Generative AI in Accounting: Practical Uses and Accuracy Limits

By Jane Ha · Updated Oct 11, 2026

Generative AI in accounting means using language models to draft and summarize text-heavy accounting work (technical memos, policy research, variance commentary, formula help, documentation) while a qualified person verifies every figure, citation and conclusion before it's used. The models write fluently and fast. They don't calculate reliably, they don't know your ledger, and they can state a wrong standard with total confidence. So the useful question isn't "can it do this task?" but "what check does its output need?"

This guide is for CFOs, controllers, finance directors and accounting managers. It covers the tasks where generative AI earns its keep, a table of task, suitability and required check, a rollout sequence, and the failure modes that matter in a profession where numbers get audited. For the broader view of AI across the function, start with AI in finance.

Key Facts

• ICAEW's generative AI guide says models can "make up quotes and references" and "contradict themselves when challenged," and warns that people are prone to automation bias, relying too heavily on technology's output (ICAEW, Generative AI guide: risks and limitations).
• The AICPA says the core duties of competence, due care, confidentiality and independence still apply when AI is used, and that responsibility for the final work product stays with the professional, whether the tool was built in-house or by a third party (Journal of Accountancy, AICPA Professional Ethics Division, August 2026).
• In PCAOB staff outreach to larger audit firms (July 2024), firms described using generative AI mainly for administrative and research work, such as initial drafts of memos and research into internal accounting and auditing guidance, and said human review of the output remains essential (PCAOB staff Spotlight, July 2024).
• In a 2025 study of GPT-4o, o1, DeepSeek-V3 and DeepSeek-R1 on arithmetic, algebra and number theory problems, procedural slips were the most frequent error type, and accuracy differed sharply between models (Mathematical Computation and Reasoning Errors by Large Language Models, arXiv).

What Generative AI Is Good At in Accounting, and What It Isn't

Generative AI predicts likely text. That's a strength for work that's mostly language: summarizing a 40-page lease, restating a policy in plain English, drafting the first version of a memo from your bullet points. It's a weakness for work that must be exactly right: a sum, a rate, a paragraph reference, a threshold.

Three limits drive almost every control in this article.

It invents. The behavior is called AI hallucination. The PCAOB's staff report notes that audit firms recognize generative output may be false or misleading, or skewed and biased, and some firms are building tools that document the source data used so a reviewer can trace it. A plausible-sounding citation to a standard paragraph that doesn't exist is the classic accounting version.

It's unreliable on arithmetic. A language model doesn't run a calculator unless you give it one. The arXiv study above found procedural slips were the most common error across the models tested. The reasoning-focused model in that study did far better than the others, so the gap is narrowing, but "better" isn't "ledger-grade." Anything the model computes should be recomputed in your spreadsheet or system.

It doesn't know your facts. Unless you supply the trial balance, the contract or the policy, the model fills gaps with generic assumptions. Grounding it in your documents, for example through retrieval-augmented generation, reduces this but doesn't remove it.

Task by Task: Suitability and Required Check

This table is the working core of the article. "Suitability" is a judgment about risk, not a promise of quality: it reflects how much damage an undetected error could do and how easy the output is to verify.

The pattern: AI is a drafting and explaining layer. Anything that produces a number, a citation or a conclusion goes through a verification step that doesn't depend on the model.

Real Examples by Accounting Function

Technical accounting and policy. A controller facing a new contract type asks the model to outline the questions a revenue recognition analysis should answer and to draft the background section from the contract summary. The controller then writes the conclusion, reads the cited guidance in the standard itself, and fixes anything the draft got wrong. The model saved the blank-page time; it didn't do the judgment.

Financial planning and close commentary. At month-end, the analyst pastes the variance table and a few lines on drivers, and the model drafts the narrative. The risk here is subtle: the model may restate a number slightly wrong or invent a plausible reason for a variance. Every figure in the paragraph gets tied back to the schedule, and every cause gets confirmed with the budget owner. The related finance workflows are covered in AI finance use cases.

Accounts payable and receivable. Models help draft vendor follow-ups, summarize a long email chain on a disputed invoice and explain why a payment didn't match. Matching itself belongs in deterministic tools, which is where AI agents for finance and rule-based automation fit.

Audit and assurance. The PCAOB's staff report describes firms using generative AI for initial drafts of audit-related memos and presentations and for research into internal guidance, and it lists areas firms see as future candidates, such as checking audit documentation for completeness against documentation requirements. In the same report, some firms said they don't permit generative AI for audit or attest procedures because of data privacy and reliability concerns. If you're a company preparing for an audit, ask your auditor about its policy, because it affects what you can hand over and how.

FP&A and Excel. Asking a model to write a lookup formula, explain an inherited macro or convert a nested IF into something readable is among the safest uses, because you can test the result immediately. The risk is shipping an untested formula into a model that feeds the board pack.

Confidentiality: What Goes Into the Prompt

The prompt is a disclosure. The AICPA's ethics guidance notes that public AI tools may use user inputs to train future models, so entered data could resurface elsewhere, and that members in public practice should consider whether information is confidential and whether consent is required before entering client data, under the Confidential Client Information Rule. A Journal of Accountancy article by a CPA risk control director recommends firms bar confidential and proprietary client and firm information from generative AI tools, and that staff treat inputs the way they would a post on a public site (Journal of Accountancy, October 2023).

For a corporate finance team the equivalent concerns are unreleased results, M&A information, payroll data and customer contracts. The practical rule: decide in writing which data classes may go into which approved tools, and make the approved tool the easy path. People paste into whatever's open when the approved option is slow.

How to Introduce Generative AI in Your Accounting Team

1. Start with a task inventory. List recurring tasks that are mostly writing or explaining: memos, commentary, emails, documentation, formula help. Rank by volume and by how easy the output is to check.
2. Choose approved tools and write the data rule. Name which tools are allowed, which data classes can go in, and who decides exceptions. Review the vendor's data terms with legal or IT. If your data isn't organized for this, work through finance data readiness for AI first.
3. Define the check for each task. Use the table above as a template. For each use, write who verifies what: numbers tied to source, citations found in the standard, conclusions owned by a named person.
4. Keep the person accountable. The AICPA is explicit that responsibility stays with the professional. Build the human-in-the-loop step into the workflow and the sign-off, not just the policy.
5. Pilot on one process for a close cycle. Pick variance commentary or memo drafting. Measure drafting time saved, corrections found in review and any error that reached a reviewer undetected by the preparer.
6. Decide how to document AI use. Some firms ask staff to disclose when AI contributed to work so reviewers know to look harder. Decide whether your workpapers record the tool, the inputs and the review.
7. Train for skepticism. ICAEW recommends professional scepticism as the main safeguard and suggests repeating the same question to see whether answers vary. Teach reviewers to treat output like work from a confident but inexperienced colleague.
8. Review quarterly. Tools, vendor terms and professional guidance change. Revisit approved tools and checks on a schedule.

Common Pitfalls

• Trusting a citation because it looks right. Fabricated references are a documented failure. Find the paragraph in the standard before it goes in the memo.
• Letting the model do the sums. Even when it shows its working, recompute. A fluent explanation of a wrong total is still wrong.
• Automation bias in review. Reviewers read AI-drafted text less critically because it's well written. ICAEW names automation bias directly. Assign the reviewer to verify, not to proofread.
• Documenting work that wasn't done. A model asked to draft a walkthrough narrative will describe a plausible procedure. If it differs from what the team did, the workpaper is misleading.
• Unapproved tools. Staff using personal accounts for client or company data is a confidentiality breach waiting to happen. Provide a sanctioned option.
• Skipping the data rule. Policies that say "be careful" don't work. Specify classes of data and approved tools.
• Hiding AI use from reviewers. If the reviewer doesn't know a section was machine-drafted, they can't calibrate their scrutiny.

Where to Go Next

If you're deciding which team members own verification and prompt standards, read AI and finance team roles. If you'd rather map the whole function before picking a pilot, AI finance use cases sorts them by value and risk. And if you only do one thing this quarter, write the one-page rule: approved tools, allowed data and the named person who verifies the numbers. Once the rule exists, give the "draft, verify, sign off" routine a home as an assigned task with an approval step, in a process tool such as Rework's BPM, rather than leaving it as an informal habit.