How I built an AI-powered financial commentary tool that reduced monthly reporting time from 4 hours to under 30 minutes — and made the output sharper than anything written manually.
Every month, after closing the books, the real work began: writing the narrative. P&L variance commentary for the CFO and board had to explain every meaningful movement — by entity, by revenue stream, by cost category — with the right tone, the right level of detail, and a clear link back to operational drivers.
This took 3 to 4 hours every month. The output was inconsistent — some months sharper than others, depending on how much time was available. And it was entirely manual: stare at a spreadsheet, think about what moved, write a sentence, move on.
The bigger issue: that 4 hours was time not spent on actual analysis. The commentary was describing the numbers — not interrogating them.
The finance team was spending its most cognitively expensive hours on a task that is, at its core, pattern recognition and language generation. Exactly what AI is built for.
I designed, built, and deployed this tool entirely independently. No IT support, no developer involvement. From identifying the problem to shipping a working app used in production monthly reporting — all within my FP&A role at Malvern International plc.
The tool needed to work with our actual data structure, understand our business context (ELT, Pathways, Juniors, deferred income, agent revenue), and produce commentary that matched the tone and depth expected at CFO and board level.
Built a Python/Pandas pipeline that ingests the monthly P&L export from our accounting system, cleans and structures it, and produces a variance table by entity and category — actuals vs budget, actuals vs prior year, with calculated variance amounts and percentages.
Designed a structured prompt system that passes the variance data alongside business context — what each entity does, what deferred income means, what drives agent revenue — so the AI generates commentary grounded in the actual business, not generic financial language.
Integrated with Groq's API (using LLaMA 3) for sub-second response time. This was critical — the tool needed to feel fast and interactive, not like waiting for a report to generate. Responses arrive in under 2 seconds for full P&L commentary.
Wrapped everything in a clean Streamlit app — upload the Excel file, select the reporting period, click Generate. Output appears with headlines, driver analysis, and flag-worthy items. Commentary can be copied directly into the board pack.
Built in an edit interface — every generated paragraph is editable before export. The AI does the first draft, the analyst adds judgement. This is the right division of labour.
The commentary is now more consistent and more insightful than what I was producing manually. Because the AI processes every line of variance simultaneously, it doesn't miss the £12k movement buried in materials costs that a tired analyst scanning at 6pm would have skipped.
The CFO now receives a first-draft commentary within minutes of the data being closed. We spend the saved time on the analysis that actually matters — scenario planning, understanding the drivers, thinking about what it means for next quarter.