Claude for accounting: bookkeeping, close and audit

What Claude can do for accountants

Claude for accounting means using Anthropic's Claude assistant to prepare and check routine finance work: answering bookkeeping questions, coding invoices, matching bank lines in a spreadsheet, drafting month-end variance commentary and reviewing audit workbooks. It matters because a well-briefed assistant can produce a first draft of this preparation work, which a qualified person then reviews, corrects where needed and signs.

The stance of this guide is simple: Claude prepares and checks, a named person reviews and stays accountable. Nothing Claude produces is posted, filed or sent to a client without that review. The sections below cover which Claude surface fits which job, three worked examples with the checks a reviewer applies, what data should never go into a chat, and how to verify outputs so they can stand as working papers.

Product details are stated as of September 2026 from Anthropic's own documentation. Plans, features and data terms change, so confirm them on Anthropic's pages before you rely on them in a policy.

Which Claude surface fits which accounting job

Claude is not one product but several surfaces on the same models. Choosing the right one matters more than prompt wording, because each surface decides what data Claude can see and what it can change. As of September 2026, Anthropic's help pages describe these options for plans from Free to Enterprise.

• Chat on claude.ai, the desktop app or mobile: questions, explanations, drafting emails and memos, reading an uploaded PDF or CSV.
• Projects: a workspace with its own chat history, uploaded knowledge (accounting policies, chart of accounts, close calendar) and project instructions; Team and Enterprise plans can share projects with view or edit permissions.
• File creation and code execution: Claude works in a sandbox, runs Python on uploaded data and returns Excel, Word, PowerPoint or PDF files, up to 30 MB per file.
• Claude for Excel: an add-in for Pro, Max, Team and Enterprise plans that answers questions about the open workbook with cell-level citations, traces formulas and fixes errors.
• Skills: folders of instructions and scripts that Claude loads for a specialised task, such as your firm's workbook review procedure; they require code execution to be enabled.
• Connectors: links to other systems through the Model Context Protocol (MCP), added from Customize, Connectors, or by an Owner on Team and Enterprise plans.
• The Claude API: for automation built by developers; see our guide on automating accounting with the Claude API.

Bookkeeping with Claude: coding a supplier invoice

The most common bookkeeping use is a second opinion on how to record a document. Give Claude the facts it needs and nothing more: the invoice lines, your capitalisation threshold, the relevant part of your chart of accounts and your VAT position. Ask for the journal, the reasoning and any assumption it had to make. Asking for assumptions is the step that turns a confident answer into a reviewable one.

Worked example. A consultancy receives an invoice for three laptops: net EUR 1,500, VAT at 20 percent EUR 300, total EUR 1,800. The firm's policy capitalises IT equipment costing EUR 1,000 or more, and it recovers input VAT in full. Reading the invoice total, Claude proposes: Dr Computer equipment 1,500 / Dr Input VAT 300 / Cr Trade payables 1,800. Debits of 1,800 equal credits of 1,800.

The reviewer checks four things before accepting it. First, the threshold: Claude compared the invoice total with it, but the policy may test each item, and at EUR 500 per laptop each item falls below EUR 1,000. Second, VAT recovery: the supplier's VAT number and the invoice details must meet the invoicing rules, or the 300 is not recoverable. Third, the period: the invoice date and delivery date decide which month it belongs to. Fourth, the supplier account: Claude cannot know whether this supplier already has an open duplicate invoice. In this case the policy is per item, so the reviewer rejects the capitalisation and records Dr IT expenses 1,500 / Dr Input VAT 300 / Cr Trade payables 1,800 instead, and adds the per-item rule to the project instructions so the next answer starts from it.

Bookkeeping in other languages: Buchhaltung and bokföring with Claude

Claude works in German, Swedish, French and many other languages, so a German bookkeeper can ask about Buchhaltung and a Swedish one about bokföring in their own language. The language is rarely the problem; the local rules are. A German business typically keys against a standard chart such as SKR 03 or SKR 04, a Swedish one against the BAS chart of accounts, and each country has its own VAT rates, invoicing rules and retention periods.

Two habits keep local answers reliable. Put your own chart of accounts and your country's current rates in the project knowledge rather than trusting the model's memory, and ask Claude to name the rule it is applying so you can check it against the official source. Every model has a knowledge cutoff: as of September 2026 Anthropic lists reliable knowledge through June 2026 for Claude Opus 5.5 and through January 2026 for Claude Sonnet 5. A rate or threshold that changed after that date will be wrong unless you supply it.

Reconciliations in spreadsheets

For reconciliations, upload the bank statement export and the ledger extract as CSV files and ask Claude to match them in the code-execution sandbox and return a workbook. Ask for formulas rather than pasted values, one tab per step, and a list of every unmatched line. A workbook with live formulas can be re-performed by the reviewer; a table of typed numbers cannot.

Worked example. The statement shows EUR 48,250 at 31 August and the cash book EUR 45,910. Claude's workbook adds a deposit in transit of 3,200 and deducts unpresented payments of 5,860, giving an adjusted bank balance of 45,590. On the book side it deducts a bank charge of 45 and an insurance direct debit of 275 not yet recorded, giving 45,590. The two sides agree with zero difference.

The reviewer re-performs the totals, traces the deposit in transit to the September statement and the unpresented payments to their payment runs, and checks that the two book adjustments are real: Dr Bank charges 45 / Cr Bank 45 and Dr Insurance expense 275 / Cr Bank 275. What gets rejected is any match Claude made on amount alone where references differ, for example two different customers paying 1,200 each on the same day. Our guide on bank reconciliation step by step covers the method itself.

Claude for month-end close: commentary and checks

At month end Claude is most useful for the writing that follows the numbers: variance commentary, close-pack narrative and the covering note to management. Give it the trial balance movement, the budget and the transaction detail behind any large variance, and instruct it to explain each variance only from lines in the data, citing the lines it relied on.

Worked example. Marketing expense is EUR 42,000 for August against a budget of 30,000, an adverse variance of 12,000 or 40 percent. The ledger detail shows a trade-show stand of 9,500 and an agency retainer of 2,500 invoiced in August for July work. Claude drafts: the variance is 12,000, of which 9,500 is the trade show and 2,500 is a timing difference on the retainer. The reviewer confirms both invoices in the ledger, checks whether the July retainer should have been accrued in July, and rejects a third sentence Claude added about higher digital advertising, because no ledger line supports it. Unsupported explanations are the typical error in AI commentary, and the rule to cite lines makes them easy to spot.

A project that holds the close calendar, the accruals policy and last month's commentary lets Claude flag items that recur every month and are missing this month. It suggests; the controller decides. Our guide on the month-end close checklist covers the close itself.

Audit working papers and an audit xls Claude skill

Auditors and reviewers use Claude to read workbooks rather than build them. Claude for Excel answers questions such as how a total in the lead schedule is derived, with clickable cell-level citations, and it can trace the source of a broken reference. Anthropic's own documentation says it is not recommended for audit-critical calculations without verification, and that data tables, macros and VBA are not supported, so it is a reviewer's aid, not evidence in itself. Two data points matter to a firm: Anthropic states that Claude for Excel does not inherit an organisation's custom data retention settings, and that its activity is not included in Enterprise audit logs.

What people search for as an audit xls Claude skill is a Skill: a folder with written instructions, and optionally scripts, that tells Claude how to review a spreadsheet the same way every time. You can write your own and upload it under Customize, Skills. Anthropic also publishes financial-services plugins for Cowork and Claude Code, and warns that AI-generated financial analysis should be reviewed by qualified professionals before it is used. A useful workbook-review skill checks the following.

• Totals that do not foot or cross-foot, and subtotals that exclude a row.
• Hard-coded numbers inside formula ranges and formulas that differ from their neighbours in the same row or column.
• Links to other workbooks, error values such as REF and DIV/0, and hidden rows or sheets.
• Ties between the lead schedule, the trial balance and the draft financial statements.
• Sign errors, such as credits entered as positive expenses.
• A findings table listing each issue with the cell reference, so the reviewer can clear or reject each one.

What to paste into Claude and what never to paste

Data protection starts with the plan. As of September 2026, Anthropic states that by default it does not use inputs or outputs from its commercial products, the Team and Enterprise plans and the API, to train models; the exception is feedback a user actively submits with the rating buttons, which organisation owners can switch off. Anthropic acts as processor for commercial customers, who remain controller. On the consumer Free, Pro and Max plans, chats are used to improve Claude when the user has the model improvement setting switched on, and also in cases such as conversations flagged for safety review or feedback the user submits; incognito chats are not used. A firm handling client data should use a commercial plan and put the processor terms on file.

Under the GDPR you still need a lawful basis for sending personal data, a processor agreement, a view on international transfers and a record of what you send. Professional duties add a further layer: the confidentiality principle in the IESBA Code applies to client information whatever tool you use. The practical rule is data minimisation: send what the task needs, remove what it does not.

• Paste freely: account names and codes, anonymised trial balances, policies, standards questions, your own templates.
• Paste with care: supplier invoices and ledger detail, after removing personal data that the task does not need.
• Pseudonymise: employee names, customer names of individuals, and anything that identifies a private person.
• Never paste: passwords, API keys, online-banking credentials, full card numbers or national ID numbers.
• Never paste without authority: client data your engagement letter or data-processing terms do not cover.
• Keep a copy: save the prompt, the files used and the output with the working paper.

Connectors, files and prompt injection

Connectors let Claude read and act in other systems, which raises the stakes. Anthropic's guidance is to connect only servers from trusted organisations, review the permission scopes you grant, review each tool approval request, click Always allow only for trusted servers, and set tools you do not need to Blocked. For accounting that means read tools on, write tools off, unless a human approves every action.

The specific risk is prompt injection: instructions hidden inside a document Claude reads. A supplier PDF could carry white-on-white text telling the assistant to update the supplier's bank details or to summarise the ledger into a web request. Anthropic warns that files from external sources can contain hidden instructions and that its file-creation sandbox could be tricked into sending data out if network access is enabled; its advice is to monitor Claude and start with network access off. Treat every external document as data, never as instructions, and never let a chat session both read supplier files and change payment details.

How to verify Claude's outputs

Confident errors look exactly like correct answers, which is why review discipline matters for any AI-prepared work. A second pass by Claude is useful but is not independent evidence. The checks below are what an experienced reviewer applies to any AI-prepared schedule.

• Re-foot every total yourself or with a spreadsheet formula; do not accept arithmetic stated in prose.
• Trace each figure to a source document or ledger line, and each explanation to a transaction.
• Check dates, periods and cut-off; AI drafts often ignore the delivery date.
• Verify every reference to a standard, paragraph, law or rate against the official text; invented or outdated references are a known failure.
• Ask Claude to list its assumptions, then confirm or reject each one.
• Record who reviewed, what was changed and what was rejected, so the file shows professional judgement was applied, as ISA 230 expects of audit documentation.

Using Claude with Skyline Nexus ERP

Skyline Nexus ERP has an AI help assistant on every screen that answers how-do-I questions from the product's own help library, in the language the user writes in. Each answer is checked before it is shown: the help page it cites must exist and must be one of the passages actually retrieved for that question, otherwise the user sees plain search results instead. Asking for figures such as sales today or sales by branch this week returns the real number from a fixed list of eight reports, limited to the branches and permissions the user already has.

For Claude Desktop and other MCP clients, Skyline Nexus ERP runs an MCP server, the Skyline ERP Assistant. An administrator issues each user a personal access token; the server then offers four tools: help search, help page, the eight live business figures (read-only and permission-checked) and creating a support ticket. Nothing else can be written through it. Claude drafting journals, bills and bank matches for human approval, and ledger reads through MCP, are on the Skyline Nexus ERP roadmap; ask us for your go-live date. Until then, the Audit Pack export gives you a single workbook of trial balance, journals and ledger detail to review with Claude for Excel under the data rules above.