Finance & FP&A AI Playbook
High-ROI AI applications for financial planning, analysis, and reporting — with prompts that work.
Where AI delivers in finance
The highest-ROI AI applications in finance are variance analysis narration, budget commentary drafting, report summarisation, and data extraction from unstructured documents (PDFs, earnings transcripts, invoices). These tasks share a key property: they require language, pattern recognition, and contextual understanding — exactly what LLMs do well. AI underperforms on tasks requiring real-time market data, complex multi-step numerical modelling, or decisions with regulatory accountability — a human must own those.
Prompt patterns that work for FP&A
For variance commentary: 'You are a senior FP&A analyst. Write a 3-sentence variance commentary for the following data: Revenue was $X vs. $Y budget (Z% variance). Context: [reason]. Tone: professional, concise, no jargon.' For budget narrative: paste your key metrics table and ask for a 200-word executive summary organised as: Performance, Drivers, Outlook. For earnings transcript analysis: 'Extract and summarise the CEO's three key priorities, any numerical guidance given, and any risk factors mentioned' before pasting the full transcript.
Connecting AI to your financial data
For Excel and Google Sheets: Microsoft 365 Copilot now works natively inside Excel for formula generation, data summarisation, and pivot analysis. For Python-based FP&A: use the Anthropic or OpenAI SDK to pipe dataframe summaries into prompts — never paste raw data files larger than ~50KB. For ERP data: build a RAG pipeline over your management reports and actuals data so AI can answer 'why did COGS increase in March?' against real documents rather than generalising from training data.
Model selection for finance
For text generation (commentary, summaries, narratives): Claude Sonnet or GPT-4o — the writing quality is measurably better than smaller models. For data extraction from PDFs and scanned documents: use a vision-capable model (Claude, GPT-4o, Gemini) — OCR-then-parse pipelines are being replaced. For high-volume classification tasks (categorise 10,000 transactions): GPT-4o-mini or Claude Haiku — 10–20× cheaper, accurate enough. Avoid using reasoning models (o3, Claude with extended thinking) for routine commentary — the token cost is 5–10× higher with minimal quality gain for this use case.
Governance and accuracy requirements
Finance is a domain where accuracy errors have material consequences. Establish clear rules: AI outputs used in board packs or external filings must be reviewed and signed off by a qualified finance professional. Never allow AI to generate numerical figures from scratch — it should always work from numbers you provide. Build a review checklist: does the commentary match the data? Are percentages calculated correctly? Are year labels accurate? The goal is AI as a first draft that reduces the blank-page problem, not as an autonomous financial reporter.
More Intermediate guides
Build Your First AI Agent
From a single LLM call to an autonomous multi-step agent — the practical path.
RAG Architecture: Ground AI in Your Data
Build retrieval-augmented generation pipelines that give models access to your documents.
Fine-Tuning vs. Prompting
When a well-crafted prompt is enough — and when you genuinely need a custom-trained model.