The short answer
Accountants get the most from AI on the work around the numbers: cleaning exports with Copilot in Excel, which gained a clean data skill in October 2026 and accounting skills rolling out worldwide in November, drafting variance commentary and client emails with Claude or ChatGPT, and summarising leases and contracts in Gemini Notebook. Use your firm's work plans, keep client data out of personal accounts, and check every figure. ICAEW guidance says confidential client information should not be loaded into public generative AI tools.
| Start with | Copilot in Excel for spreadsheets, Claude for writing |
|---|---|
| Best free option | Gemini Notebook Standard for non-client documents |
| Keep human | Judgement calls, sign-off, anything filed |
| Biggest risk | Client data in a personal AI account |
| Rules to know | IESBA confidentiality principle; IRS section 7216 for US preparers |
How accountants are using AI at work in 2026
Most accountants do not need an AI that does accounting. They need help with the hours around it: cleaning exports that arrive in the wrong shape, writing the same explanation for the fifth client this month, reading a 90-page lease to find the three clauses that matter, and turning a variance table into commentary a director will actually read. That is where AI earns its keep, and it is also where the risk sits, because almost everything an accountant touches is confidential.
The tools have moved quickly this year. In October 2026 Microsoft added a clean data skill to Copilot in Excel, and its accounting skills, which reached Microsoft's Frontier programme in October, roll out worldwide in November, so the assistant works inside the workbook rather than in a separate chat. Claude from Anthropic runs Claude Opus 5.5 on Pro and above and is strong at reading long documents and explaining its working. Google renamed NotebookLM to Gemini Notebook on 16 July 2026; it answers only from the sources you upload, with citations. The rule for all of them is the same: use the work version your firm pays for, keep client data out of personal accounts, and check every number.
The 7 accounting tasks worth handing to AI
Each task below names the tool we would start with and why. The matching prompts are further down the page, ready to copy.
1. Clean and reshape messy exports (Copilot in Excel)
Bank exports, payroll reports and trial balances rarely arrive tidy. The clean data skill Microsoft added to Copilot in Excel in October 2026 scans the workbook for duplicates, inconsistent formatting and incorrect data types, then recommends fixes that you review and approve one by one. Your firm needs Microsoft 365 Copilot licences; for your own files, Copilot in Excel also comes with Microsoft 365 Personal, Family (subscription owner only), Premium and Pro. Keep the original sheet untouched and work on a copy.
2. Draft variance and management commentary (Claude or Copilot)
Give the model the numbers and your notes on what drove them, and it writes the first version of the commentary. Claude is better at careful, readable prose; Copilot is more convenient if the figures already sit in Excel. The model must only explain movements you have already understood. It cannot know why marketing spend rose, and it will guess if you let it.
3. Summarise long documents: leases, contracts, standards (Gemini Notebook)
Upload the lease, the loan agreement or your firm's technical manual to a notebook and ask questions. Gemini Notebook answers only from those sources and cites the passage, so you can click through and check. The free Standard plan allows 100 notebooks with up to 50 sources each. On a work or school account, Google says your uploads and questions are not reviewed by humans or used to train its models.
4. Write client emails that explain finance in plain English (ChatGPT or Claude)
Clients ask the same questions every year: why do I owe tax on account, what is this accrual, why is my profit different from my bank balance. Ask the AI to explain your answer at the client's level, then you correct the facts. This is one of the safest and most time-saving uses, as long as you write the technical point yourself and remove names and figures. Our email prompts cover the follow-ups too.
5. Check formulas and audit a workbook (Copilot in Excel or ChatGPT)
Ask the AI to explain what a nested formula does, find hard-coded numbers inside formulas, or list cells that break a pattern. It is a fast second pair of eyes on a model someone else built. ChatGPT's data analysis works well on an anonymised copy; Copilot works on the live file. Our Copilot prompts for Excel go further.
6. Research a technical question, with sources (Perplexity)
Perplexity searches first and numbers its sources, which makes it a good starting point for "what changed in this rule" questions. Treat it as a map to the primary source, never as the source. Ask the question in general terms with no client details, and on Free, Pro or Max turn off AI data retention under Account settings, Preferences, because it is on by default.
7. Notes from client meetings (Fireflies.ai, with consent)
A meeting assistant turns a one-hour planning call into a summary and action list. Fireflies.ai says meeting content is never used to train AI models. Always tell the client the call is being recorded and transcribed, and switch off auto-join so the bot does not arrive at meetings where it is not welcome.
What accountants should keep human
- Judgement calls. Revenue recognition, impairment, going concern and tax positions need a qualified person who owns the decision. AI can list the considerations; it cannot carry the responsibility.
- Anything you sign or file. Returns, accounts, audit opinions and regulatory submissions are reviewed line by line by a person, every time.
- Professional scepticism. ICAEW's guidance warns against over-reliance on generative AI output and against restating as fact answers that may be invented. Every figure an AI produces gets traced back to the ledger.
- Difficult conversations. Telling a client their business is in trouble, or that they have a penalty coming, is a relationship moment. Draft notes with AI if you like; have the conversation yourself.
- Ethics and conflicts. Whether to take on a client, or how to handle a suspected irregularity, stays with partners and your firm's procedures.
A starter AI stack for accountants
| Job | Tool | Plan to use | Why |
|---|---|---|---|
| Spreadsheets and Outlook | Microsoft 365 Copilot | Your firm's licence, or Copilot Chat with your work account | Works in Excel and Outlook with enterprise data protection |
| Writing and long documents | Claude | Team (2 to 150 users) | Strong writing; Team content is not used for training by default |
| Questions about your own documents | Gemini Notebook | Work account, or free Standard for non-client material | Answers only from your sources, with citations |
| Technical research | Perplexity | Free, with AI data retention turned off | Numbered sources you can check |
| Client meetings | Fireflies.ai | Plan that fits your call volume | Says it never trains on meeting content |
If your firm runs on Google Workspace instead, swap Copilot for Gemini in Gmail, Docs and Sheets; Google says Workspace content is not reviewed by humans or used for training outside your domain without permission. For invoices that arrive by email, our invoice-processing stack shows how Zapier, Claude and Notion log and route them for approval. The wider picture of tools for finance work is in the best AI for finance and the best AI for spreadsheets.
A 30-day AI rollout plan for an accounting team
- Days 1 to 7: set the rules first. Agree which tools are approved and on which plans. Confirm that personal accounts are off limits for client work. Write a one-page policy covering what may be pasted (anonymised, internal or public material) and what never may (client identifiers, return data, payroll). Our guide to writing an AI policy has a template.
- Days 8 to 14: two low-risk pilots. Pick two tasks from the list above that use internal or anonymised data, such as formula checks and client email drafts. Three to five people try them on real work and note the time taken and every error they catch.
- Days 15 to 21: build the review habit. Turn the prompts that worked into a shared library. Add a review checklist: numbers traced to source, technical references opened and read, client name and figures checked by a person.
- Days 22 to 30: decide with evidence. Compare time saved against errors found. Decide which licences are worth buying, which tasks to add next (document summaries are usually third), and who owns the policy. Repeat the review every quarter, because plans and features change monthly.
The risks for accountants: confidentiality comes first
Confidentiality is one of the five fundamental principles in the IESBA Code, alongside integrity, objectivity, professional competence and due care, and professional behaviour. ICAEW's generative AI guide applies them directly: respecting client confidentiality means not loading confidential information into public generative AI tools, even when that information has become public. The same guide notes the Code was updated in 2022 to name threats such as undue reliance on technology.
In the United States, tax preparers have an extra rule. Section 7216 of the Internal Revenue Code limits how preparers disclose or use tax return information, and the IRS notes criminal and civil penalties under sections 7216 and 6713. Before any return data goes into an outside AI service, get your firm's view on whether it counts as a disclosure that needs the client's written consent.
Check the data settings of every tool. ChatGPT Free, Plus and Pro use your chats for training unless you switch off "Improve the model for everyone" under Settings, Data controls; ChatGPT Business and Enterprise do not train on your data by default. Claude Free, Pro, Max and Ultra can train on your chats unless you turn off the model improvement toggle under Settings, Privacy; Team and Enterprise do not by default. Microsoft says prompts from work and school accounts are not used to train its foundation models.
The last risk is invented detail. In October 2025 the Associated Press reported that Deloitte Australia agreed to repay part of its fee for a 237-page government report that contained a made-up court quote and references to research that did not exist; the revised version disclosed that Azure OpenAI had been used. A standard reference or tax rate that looks right is the most dangerous kind of error, so open every source an AI cites before you rely on it. Our guide on reducing AI hallucinations shows how.
Prompts to copy
Clean a messy export
This sheet is a bank export for [period]. Clean it on a new sheet: split the description column into payee and reference, convert dates stored as text into real dates in DD/MM/YYYY, remove exact duplicate rows, and add a column flagging any row where the amount is blank or the date falls outside [start date] to [end date]. Do not change the original sheet. List every change you made.
You get: A clean copy plus a change log you can review
Draft variance commentary
Here is a budget versus actual table for [period] and my notes on the reasons for each movement. Write management commentary of no more than 250 words for [audience]. Explain only the movements I have explained in my notes, in order of size. If a variance over [threshold] has no explanation in my notes, list it at the end under 'Needs explanation' instead of guessing.
You get: Commentary that sticks to your explanations, with gaps flagged
Find what matters in a lease
Using only the lease in this notebook, list: the lease term and any break clauses, rent review dates and method, any rent-free period or incentive, who is responsible for repairs and dilapidations, and any clause that could affect how the lease is measured for [IFRS 16 or your framework]. Quote the clause number for each point. Say 'not stated' where the lease is silent.
You get: A cited summary of the clauses an accountant needs
Explain a tax point to a client
Rewrite my explanation below for a small business owner with no finance background. Keep every fact and number exactly as I wrote them, use short sentences, no jargon, and end with one clear action and its deadline. Under 180 words. My explanation: [paste your explanation with the client's name and figures removed]
You get: A plain-English client email that keeps your facts
Audit a workbook's formulas
Review the formulas in this workbook. List: any formula containing a typed number instead of a cell reference, any cell in a column whose formula differs from the cells around it, any reference to an empty cell, and any circular reference. For each, give the sheet, the cell and a one-line explanation. Do not change anything.
You get: A list of formula risks to check by hand
Start technical research
What changed in [rule, standard or tax provision] in [jurisdiction] in the last 12 months? List each change with its effective date and link to the official source (regulator, standard setter or tax authority) for each. If you can only find secondary sources for a point, say so.
You get: A sourced starting list to verify against primary sources
Summarise a client planning call
From this transcript of a planning meeting with [client role], list: decisions made, information the client promised to send and by when, actions for our team with owners, and any question the client asked that we did not fully answer. Leave out small talk. Keep figures exactly as spoken and mark any you are unsure of.
You get: Clean meeting notes with owners and open questions