AI for accountants: a practical guide
AI is now built into QuickBooks, Xero, Sage and the big tax and audit platforms. Here is what it does well, where it gets numbers wrong, and what the confidentiality rules mean for client data.

In short
- Nearly every accounting firm now uses some AI, but only about one in five has a written policy for it, according to Karbon's 2026 survey.
- The main accounting platforms have added AI agents for reconciliation, data entry, return preparation and audit testing. Most of the time savings come from the dull work, not from judgement.
- AI is weakest exactly where accountants need it most: adding up across large tables. In one 2026 benchmark, top models were right 94% of the time on simple look-ups but only 33% on complex aggregation.
- In the US, sharing tax return information with an outside AI service can need the client's consent under section 7216. Use approved tools with contracts that forbid training on your data.
- Deloitte Australia refunded part of a A$440,000 government fee in 2025 after a report it produced with AI help contained a fabricated court quote and non-existent references.
- Regulators apply the rules you already follow: due care, confidentiality and supervision. The FRC and PCAOB expect firms to test and control the AI tools they use in audits.
Accounting was always going to be an early target for AI. Much of the work is structured, repetitive and already digital. In 2026 the AI is no longer a separate chatbot: it is built into the ledger, the tax software and the audit platform. This guide covers what it does well, where it fails, and the rules on client data.
Everyone uses it, few have rules
Surveys agree that AI use in accounting firms is close to universal. Karbon’s 2026 survey of nearly 600 firms found that 98% use AI, most of them daily, saving around an hour per employee per day. Intuit’s survey of 725 US accountants found 88% had used AI for at least one client service. Thomson Reuters found 81% of tax and audit professionals use it regularly.
Governance has not kept up. Only 21% of the firms in Karbon’s survey had a documented AI policy, and more than a third of the professionals in the Thomson Reuters study used AI tools their firm had not approved.
Almost every firm uses AI; few have written rules
Share of accounting firms, Karbon survey of nearly 600 firms, 2026
Source: Karbon, State of AI in Accounting 2026
The tools
| Tool | Made by | What the AI does |
|---|---|---|
| QuickBooks and Intuit Accountant Suite | Intuit | Agents for bookkeeping, payroll, payments and client work; a platform for firms managing many clients |
| JAX (Just Ask Xero) | Xero | Bank reconciliation, data entry and getting paid, with controls meant to reduce errors |
| Sage Copilot | Sage | Assistant and agents across Sage’s accounting products, including tools for UK Making Tax Digital |
| Copilot in Excel | Microsoft | Analysis, formulas and charts from the Copilot pane |
| CoCounsel Tax and Audit | Thomson Reuters | Tax research, US 1040 return preparation and audit document analysis |
| CCH Axcess Expert AI | Wolters Kluwer | AI inside tax, audit and workflow software, including reading tax forms |
The Big Four have built their own platforms, such as KPMG Workbench and agentic AI in EY’s audit tool Canvas. If you work at one, you will use those rather than public chatbots.
Tools change fast. Microsoft retired the =COPILOT() function in Excel on 14 September 2026, after a preview that ran only for early-access users. Spreadsheets that used it show errors when they recalculate. Don’t build client work on features marked preview or beta.
Where AI gets numbers wrong
Language models are built to predict words, not to add up. They are good at reading documents, drafting emails and explaining a rule, and much less reliable at arithmetic across a big table.
A 2026 study by researchers in Zurich tested leading models on questions about real financial spreadsheets. The best model was wrong about once in every six questions. Models did well on simple look-ups but badly when they had to combine many cells.
AI is good at finding a number, poor at adding them up
Accuracy of the best-performing models on questions about financial spreadsheets, by task type
Source: FinSheet-Bench (arXiv, March 2026)
The same problem applies to citations. In 2025 Deloitte Australia agreed to refund part of its A$440,000 fee to the Department of Employment and Workplace Relations after a report it had produced with the help of AI turned out to contain a made-up quote from a court judgment and references to papers that do not exist.
Practical rules follow from this:
- Let AI find, you compute. Use it to locate figures, summarise documents and draft formulas. Do the totals in the spreadsheet or the ledger, where you can trace them.
- Tie out everything. Reconcile AI-prepared numbers to source records, just as you would for a junior’s work.
- Check every reference to a standard, a ruling or a paper before it goes into a report.
Client confidentiality
- United States, tax. Section 7216 of the Internal Revenue Code makes it an offence for preparers to use or disclose tax return information without authorisation, and the rules set out when the client’s written consent is needed. The IRS has not issued AI-specific guidance, so the careful reading is to treat sending return data to an outside AI service like any other disclosure: use a vendor under contract, or get consent.
- United States, CPAs. The AICPA Code has no AI-specific rule, but its due care and confidential client information rules apply. The AICPA has published a generative AI policy template for small firms.
- United Kingdom. ICAEW’s guidance says not to put confidential or personal data into public AI tools and ties AI use to the five fundamental ethical principles. UK GDPR applies to client personal data.
- Everywhere. Prefer AI built into the software your firm already uses under contract, where client data stays inside the system and is not used to train models.
What regulators expect in audit
Audit regulators have not banned AI, but they expect firms to control it. The UK’s Financial Reporting Council published its first guidance on AI in audit in June 2025. It adds no new requirements but shows how firms should document and test the AI tools they rely on. In the US, PCAOB staff found in 2024 that firms mostly use generative AI for administrative and research tasks and stressed the need for strong supervision. Under Circular 230, US tax practitioners must still exercise due diligence when they rely on others’ work, and that includes a machine’s.
In the EU, the AI Act treats AI that scores people’s creditworthiness as high-risk. Ordinary bookkeeping and tax tools are not high-risk, but firms are expected to support AI literacy among their staff.
Getting started
- Write a one-page policy. Which tools are approved, what data may go in, and who reviews the output. The AICPA template is a good starting point.
- Start with the pain points. Bank feeds, receipt capture, reconciliations and first drafts of client emails are where firms report the biggest savings.
- Keep the controls. Review, sign-off and an audit trail apply to AI-prepared work exactly as to anyone else’s.
- Train the team. Most mistakes come from people who don’t know the tool’s limits, not from the tool itself.
Terms unclear? See our AI glossary, or browse the other guides.
Sources
- Karbon: State of AI in Accounting 2026 report (January 2026)
- Intuit QuickBooks: 2026 Accountant Tech Survey (June 2026)
- Thomson Reuters: Future of Professionals 2026, tax and audit firms (July 2026)
- Intuit: Intuit Accountant Suite, an AI-native platform for accounting firms (October 2025)
- Xero: JAX, Xero's AI financial superagent (September 2025)
- Sage: Sage expands AI agents across finance, HR and operations (April 2026)
- Microsoft Support: COPILOT function (retired 14 September 2026)
- Thomson Reuters: Agentic AI solutions for tax, audit and accounting (November 2025)
- Wolters Kluwer: CCH Axcess Expert AI
- KPMG: KPMG launches multi-agent AI platform Workbench (June 2025)
- EY: Enterprise-scale agentic AI for audit in EY Canvas (April 2026)
- FinSheet-Bench: evaluating LLMs on financial spreadsheets (arXiv, March 2026)
- CFO Dive: Deloitte refunds part of fee over AI errors in Australian government report (October 2025)
- IRS: Section 7216 information center
- 31 CFR 10.22: Diligence as to accuracy (Circular 230)
- AICPA: Small firm generative AI policy template (April 2026)
- ICAEW: Generative AI guide, ethics
- FRC: Guidance on the use of AI in audit (June 2025)
- PCAOB: Staff observations on generative AI in audits and financial reporting (July 2024)
- EU AI Act, Annex III: high-risk AI systems
More guides
- AI for teachers: a practical guide
- AI for lawyers: a practical guide
- AI for doctors: a practical guide
- AI for HR and recruiters: a practical guide
- Unusual jobs for 2027: penguins, puffins, particle physics, and who can actually apply
- AI presentation tools compared: Gamma, Beautiful.ai, Copilot, Gemini and Canva
- AI meeting note-takers compared: Otter, Fireflies, Fathom, Granola, MeetGeek, Notuly and the built-ins
- We gave one meeting to four AI note-takers. Two knew who had to do what
Comments
What do you think? No account needed. Please be kind; comments that attack people are removed.
No comments yet. Be the first.