How is AI being used in accounting work?
In practice and in-house teams, AI is mostly used for the writing and explaining that surround the numbers rather than for the numbers themselves. That is where it is most reliable and where a mistake is easiest to catch.
Many accounting and bookkeeping platforms also include AI features, such as suggested categories or automated matching. Those are configured by the software vendor and your administrator; this guide is about how individual accountants use AI assistants and review AI suggestions in their own work.
Client communication
Drafting requests for missing records, explanations of a filing, or a plain-English summary of a set of accounts for a client who is not an accountant.
Research starting points
Summarising a long piece of guidance you have obtained, or listing questions to research, before you go to the primary source.
Procedure notes
Turning your own steps into a clear checklist or process document for a junior colleague.
Reconciliation notes
Writing up unreconciled items clearly, with what is known and what is still open.
Spreadsheet help
Explaining an inherited formula or suggesting a cleaner one, tested before it is relied on.
What does professional guidance say about accountants using AI?
Professional bodies are applying existing ethical principles to AI rather than writing an entirely separate rulebook. ICAEW's generative AI guide applies the five fundamental principles of integrity, objectivity, professional competence and due care, confidentiality and professional behaviour to the use of these tools.
Its practical points are ones every accountant can act on today. It advises against loading confidential information into public generative AI tools, because users have limited control over who sees the data and how long it is kept. It calls for professional scepticism when reviewing output and warns against automation bias. It encourages openness about using generative AI in client work. And it says plainly that these tools do not make up for a lack of professional knowledge and skill, noting that they may not reflect current legislation.
If you belong to another body, check its own guidance and code of ethics, which may phrase the expectations differently. This page is general information, not professional or legal advice.
What does a reviewed AI client email look like?
Priya Nair runs a fictional interior design business as a sole trader, and her accountant needs several missing items before preparing her year-end accounts: three months of bank statements, receipts for two large equipment purchases and confirmation of how much of her home she uses for work.
Asked to 'write a friendly email to a client asking for missing year-end documents', the assistant produced a pleasant email that also stated a filing deadline, warned of a specific penalty amount and said home office costs could be claimed 'at a flat rate of 30 percent'. None of that came from the accountant. The deadline and penalty depended on her jurisdiction and filing status, and the home office line was simply wrong as a general statement.
The reviewed prompt listed the three items, the date by which the accountant needed them and the upload method, and added: 'Do not state any tax rule, deadline, rate or penalty.' The accountant then added one sentence of their own about the deadline, checked against the tax authority's published dates. The final email was shorter, specific and contained nothing the accountant could not stand behind.
Can AI help with bank reconciliation notes?
It can help you write them clearly, but it should not resolve them. Suppose a fictional client's bank reconciliation has two unmatched items: a payment of 1,240 to a supplier that appears twice in the bank feed, and a receipt of 380 with no matching invoice.
An AI draft described the first as 'a duplicate payment that should be reclaimed from the supplier' and the second as 'likely a customer prepayment'. Both may be true, and both are guesses. The double payment could be two genuine invoices of the same value; the receipt could be a refund, a transfer or a deposit. The useful version records each item, its date and amount, the possible explanations and the evidence needed to close it, and leaves the conclusion to the person who checks the supplier statement and the client's records.
Want to see how this connects to a role?
Explore the planned team trial ↗Can AI answer tax and technical accounting questions?
Treat any AI answer on tax or accounting standards as a research lead, not an answer. General assistants may draw on outdated or incomplete material, mix rules from different countries and state thresholds that have since changed, all in a confident tone. The ICAEW guide makes the same point about current legislation.
Use AI to frame the question, identify the relevant area and list what to look up, then go to the primary source: the legislation, the tax authority's own guidance or the accounting standard. Record which source you relied on. If the tool cites a source, open it and find the passage; a real-looking reference is not the same as a correct one.
How should bookkeepers review AI categorisation?
Review the rules, not just the transactions. AI categorisation in bookkeeping software usually learns from past coding, so an early mistake can repeat every month. Check new suppliers, mixed-use expenses, anything near a capital or revenue boundary and transactions with vague descriptions such as card payments to marketplaces.
Keep an exceptions list of the payees and patterns the suggestions get wrong, and check those every period. Suggested matching can save time on clear cases; unclear cases still need the receipt.
Will AI replace accountants?
Nobody can answer that reliably, and predictions presented as fact should be treated with caution. What can be said is that AI is taking on more of the drafting, summarising and pattern matching in accounting work, and that responsibility for judgement, professional scepticism, client confidentiality and signing off the numbers remains with qualified people under professional rules.
A related question, whether AI is better than a qualified accountant, rests on a category error. An AI tool can produce text that sounds like advice; it cannot take professional responsibility for it, know the client's full position or be held to a code of ethics. The accountants likely to benefit are those who learn to use these tools while keeping those responsibilities firmly in hand.
What client information can go into an AI tool?
Only what your firm has approved for that specific tool and plan, and nothing at all into a public consumer tool if the client's information is confidential. Engagement terms may also restrict how client data is processed and by whom.
For practice and for building templates, use fictional data like the examples on this page. When AI is used on real client work inside an approved system, keep the same review trail you would keep for any other work: the source, the checks and who approved the final version.
How can an accounting team practise these habits?
Use the client email and reconciliation examples as team exercises with fictional data. Each person drafts with AI, then lists what they removed or left open. The lists show who catches invented rules and who resolves items the evidence does not support.
Tosheo.ai is being built with a planned finance and accounting path that uses tasks like these, feedback against a visible rubric and a manager view of skill progress without chat logs. It is pre-build, so this is planned scope. It does not provide accounting, tax or regulatory training or count towards professional CPD requirements unless your body says so.