reshaping the work. The practitioners who get the most out of it aren’t the ones who trust it most. They’re the ones who know exactly where to trust it and where to stay skeptical.

Sources: AICPA Code of Professional Conduct, Section 1.700.001 (Confidential Client Information Rule), pub.aicpa.org; AICPA Guidelines for Responsible Use of AI in Forensic and Valuation Services Engagements (2025), aicpa-cima.com; Journal of Accountancy, How AI Is Transforming the Audit (February 2026), journalofaccountancy.com.