Claude for audit.

Where Claude’s long-context window and careful drafting fit into a statutory audit. What it’s genuinely good at, where it falls short, and the prompts that play to its strengths.

Claude is a large language model built by Anthropic. From an auditor’s practical standpoint, three things differentiate it from the rest of the consumer LLM pack: a very large context window (up to a million tokens on the latest Opus tier), a drafting voice that skews formal and structured, and a refusal pattern that errs towards caution on grey-area requests.

The context window matters because most audit reference material — Standards on Auditing, Ind AS, CARO 2020, the Companies Act 2013 — is long. Claude can hold an entire SA in working memory, cross- reference it against a fact pattern, and produce a structured response without the “wait, what did paragraph 17 say?” problem that smaller-context models have. For drafting, it tends to produce file-grade prose on the first pass.

2.1 · Reading long regulatory text

Paste the full text of SA 540 (Auditing Accounting Estimates) and ask Claude to walk you through how the new requirements differ from the old SA 540 — it handles the entire 60-page standard in one go. Same for Ind AS 115 (Revenue), Ind AS 116 (Leases), or the full CARO 2020 order with all 21 clauses. This is the single biggest workflow shift it enables.

2.2 · Drafting long-form audit documents

Engagement letters, MRLs, communications under SA 260, the basis-for-opinion paragraphs in a modified audit report — Claude produces clean drafts that tend to need only light editing. The voice is formal without being pompous, which is hard to get right.

2.3 · Multi-document reasoning

Drop in a board resolution, a related-party policy, and a sample transaction — ask Claude whether the transaction was board-approved under Section 188. It will hold all three documents in context, check the policy against the resolution, identify gaps, and explain its reasoning step by step. This is audit work that used to require an article assistant’s full afternoon.

2.4 · Stress-testing your conclusions

State your audit conclusion, then ask Claude to argue the opposite. It will produce a counter-argument of comparable rigour, which is exactly what an EQR partner does mentally. Genuinely useful for going- concern conclusions under SA 570 and KAM judgements under SA 701.

2.5 · Structured outputs and tables

Ask for a working-paper-grade table — rows, columns, exact prescribed labels — and Claude tends to format cleanly. Useful for things like the CARO clause-by-clause checklist, the IRAC classification matrix in bank audits, or the Section 13(3) specified-person tracker in trust audits.

Three honest weaknesses for audit-specific work:

• Live web access in the free tier is limited. Claude doesn’t browse by default in the free Sonnet tier. Paid tiers have web search, but for “what did MCA notify last week” queries, Perplexity is faster.
• Image generation is essentially absent. If you need a process flowchart for an audit working paper, you’ll draw it yourself or use a different tool.
• Indian-context sometimes lags. Generic content (SAs, Ind AS framework, Companies Act provisions) is solid; very recent ICAI notifications or RBI Master Direction updates can be stale or absent depending on when the model was last updated.
• Can be conservative on legitimate audit questions. Asking Claude to draft a Section 13(3) test query that scrutinises related-party transactions sometimes triggers caution; rephrasing as a methodology question fixes it.

• Projects. Claude lets you create a Project for a single client engagement (or a single methodology area). Drop in your firm’s methodology notes, working-paper standards, prior-year files (anonymised or generic). Every conversation in that Project inherits the context, so you don’t re-explain every time.
• Artifacts. Outputs render as editable canvases — a draft engagement letter, a CARO clause table, a working-paper template — that you can iterate on in-place. Cuts down on copy-paste churn.
• Analysis tool (code interpreter). Claude can write and run Python on your behalf for ledger calculations, ratio analysis, sample-size determinations. Useful when you’ve already exported the data and want a quick analytical pass.
• Long-context mode (Opus). When you need to drop multiple long documents at once — a 200-page regulatory order plus the client’s response plus the audit team’s prior memo — Opus with the 1M token window holds all of it.

A rough fit map. The strongest fits are bolded — the rest are usable but other tools may serve better.

• Planning. Strong fit. Industry risk briefings, materiality benchmark discussions, audit-strategy memo skeletons.
• Risk assessment. Strong fit. ROMM brainstorming, SA 240 fraud-scheme generation, ICFR control matrix skeletons.
• Sampling. Usable. Methodology defence memos under SA 530; for sample-size math an Excel template or an audit-grade tool serves better.
• Substantive testing — pattern detection. Not the right tool. Use an audit-grade engine like MindBridge or Inflo for full-population scoring.
• Substantive testing — reading documents. Strong fit. Cross-referencing a board minute against a contract against the underlying ledger entry.
• CARO drafting. Strong fit. Clause-by-clause language, not-applicable reasoning, consistent voice across the annexure.
• Communications and drafting. Strongest fit. Engagement letters, MRLs, SA 260 communications, management letter findings.
• Concluding work. Not used. The audit opinion is the partner’s judgement.

These play to Claude’s structured-reasoning and long-context strengths. More live in the Audit Prompt Library.

I'm going to paste the full text of SA 540 (Revised), and then a generic fact pattern about a manufacturing entity's inventory provisioning. Walk me through:
1. Which paragraphs of SA 540 are directly relevant
2. The specific procedures the standard requires for this kind of estimate
3. What documentation under SA 230 the auditor would need to keep
4. Three additional questions the auditor should put to management

[Paste SA 540 text here]
[Paste anonymised fact pattern here]

My engagement team has concluded that going concern is appropriate without a material uncertainty disclosure under SA 570 (Revised). The supporting evidence: management's 12-month cash flow forecast assuming the working-capital line is renewed and a key debtor pays within 60 days.

Argue the opposite case. What counter-considerations under SA 570.16-18 would suggest a material uncertainty exists? What additional evidence should the EQR partner reasonably ask for? What disclosure would be appropriate if the counter-case prevailed?

Be rigorous, not contrarian for its own sake.

Draft the "Basis for Qualified Opinion" paragraph for an Indian auditor's report under SA 705 (Revised). The qualification is: the company has not provided for diminution in the value of a long-term investment in a wholly-owned subsidiary, even though the subsidiary has incurred losses for three consecutive years and its net worth has eroded by 80%. Quantum of the resulting misstatement is material but not pervasive.

Use the SA 705 illustrative format. Keep it tight — three short paragraphs. Cite the standard.

Three tiers matter for an Indian audit firm:

• Claude.ai free. Sonnet model, web-search-light, no Projects. Anthropic’s policy on free conversations is that they may be used to improve models unless you opt out in account settings. For methodology questions, fine. Not for anything that could conceivably identify a client.
• Claude Pro / Team. Paid consumer / small-team tier. Higher rate limits, Projects, Artifacts, analysis tool. Same data-handling posture as the free tier — opt-out available in settings.
• Claude for Work / Enterprise / API. Contracted enterprise tier. Anthropic does not train on customer inputs or outputs by contract. This is the tier most CA firms should be on if they intend to put any work-adjacent (still not real-client) information in.

Regardless of tier, treat the SA 230 documentation discipline as a constant: if Claude was used to get to a conclusion, that fact and the prompt approach belong in the working paper.

• It will sometimes hallucinate paragraph numbers. If Claude cites “SA 540, paragraph A47”, open the standard and verify. This is the most common error pattern.
• It can mix up Ind AS and IFRS where the two diverge. Especially on revenue recognition and leases, double-check that the response is consistent with the Indian carve-outs.
• Date ambiguity. Claude doesn’t know “today’s date” reliably in the free tier; for date-sensitive questions (e.g. “is this circular still in force”) you need to either tell it the date or verify against the source.
• Refusal patterns on legitimate audit work. Asking it to “draft a query questioning whether the client has misstated revenue” sometimes hits a refusal. Reframe as “draft an audit query on revenue recognition for management” — same goal, no refusal.

• Pick one drafting task in a current engagement. Open Claude (any tier), give it the methodology + anonymised fact pattern, ask for a draft. Edit and ship.
• Create a Project for your firm’s audit methodology — add your standard working-paper templates and your house style. Every subsequent prompt in that Project inherits the context.
• Pick one SA you find tricky. Paste the full text, ask Claude to give you a walkthrough with examples. Notice which examples ring true to your practice; you’ve just sharpened your grasp of the standard in 15 minutes.