AI in Accounting: What It Does Well, What It Doesn't, and What Changes for Your Job

8/28/2026

Most writing about artificial intelligence in accounting falls into one of two camps. One says the profession is about to be transformed beyond recognition. The other says it is a fancy autocomplete that no serious practitioner should trust. Both are wrong in the same way: they talk about "AI" as if it were a single thing with a single verdict.

It is not. The tools now sold to accountants do several quite different jobs, and each one has its own failure pattern. A tool that reads invoices well can be useless at research. A tool that drafts a clean client letter can invent a citation with complete confidence. The practical question for an accountant is never "is AI good?" It is "for this task, what does the tool get right, how does it fail, and what review step catches the failure?"

This post is the broad map. It walks through the main areas of accounting work (the close, tax, audit and advisory), sets out where AI genuinely helps in each, where it does not, and what the risk and control side looks like. Our narrower posts go deeper on individual use cases, and they are linked where they apply.

What "AI" Actually Means in an Accounting Context

The label covers at least four kinds of technology, and separating them clears up most of the confusion.

Machine learning classifiers. Software that learns patterns from historical data and applies them to new items. In accounting this is the engine behind transaction coding suggestions, anomaly flags and matching. It is good at volume and consistency. It is bad at anything it has not seen before, and it inherits whatever errors sit in the history it learned from.

Document extraction. Tools that read semi-structured documents (invoices, receipts, bank statements, tax forms) and pull out fields. Accuracy on clean, common documents is high. The errors that remain are the dangerous kind: plausible numbers in the right field that happen to be wrong.

Generative language models. The chat-style tools that write, summarize, explain and answer questions. They produce fluent text from patterns in their training. They do not look anything up unless they are connected to a source, and they do not know when they are wrong.

Rule-based automation. Not AI at all, strictly, but it is often sold alongside it. Bots and workflow rules that do exactly the same steps every time. Our post on robotic process automation in bookkeeping covers that category on its own terms.

The reason the distinction matters: each category fails differently, so each needs a different control. A classifier fails quietly at the edges. An extractor fails with believable numbers. A language model fails with believable prose. A rule fails loudly when the input changes. A firm that applies one review approach to all four will over-review some work and under-review the rest.

Where AI Helps Across the Main Areas of Accounting Work

The monthly and year-end close

The close is the most mature use case because so much of it is repetitive, data-heavy and rule-bound.

What works well:

• Transaction coding suggestions for recurring vendors and customers, where history is long and consistent
• Bank and subledger matching, including fuzzy matches on amount, date and description that a strict rule would miss
• Flux analysis drafts, where a tool compares balances period over period and drafts the first-pass explanation for the reviewer to correct
• Accrual and reclass suggestions based on prior-period patterns

What does not: anything that depends on knowing something the data does not contain. A tool cannot know a contract was amended last Tuesday, that a vendor changed its billing cycle, or that a large payment was a deposit rather than an expense. Those facts live in people's heads and email, not in the ledger.

The sequence that makes close automation pay off (measure, redesign, standardize, then automate) is set out step by step in our post on automating the month-end close with AI.

Tax preparation and research

Tax work splits sharply into two halves, and AI's usefulness splits with it.

Data gathering and input is where the real savings sit. Extracting figures from client documents, comparing this year's data to last year's, and flagging missing items all suit current tools well, provided a preparer checks extracted figures against the source.

Research and positions are where the risk sits. A general-purpose language model can produce a confident answer with a code section that does not say what it claims, or a case that does not exist. Retrieval-based tools that search a real authority database and cite what they found are a better design, but the citation still has to be read before it goes in the file. Our post on generative AI for tax research explains how to use these tools without putting an unverified authority in a workpaper.

Review sits in the middle. A tool can compare a return to the prior year and to the source documents and list differences worth a look. It cannot decide whether a position is correct. Our post on using AI to review tax returns covers where that line falls.

Audit and assurance

Audit has used data analytics for years, and AI extends the same idea: test whole populations instead of samples, and point the auditor's attention at the unusual items.

• Full-population journal entry testing for entries posted at odd times, by unusual users, to rarely used accounts, or in round amounts
• Contract and document review, pulling key terms out of leases or revenue contracts for the auditor to evaluate
• Anomaly scoring that ranks transactions by how far they sit from the norm

The limit is the same in every case. A tool can tell you an item is unusual. It cannot tell you whether it is wrong, and it cannot replace the auditor's judgment about what the evidence supports. Independence also needs thought when a firm helps an attest client build its own AI monitoring. Our post on AI and data analytics for audit teams goes further.

Advisory and client communication

This is where generative tools save the most time per hour of use, because so much advisory work is writing.

• Turning a technical conclusion into a plain-English client email
• Drafting a memo structure from your own bullet points
• Summarizing a long document into the questions worth asking the client
• Producing the strongest counterargument to your own position before a partner does

None of this requires the tool to know accounting. It requires the tool to write, while the accountant supplies the substance and checks the result.

Forecasting and analysis

Forecasting tools can fit trends and seasonality to historical data quickly, and they are useful for a first-pass projection. They are weak at turning points, because the data that would signal a change in direction usually is not in the history yet. Treat a model's forecast as a baseline that the accountant adjusts with what they know about the business, not as the forecast itself.

What AI Does Not Do Well

A short list, and every item on it has caused a real problem somewhere.

It does not know when it is wrong. A language model gives a wrong answer in the same confident tone as a right one. There is no built-in signal of doubt, which is why verification cannot be optional.

It does not reason from facts it was not given. If the relevant fact is not in the prompt or the data, the tool fills the gap with what is typical. In accounting, the atypical item is usually the one that matters.

It does not stay current on its own. A model trained on data up to a certain date does not know about later legislation, guidance or rate changes unless it is connected to a current source. Tax law that changed after the training cutoff is simply missing. That matters a great deal in a year like this one, after the 2025 law changes made many individual provisions permanent.

It does not take responsibility. The signature on the return, the opinion or the financial statements belongs to a person. No professional standard lets a practitioner point at a tool.

It does not handle edge cases well. Classifiers are trained on the common case. The unusual transaction, the one-off contract or the related-party payment is exactly where they are least reliable and where the stakes are highest.

The Risk Side: Confidentiality, Security and Professional Duties

Adoption decisions often get made on capability alone. The risk side should come first, because several of these obligations carry legal consequences.

Tax return information

For tax practitioners, the sharpest constraint is the federal restriction on disclosure and use of tax return information. Internal Revenue Code section 7216 makes knowing or reckless unauthorized disclosure or use a misdemeanor, with a fine of up to $1,000 (up to $100,000 where the disclosure relates to identity theft) and up to a year in prison. Section 6713 adds a civil penalty of $250 per disclosure or use, capped at $10,000 a year, with higher amounts in identity-theft cases. The regulations set out when client consent is needed and the form that consent must take.

Entering client tax information into a tool run by an outside party may be a disclosure, depending on the arrangement. Many firms decide the simplest control is to keep identifiable return information out of AI tools entirely and work from de-identified input.

Data security obligations

The FTC treats tax preparation firms as financial institutions under its Safeguards Rule. The rule requires a written information security program, a designated qualified individual, a written risk assessment, multi-factor authentication and encryption of customer information, among other elements. Since May 2024 it has also required notifying the FTC, within 30 days of discovery, of a security event involving the unencrypted information of at least 500 consumers. The IRS publishes a template for small practices, Publication 5708, to help build a written information security plan.

Any AI tool that touches client data belongs inside that program: who may use it, what data may go in, how the vendor stores and retains it, and whether the vendor trains its models on your inputs.

Professional standards

The existing standards already cover AI, even though they were not written with it in mind. Competence means understanding a tool well enough to evaluate what it produces. Due diligence, including the diligence rules for practice before the IRS in Circular 230, means checking work before relying on it. Confidentiality rules apply to client information wherever it is sent.

Our post on writing an AI governance policy for an accounting firm turns these duties into a working policy with twelve sections.

The Control Framework That Makes AI Usable

Firms that get real value from AI share a small number of habits.

Match the review to the failure mode. Extracted figures get checked against source documents. Generated text gets read for substance, not just style. Citations get opened and read. Classifier output gets reviewed on exceptions and low-confidence items, plus a sample of the high-confidence ones.

Keep a human sign-off on anything relied upon. Drafts can be machine-made. Conclusions, positions and anything sent to a client or regulator need a named reviewer.

Document the use. If a tool produced part of a workpaper, the file should say so, along with what was checked. A reviewer, a peer reviewer or an examiner should be able to see what the machine did and what the person did.

Approve tools deliberately. A short list of approved tools, each with a defined data tier, beats a blanket ban that people work around and a free-for-all that nobody can audit.

Start where errors are cheap. Internal drafts, summaries and prior-year comparisons are a far better first project than anything client-facing.

What Changes for Your Job

The honest summary is that AI moves time from production to review.

Tasks that consisted mainly of moving, keying or matching data are shrinking. Tasks that consist of deciding (whether a position holds, whether an anomaly is an error, what the client should do next) are not, and the volume of decisions per person tends to rise because the production work in front of them gets faster.

That shift has three practical effects for an individual accountant:

1. Review skill becomes a core skill earlier in a career. Staff who once spent two years preparing before they reviewed anything will be reviewing machine output from the start. Knowing what a wrong answer looks like matters more than knowing how to produce a right one by hand.
2. Explaining becomes more valuable. When the numbers arrive faster, the remaining bottleneck is the conversation with the client about what they mean.
3. Comfort with data becomes a baseline expectation. Not programming, but structuring data, questioning it and spotting when a result does not make sense.

Firms are also changing how they bill and staff work as production time falls, which is a separate conversation about pricing models and capacity.

Where to Start

A practical first ninety days looks like this:

• Pick one internal, low-risk use case (meeting summaries, first-draft client emails from your own notes, or a prior-year comparison) and use it consistently.
• Write down the review step you apply to it, and apply it every time.
• Read your firm's policy on data entry into AI tools. If there is not one, that is the first project, not the second.
• Learn the vocabulary: prompts, retrieval, extraction confidence, hallucination, data retention. You need it to evaluate any vendor claim.

If you want a structured grounding before you start choosing tools, the AI Essentials for Accountants Certificate Program is the foundation-level option among our AI certificates, and the full set is on the AI courses for accountants and CPAs page.

Build the Foundation First

AI in accounting rewards the practitioner who understands both what a tool can do and how it fails. The gap between firms that use these tools well and those that do not is mostly a gap in that understanding, not in software budgets. The AI Essentials for Accountants Certificate Program is a sensible place to build it, and our accounting productivity FAQs answer common questions about the wider set of productivity skills.

Frequently Asked Questions

How is artificial intelligence used in accounting today?

The most common uses are transaction coding suggestions, bank and subledger matching, extraction of data from invoices, receipts and tax documents, anomaly detection in audit, and drafting of client communications and memos. Research assistance is growing, especially with tools that search a real authority database and cite their sources. In every case the pattern is the same: the tool handles volume and first drafts, and a person reviews the output and makes the decisions that carry professional responsibility.

Is AI accurate enough to rely on for accounting work?

Accurate enough to save time, not accurate enough to rely on without review. Extraction tools are highly accurate on clean, common documents but make plausible-looking errors on the rest. Language models can produce confident answers that are simply wrong, including invented citations. The right approach is to match the review to the failure mode: check extracted figures against source, read every citation, and review classifier exceptions plus a sample of the items it was confident about.

Can I put client information into an AI tool?

Only within your firm's policy and the law. For tax practitioners, section 7216 restricts disclosure and use of tax return information and carries criminal and civil penalties, and the regulations set the consent requirements. The FTC Safeguards Rule also requires a written security program that covers how client data is handled. Many firms prohibit entering identifiable client data into AI tools and work from de-identified input, which keeps most of the benefit with far less risk.

What are the biggest risks of AI in accounting?

Four stand out. Confidentiality breaches when client data goes to a vendor without proper terms or consent. Unverified output, especially invented or misread citations, ending up in a workpaper or client advice. Out-of-date answers when a model does not know about recent law changes. And quiet classification errors at the edges, where unusual transactions get coded like typical ones. Each has a known control, but only if someone applies it consistently.

Does AI change what accounting students and new staff should learn?

Yes, at the margin. Technical accounting and tax knowledge matter as much as ever, because they are how you recognize a wrong answer. What rises in importance is review skill, data literacy, and the ability to explain results to clients in plain language. New staff will often review machine output early in their careers, so learning what errors look like, and how to check them efficiently, is a practical priority.

Where should a small firm start with AI?

Start with an internal, low-risk task where an error is cheap to catch, such as first-draft emails from your own notes, meeting summaries, or prior-year comparisons. Write down the review step and apply it every time. Before any client data goes near a tool, put a short policy in place covering approved tools, what may be entered, and how output is verified. Then expand to bookkeeping or tax workflows once the habits are established.