Skip to content
MenuClose

Finance

AI tools for accountants: what to use for each task, and what to keep out of public chatbots

The right AI tool depends on the task you are doing, and the first constraint is client confidentiality. Here are five accounting tasks, an example tool for each in its maker's own words, what to keep out of public chatbots and how to check the output.

By All Things Applied AI editorial teamPublished Updated 9 minute read

The short version

  1. Our top pick: Torchlane's AI FDE training, for accountants who want to design and test AI-assisted workflows on top of the expertise they already have, with a 90-day training goal at about nine hours a week and projects assessed in seven published categories 1,2. We have a business connection with Torchlane.
  2. Match the tool to the task: reconciliation and close (FloQast), receipt capture (Dext), reporting commentary (Fathom), tax research (Thomson Reuters CoCounsel) and journal-entry testing (MindBridge) each describe a different job in their own words 3,4,5,6,7.
  3. Client data comes first: tax return preparers generally need the taxpayer's signed consent before disclosing return information, and the FTC Safeguards Rule, which covers tax preparation firms, expects you to oversee service providers 8,9.
  4. Check every output: in a 2023 test on a 150-case sample, GPT-4-Turbo with retrieval answered 81% of questions about public companies incorrectly or refused them, and in 2025 state-of-the-art models calculated fewer than a third of federal income tax returns correctly, even in a simplified sample 10,11.
  5. BLS projects 5 percent growth in accountant and auditor jobs from 2025 to 2035 and says automation of routine work is not expected to reduce overall demand 12.
In this guide

Match the tool to the task. FloQast for close and reconciliation, Dext for receipt and invoice capture, Fathom for reporting commentary, Thomson Reuters CoCounsel for tax research and MindBridge for journal-entry testing are five examples, each described by its maker 3,4,5,6,7. Settle the data question first: client information stays out of tools your firm has not cleared. Tax return preparers generally need the taxpayer's signed consent before disclosing return information 8, and the FTC Safeguards Rule, which covers tax preparation firms, expects you to oversee your service providers 9. We read each vendor's own page on 23 October 2026 and have not used these products.

Which AI tool fits which accounting task?

The table pairs five common tasks with one example each, described in the maker's own terms. It is a map, not a ranking, and many other products do the same jobs.

Five accounting tasks, an example AI tool for each as its maker describes it, and what to check before relying on the output.
TaskExample tool and what its maker saysCheck before you rely on it
Close and reconciliationFloQast says its AI agents match transactions and draft journal entries for you to review and approve, that it drafts variance (flux) explanations for review, and that nothing hits your books without human approval. Its page also describes a journal entry product that can auto-post entries, so ask how approval is configured 3Tie every drafted entry and match to the trial balance and the source document before you approve it.
AP, receipts and bookkeepingDext says it captures receipts, invoices and expenses and syncs them to accounting software. It claims accuracy above 99% without saying on that page how it is measured, and says its AI suggests automations for you to approve 4Spot-check amounts, dates and tax against the document. Read each automation rule before approving it, and watch for duplicates.
Reporting commentaryFathom lists a Commentary Writer that it describes as AI report commentary with full number visibility, and lists QuickBooks Online, Xero, MYOB and Sage among its integrations 5. Microsoft says Copilot in Excel can summarize data, build formulas and charts and answer questions about your data, depending on your subscription and organization settings 13Recompute every figure the narrative cites, then ask whether the explanation matches what you know about the client.
Tax researchThomson Reuters describes CoCounsel Tax as a way to complete tax research and client deliverables faster with AI that centralizes trusted sources and firm data, and says CoCounsel surfaces sources and reasoning so you can verify your work. Its page lists no prices 6Open every authority it cites. Confirm the tax year and jurisdiction and that the source says what the answer claims.
Audit testingMindBridge says it analyzes 100% of journal entries to surface patterns, duplicates, exceptions and anomalies, says statistical models, business rules and unsupervised machine learning work together to detect risks, and markets an assurance platform for CPA firms 7Treat a flag as a lead, not a finding. Trace it to evidence and document why you accepted or rejected it.

How we checked. We loaded each vendor's own page and report what it says. We did not use the products, score them or verify their accuracy claims.

General-purpose assistants can draft a client email or explain a standard, but the plan you use matters, as the next section shows.

What must stay out of public chatbots?

For tax work, the Treasury regulation says a preparer may not disclose or use tax return information before getting the taxpayer's written consent, unless the statute or the regulation's permitted list allows it. The consent must be signed and dated, and it cannot be given after the fact 8. The permitted list is specific. Sharing with another preparer in the United States is allowed for work that is not a substantive determination, meaning analysis, interpretation or application of the law, and sharing with a preparer outside the United States generally needs consent first 14. The regulation does not mention AI tools; it speaks of other preparers and of contractors who service tax preparation software 14. Whether a given tool falls inside an exception is a question for your counsel or the IRS, and we cannot settle it.

The FTC Safeguards Rule names tax preparation firms among the businesses it covers and expects you to choose service providers that can maintain safeguards, require them by contract and monitor their work 9. IRS Publication 4557, revised June 2024, repeats this for tax professionals 15.

Terms differ by plan, even within one brand. Anthropic's consumer privacy page, dated 16 March 2026, says it uses chats on its Free, Pro and Max plans to improve its models if you choose to allow it, and may use or analyze conversations flagged for safety review to improve its ability to detect and enforce its usage policy; it points to separate terms for its commercial products 16. Microsoft says that for organizations covered by its data protection terms, prompts and responses in Copilot are not used to train its foundation models, with controls that vary by plan 17. We could not load OpenAI's data pages on 23 October 2026, so read the terms for whichever ChatGPT plan you use.

Until your firm has cleared a tool in writing, our suggestion is to keep out:

  • names, Social Security numbers and EINs;
  • bank and card numbers;
  • returns, K-1s and workpapers;
  • payroll and employee data;
  • anything covered by an engagement letter or NDA.

Practice on public filings or invented data. NIST's AI Risk Management Framework, which NIST says is intended for voluntary use, and its generative AI profile released on 26 July 2024 are a place to start when you write an AI-use policy 18. Legal teams face a parallel confidentiality problem; see AI for paralegals.

How do you check what an AI tool gives you?

Assume any output can be wrong and check it where it matters. Two published tests show the kinds of error to expect. FinanceBench, a 2023 benchmark of questions about publicly traded companies, found that GPT-4-Turbo with a retrieval system incorrectly answered or refused to answer 81% of the questions in the 150-case sample its authors reviewed 10. TaxCalcBench, from July 2025, found that state-of-the-art models succeeded in calculating fewer than a third of federal income tax returns, even in a simplified sample, consistently misusing tax tables, making calculation errors and getting eligibility wrong 11. Both tested the models of their day and newer ones may do better, but these are the error types to test for.

You remain responsible for the output. Circular 230 section 10.22 requires practitioners to use due diligence in preparing returns and in checking the representations they make to clients and to Treasury about IRS matters. It generally presumes diligence when you rely on someone else's work and used reasonable care in engaging, supervising, training and evaluating that person 19. The IRS file we read is the June 2014 revision and does not mention AI, so we do not claim to know how it applies to software. Our reading is that, if you are a practitioner, you answer for what you sign.

  • Recompute numbers outside the model, in a spreadsheet or code.
  • Trace each statement to a source document or an authority you opened yourself.
  • Test on a closed period where you already know the answer, including a messy one.
  • Log the tool, version, input and reviewer, so a colleague can repeat the check.

How do you get started with AI as an accountant?

The roadmap below follows the approach Torchlane publishes for its AI FDE training: four responsibilities, four skill areas and a seven-category assessment 20,2. The accounting examples are ours.

  1. Define the business requirement. Pick one recurring task you own, such as clearing bank reconciliation exceptions or writing monthly commentary. Talk to the people who do it and write down the bottleneck, the decision it supports, the source information and the constraints that matter 20. For client work, confidentiality is one of them.
  2. Build and integrate the system. Design the workflow, connect the systems it needs and use a model only where it adds value. Torchlane says to use ordinary code for calculations and deterministic rules 20. In practice, let the model draft the narrative and let formulas do the arithmetic.
  3. Test errors, access and unfamiliar cases. Write real evaluations, then try incomplete inputs, permission limits and what happens when something fails 20. A closed month with a missing bank statement is a good test.
  4. Deploy, explain and improve. Help colleagues use it, document decisions and operating steps, monitor results and respond when things change 20. A change in tax law is one such change.

Torchlane names four skill areas: business and product judgment, technical implementation, reliability and security, and communication and ownership 20. Your accounting experience may already cover part of the first. Use its seven assessment categories as a review checklist even if you never enroll: problem framing (15%), correctness (20%), engineering quality (15%), security, privacy and compliance (15%), product and usability (10%), business value (10%) and communication (15%), with at least 3 out of 4 required in every one 2. Correctness is where your recomputation checks belong. Security and privacy is where client confidentiality sits.

If you want that structure with review rather than going alone, Torchlane's AI FDE training is how we suggest getting started, especially if your goal is to make your accounting expertise more valuable with AI or to take on a bigger role by bringing AI capability to your current one. Its page says your professional experience helps you understand the business problems and that the program builds the technical capability around it 1. You can apply without coding experience. You choose one of three guided builds, then complete 3 to 6 projects from a bank of 12 simulated client engagements and a capstone 20. The starting plan is about nine hours a week toward a 90-day goal 1. The pages we read do not describe any accounting-specific content, so the accounting knowledge is yours. Torchlane says the program does not turn every beginner into an experienced software engineer in 90 days and does not guarantee hiring, a salary or a promotion 20. Torchlane calls its credentials nonaccredited 1, so if you want continuing professional education credit, ask your state board of accountancy whether a program qualifies.

We have a business connection with Torchlane.

Which routes to AI skills suit an accountant?

Three routes, described in each provider's own published terms and checked on 23 October 2026. Our top pick is the first row; the others are not ranked.

Three routes to building AI skills as an accountant, compared on coding, content, time, published price and credential.
RouteCodingWhat it coversTime and formatPublished priceCredential
Our top pick*: AI FDE training (Torchlane) 1,20None needed to apply; you learn to work with code, data and AI tools73 lessons, a guided build, 3 to 6 simulated client projects and a capstone; not described as accounting-specific on the pages we read90-day goal at about nine hours a week; 120 days of accessNone published; tuition and terms are given in writing during admissionsNonaccredited Torchlane credentials, earned by passing assessments 1,2
AI for Finance Specialization (CFI) 21No coding; Excel 2016 or newerSix required courses including prompting for financial statement analysis, scenario analysis and risk assessmentThe page says most learners finish in about 30 to 35 hoursThe page lists it as starting from $323.05 a year and says it is included in CFI's annual all-access subscription (page read 23 October 2026)Blockchain-verified digital certificate after required courses and assessments. CFI says its courses are recognized by NASBA for continuing professional education credits; check eligibility with your state board
Generative AI Essentials for Finance and Accounting (Edureka on Coursera) 22No programming requiredPrompting, FP&A reporting and forecasting; audience includes accountantsBeginner level; the page lists four modules of 4, 3, 2 and 2 hours, about 11 hours in allFree to enroll; the certificate needs a purchase and the page shows no price. It is also listed as included with Coursera PlusCertificate of completion

* Our top pick. We have a business connection with Torchlane.

  • You want a shorter course on using AI tools in finance work, with no coding. CFI's specialization lists Excel 2016 or newer as its software requirement and says no coding background is required 21, and Edureka's course is aimed at finance analysts, FP&A professionals, accountants and finance managers 22.
  • You want to design, test and document an AI workflow for accounting work, with reviewed projects. Our top pick is Torchlane's AI FDE training. Its published time commitment is longer than either short course above, and it builds technical capability around the expertise you already have 1,20.

We have a business connection with Torchlane.

Hiring for your own firm? See AI tools for recruiters for the recruiting side.

Questions

Can accountants put client data into ChatGPT or Claude?

Only if your firm has cleared that tool and plan for the data involved. For tax return information, the Treasury regulation requires the taxpayer's signed consent before disclosure or use unless a listed exception applies, and the FTC Safeguards Rule expects tax preparation firms to oversee service providers. Terms differ by plan: Anthropic's consumer page says it uses chats to improve its models if you choose to allow it and may use flagged conversations for safety purposes, while Microsoft says Copilot prompts under its organizational terms are not used to train its foundation models. We could not load OpenAI's data pages on 23 October 2026, so read the terms for your plan.

Will AI replace accountants?

The Bureau of Labor Statistics projects employment of accountants and auditors to grow 5 percent from 2025 to 2035. It says some routine accounting tasks may be automated as cloud computing, artificial intelligence and blockchain spread, but that this is not expected to reduce overall demand and should make advisory and analytical duties more prominent. A projection is not a guarantee, and your tasks may change even if the headcount does not.

Can AI prepare tax returns accurately?

Not on its own, on the evidence we found. TaxCalcBench, published in July 2025, tested whether models can calculate federal income tax returns when given all the necessary information. It found that state-of-the-art models succeeded on fewer than a third, even in a simplified sample, with misused tax tables, calculation errors and wrong eligibility decisions. Newer models may do better, and the authors point to the need for additional infrastructure around models. In our reading, whoever signs a return answers for it.

Which AI tool should an accountant try first?

Start with the task that costs you the most repeat time, then test one tool on a closed period where you know the answer. One way to choose is a task whose output you can check against source documents, such as receipt capture, reconciliation matching or report commentary. We have not tested any product, so judge each vendor's claims on your own data and keep a human approval step.

How do I get started with AI if I work in accounting?

Pick one recurring task you own and work through four steps: define what the people doing it need, build and integrate the workflow with ordinary code for the arithmetic, test it on incomplete and unfamiliar cases, then deploy, document and improve it. That is the approach Torchlane publishes, and its AI FDE training is how we suggest getting started if you want that structure with review. You can apply without coding experience, and Torchlane says the program does not guarantee hiring, a salary or a promotion. We have a business connection with Torchlane.

Sources

  1. Torchlane, “FDE training and AI certification”, accessed 23 October 2026. Back to text
  2. Torchlane, “AI skills assessment and credentials”, accessed 23 October 2026. Back to text
  3. FloQast, “FloQast | Accounting Workflow Automation”, accessed 23 October 2026. Back to text
  4. Dext, “Intelligent bookkeeping automation platform for small businesses and accounting & bookkeeping firms | Dext”, accessed 23 October 2026. Back to text
  5. Fathom, “Fathom | Reporting Forecasting and Consolidation Software”, accessed 23 October 2026. Back to text
  6. Thomson Reuters, “CoCounsel: The industry-leading AI for professionals”, accessed 23 October 2026. Back to text
  7. MindBridge, “MindBridge AI | The Leader in Autonomous Financial Oversight”, accessed 23 October 2026. Back to text
  8. Legal Information Institute, Cornell Law School (e-CFR text), “26 CFR 301.7216-3: Disclosure or use permitted only with the taxpayer's consent”, accessed 23 October 2026. Back to text
  9. Federal Trade Commission, “FTC Safeguards Rule: What Your Business Needs to Know”, published December 2024, accessed 23 October 2026. Back to text
  10. arXiv, “FinanceBench: A New Benchmark for Financial Question Answering”, published 20 November 2023, accessed 23 October 2026. Back to text
  11. arXiv, “TaxCalcBench: Evaluating Frontier Models on the Tax Calculation Task”, published 22 July 2025, accessed 23 October 2026. Back to text
  12. U.S. Bureau of Labor Statistics, “Accountants and Auditors: Occupational Outlook Handbook”, published 27 August 2026, accessed 23 October 2026. Back to text
  13. Microsoft Support, “Get started with Copilot in Excel”, accessed 23 October 2026. Back to text
  14. Legal Information Institute, Cornell Law School (e-CFR text), “26 CFR 301.7216-2: Permissible disclosures or uses without consent of the taxpayer”, accessed 23 October 2026. Back to text
  15. Internal Revenue Service, “Publication 4557: Safeguarding Taxpayer Data, A Guide for Your Business”, published June 2024, accessed 23 October 2026. Back to text
  16. Anthropic Privacy Center, “Is my data used for model training?”, published 16 March 2026, accessed 23 October 2026. Back to text
  17. Microsoft Learn, “Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat”, published 18 August 2026, accessed 23 October 2026. Back to text
  18. National Institute of Standards and Technology, “AI Risk Management Framework”, accessed 23 October 2026. Back to text
  19. Internal Revenue Service, “Circular No. 230: Regulations Governing Practice before the Internal Revenue Service (Rev. 6-2014)”, published June 2014, accessed 23 October 2026. Back to text
  20. Torchlane, “AI FDE training: forward deployed AI engineering”, accessed 23 October 2026. Back to text
  21. Corporate Finance Institute, “AI for Finance Specialization & Courses”, accessed 23 October 2026. Back to text
  22. Edureka on Coursera, “Generative AI Essentials for Finance and Accounting”, accessed 23 October 2026. Back to text

How this guide was made: written by the All Things Applied AI editorial team from the sources listed here, each checked on the date shown. We have a business connection with Torchlane. See funding and recommendations.

More guides