What AI actually looks like for a finance team

Finance teams sit in an unusual position with AI. On one hand, their work is heavily document-based, analytical, and repetitive in structure — exactly the kind of work Claude handles well. On the other hand, the consequences of errors are real: wrong numbers in a board report, a miscommunicated position in an investor update, a compliance gap. The tolerance for "approximately right" is lower than in almost any other function.

This creates a specific set of use cases that work well and a clear set that should stay human.

Where Claude delivers for finance

First drafts of recurring reports. Monthly board packs, investor updates, departmental summaries — these follow a template every time. The structure is the same; the numbers change. Claude can take the numbers, the prior period's report, and the key narratives, and produce a first draft that is 80% there. The finance lead reviews, corrects the framing, and adds judgment. Two hours of work becomes thirty minutes.

Variance commentary. "Revenue is up 12% vs. plan. Write three sentences explaining why, based on this data." Claude is good at this — turning numbers into prose that non-finance stakeholders can read. The human still owns the interpretation; Claude handles the writing.

Document summarisation and extraction. Contracts, vendor agreements, regulatory filings — there is always more to read than there is time to read it. Claude can extract key terms, flag clauses that need attention, and produce a structured summary. This is not legal advice; it is triage.

Internal Q&A and policy lookup. "What is our capitalization policy for software development costs?" "What does the FX hedging policy say about threshold triggers?" If you upload your finance policies and procedures to a Project, your team can ask questions in plain language and get accurate answers sourced from your own documents.

Drafting financial communications. Emails to auditors, responses to investor queries, CFO talking points for earnings calls. The financial content is human-generated; Claude handles the writing and structure.

Where Claude has no place in finance

Producing numbers. Claude generates text. It does not crunch spreadsheets reliably — it can appear to, but it makes calculation errors that are hard to spot. For any actual arithmetic, your spreadsheet or BI tool is the right system. Use Claude downstream of the numbers, not to produce them.

Making judgments about accounting treatment. "Should we capitalise or expense this?" is an accounting judgment that requires your policies, your auditor's position, and professional judgment. Claude can tell you what the general standard says. It cannot tell you what the right answer is for your specific situation.

Compliance sign-off. No AI output should serve as the final word on whether something is compliant. Claude can help you understand a regulation in plain English; it cannot tell you whether your specific practice is compliant.

The Project setup that works

Create one Project for finance. Upload:

• Your chart of accounts and key definitions
• Reporting templates and prior period examples
• Finance policies and procedures
• Key metrics definitions (so Claude uses your definitions, not generic ones)

System prompt: "You are a finance assistant for [Company]. You help the finance team draft communications, summarise documents, and produce first drafts of reports. You never produce or verify numbers — you work with numbers the team provides. You never give accounting or compliance advice. When in doubt about financial accuracy, flag it explicitly."

That last instruction — flagging uncertainty explicitly — is the most important one. A finance Claude that confidently produces wrong information is more dangerous than one that says "I'm not certain about this figure, please verify." Make uncertainty visible.

The pattern that works

Use Claude where your finance team is spending time on writing, synthesis, and communication — not where they are making financial judgments. The leverage is real; it just requires knowing which side of that line you are on.