Accounting and Artificial Intelligence: What to Delegate, What to Review, What to Keep

In practice, accounting and artificial intelligence means software that classifies transactions, drafts documents, retrieves authority, and summarizes records, all under review by a professional who remains responsible for the output. The open question in most firms is no longer whether to use AI. It is where the review line sits, task by task. This framework draws that line.

What people actually mean by AI in accounting

Three different technologies get collapsed into one label, and each one earns a different level of trust.

Rules-based automation applies fixed logic: bank feeds, recurring journal entries, approval routing. It does what it is configured to do, every time. The risk is a rule that was wrong from the start or has gone stale, bad source data, or an exception nobody anticipated, not a creative error.

Classification and analytical models make probabilistic calls: coding a transaction to a 150-line chart of accounts, flagging an outlier in a variance report, matching a payment to an invoice. These models carry confidence levels, and that matters later in this framework.

Generative drafting produces new text: a research summary, a client letter, a flux explanation, a memo. It is the newest layer and the one where output can sound authoritative while being wrong.

Modern platforms blend the three, and generative models increasingly do classification work too, but the failure modes stay distinct. A delegation decision that treats them as one thing will be wrong for at least two of them.

The accounting workflow in five stages

Field research on AI-enabled accounting platforms describes the work as a chain of five stages (Choi and Xie, Journal of Accounting Research, 2026):

1. Data collection: pulling bank feeds, documents, and client records into the system.
2. Data processing: cleaning, matching, and reconciling what came in.
3. Account categorization: coding transactions to the ledger.
4. Confirmation and review: checking the work before it hardens.
5. Financial reporting: producing the statements and the narrative around them.

The boundaries matter because an error gets harder to detect the further it travels downstream. A misclassified transaction caught at stage three is a recode; the same error found after the financial statements go out may mean correcting statements already in clients' hands. The review gate belongs at the stage where the error would stop being visible.

A delegation framework for accounting and artificial intelligence

The table below sets a review line for the core tasks of an accounting and tax practice. The review steps in every row are my recommended controls; the evidence grade names what stands behind the row, not who prescribed the procedure. The grades:

• Field evidence: a peer-reviewed study measured AI performance or human-AI interaction on this task.
• Professional standard: a rule of professional conduct or federal practice sets the underlying duty; the implementation is mine.
• Firm practice: this is my own review process, offered as a working default and nothing more.

Two rules sit on top of the table. First, confidence signals change the gate: a low-confidence item needs human resolution before it is accepted or posted. Second, a number generated inside an AI's response is never accepted as a computed result; more on both below. For the research row, how you prompt matters as much as how you review: see Prompting AI for Tax Research.

What the research shows about artificial intelligence in accounting, and what it does not

Field evidence on artificial intelligence in accounting now has a peer-reviewed anchor: Choi and Xie, "Human + AI in Accounting: Early Evidence from the Field," Journal of Accounting Research 64(3), 1333-1373 (2026). The design matters as much as the findings. The study combines two survey waves of 277 accountants producing 1,038 weekly observations, transaction-level data from 79 US private firms covering roughly 200,000 records, and a randomized experiment in which 99 accountants classified 43 transactions against a 153-category chart of accounts.

The survey and platform findings are associations, not guarantees for your firm. A one standard deviation increase in AI use was associated with accountants reporting about 19% more clients supported per week, and with a 2.6 percentage point reduction in time spent on data entry, roughly one hour out of a 40-hour week. Firms whose accountants used generative AI showed more granular ledgers and faster month-end closes. The randomized experiment, the study's one causal component, found that AI assistance improved classification accuracy on average, but that leaning on non-consensus AI recommendations increased the risk of error.

The authors' conclusion is the one worth carrying into partner meetings: the evidence suggests AI augments rather than replaces professional judgment. Nothing in the study supports a claim that AI files returns, signs statements, or runs an engagement on its own. And none of these figures is a benchmark to promise a client or a managing partner; they describe the sampled firms, not yours.

For anyone weighing an accounting career against the headlines, the honest reading is narrower: the measured effects are task-level, more client work supported and less time on data entry. What happens to accounting employment over time is a question this study was not built to answer.

Where model uncertainty should trigger a human

The study's override analysis is small but instructive. Within one firm's nine internal accountants, a drop in the model's confidence score from 1.0 to 0.7 was associated with roughly a 10 percentage point increase in the likelihood that the accountant overrode the AI's suggestion. That effect size belongs to that firm; do not import it as a universal threshold.

The design principle travels even if the number does not. Treat a confidence score as one review input, not a verdict: a displayed score is not automatically calibrated to real accuracy, so ask how the vendor validated it and weigh it alongside materiality, unusual facts, and your own observed error rates. Route low-confidence items to an experienced reviewer rather than the least expensive one, sample accepted items on a schedule, high-confidence ones included, and document the reason each time a human overrides the model. The override log becomes your evidence of supervision, and over time it tells you which task classes the tool has actually earned.

The three review traps

The AICPA Professional Ethics Division has named the three habits that corrode AI review (Journal of Accountancy, August 2026):

• Automation bias: trusting the machine over your own contrary reasoning. Counter it by asking what evidence would change the answer, then looking for it.
• Overconfidence bias: treating the tool as infallible. Counter it by sampling: re-perform a slice of delegated work cold, on a schedule.
• Anchoring bias: letting the tool's first answer frame everything after it. Counter it by forming your own expectation before you open the output.

One more constraint belongs in every firm's AI policy. Digits CEO Jeff Seibert's warning on CPA.com's AI resource hub is blunt: large language models produce "generated output, not computed answers" (CPA.com). His warning is broad; the rule I draw from it is narrower, because some AI systems can run real calculations by operating a spreadsheet, a calculation engine, or executed code. Never rely on a number that exists only in a model's prose. Every figure in an AI-assisted deliverable traces to a calculation someone can inspect, and someone inspects it.

What cannot be delegated, under the standards that apply to you

The AICPA Code of Professional Conduct does not bend for new technology. The General Standards Rule (ET Sections 1.300.001 and 2.300.001) requires members to undertake only services they can complete with professional competence and due professional care, and competence here includes understanding the tool's limits in the specific context you are using it (Journal of Accountancy, August 2026). The Confidential Client Information Rule (ET Section 1.700.001) governs what goes into a tool in the first place, including whether client consent is required before client data touches a third-party system. The interpretation on disclosing information to a third-party service provider (ET Section 1.700.040) sets out the alternatives: contractual confidentiality assurances from the provider, or specific client consent. Members in business carry the parallel obligations in Part 2 of the Code.

For federal tax practice, the IRS Office of Professional Responsibility has set out how Circular 230 applies when AI assists the work; the short version is that due diligence, competence, and responsibility for written advice all remain with the practitioner. We cover that guidance section by section in Responsible AI Use in Tax Practice.

The line that holds across every standard: no filing, position, opinion, or signed deliverable is issued by a tool. The responsible practitioner or preparer issues it, after review, and owns it. Using AI never transfers that responsibility; it only changes what the review looks like.

Client data deserves its own gate before any of this starts. Section 7216 and the FTC Safeguards Rule set the rules for what client tax information can flow into which systems, covered in Using AI With Client Tax Data. And vendor selection, including the security certifications worth demanding, is a separate exercise we have already mapped in AI Tax Software Selection.

What this looks like in one engagement

The cleanest example I have is my own return. I ran it through Marble, and it kept flagging that my return said someone else could claim me as a dependent. My first reaction was the one this article warns about, just in reverse: I assumed the tool had invented a checked box, and I nearly dismissed the flag.

It had not. Years earlier, preparing my own return, I had accidentally checked the box indicating that someone else could claim me as a dependent, and that answer carried forward through multiple filing years without anyone catching it, me included. The software rolled it forward faithfully each year, because preserving prior-year answers is exactly what carryforward data is built to do. For a status that feeds the standard deduction formula and a list of credit eligibility tests, that could be an expensive oversight.

Two lessons came home. First, the review gate runs in both directions: verify a flag before you act on it, and verify it before you dismiss it. A flag is not an answer; it is an instruction to go look at the source. Second, prior-period consistency is not evidence of correctness. An error entered once and rolled forward will agree with itself every year, which is precisely why a fresh check against the underlying facts catches what year-over-year review never will.

An adoption sequence that holds up

Rolling the framework out is less about the tool and more about the order of operations:

1. Pick one low-risk task class, such as routine transaction classification, not the whole workflow.
2. Define the review gate before rollout: who reviews, what they verify, what triggers escalation.
3. Train the staff who will do the reviewing, not just the staff who will do the prompting.
4. Document approvals and overrides from day one, so supervision is evidenced, not asserted.
5. Revisit the delegation table annually, and after any material change to the tool, the workflow, or your observed error rates. The gates should be a decision, not a leftover.

A one-page delegation checklist with fields for the task, permitted inputs, named reviewer, verification steps, and sign-off turns this from an article into a control. The economics of the rollout, including what the time savings are actually worth, are a separate question we take up in AI ROI for Accounting Firms.

For the research and drafting rows of the table, this is the workflow Marble was built for: ask federal or state tax questions in plain English, get citation-backed answers that link directly to the primary authority, and generate memos and client communications that arrive as drafts for your review, not as finished positions. The review gate stays where the standards put it, with you. Sign up for Marble.

Frequently asked questions about accounting and artificial intelligence

Is AI going to replace accounting?

The field evidence measures tasks, not headcount. A 2026 Journal of Accounting Research study found AI use associated with accountants supporting more clients and spending less time on data entry, and its authors describe augmentation of professional judgment rather than replacement. What that means for accounting employment over time is an open question; what it means today is that the work is shifting toward review, communication, and responsibility for AI-assisted output.

How is AI being used for accounting?

Across three layers: rules-based automation for feeds and recurring entries, classification models for transaction coding and anomaly detection, and generative tools for drafting research summaries, memos, and client communications. In every layer, firms with durable results put a defined review gate between the AI's output and anything a client or regulator sees.

Can ChatGPT do my bookkeeping?

It can assist with bookkeeping tasks, but its output is not a completed set of books. Some chatbots can run code to perform real calculations; a number that exists only in a chat response was generated, not computed. Before anything posts or issues, require traceable source data, inspectable calculations, reconciliation against the records, and a human approval step.

Is accounting a good field to go into with AI?

The field evidence points to augmentation: accountants using AI served more clients and shed data-entry hours, while judgment-heavy work stayed human. The skill mix is shifting toward review, client communication, and knowing where the review line sits, which arguably raises the value of professional judgment rather than lowering it.

What has to stay with the professional?

Responsibility. Using AI does not transfer the competence and due care obligations under the AICPA General Standards Rule, decisions about confidential client information, or the review, approval, and signature requirements that apply to the engagement. AI can draft toward those moments. The responsible preparer or practitioner owns them.