I built the AI Profitability Dashboard to answer a question every commercial finance leader knows: where is margin moving, and why? It runs in a browser, as a Claude artifact or a single web page, and its demo uses synthetic data for a fictional toy company across 13 markets, 16 brands and 161 SKUs.
The tool itself is only part of the story. Building it taught me a lot about how finance and AI should work together. These are the five lessons I would pass on.
1. Define the finance logic first
AI can write a dashboard quickly, but it cannot decide how your business defines variable margin. I specified the P&L structure, the like-for-like rules and the bridge method first. The variable margin bridge splits a change into volume, price and mix, allowances, product cost, royalties and logistics, and the steps must add up exactly. That rule is what makes the output credible in a review meeting.
2. Limit what the AI can use
The Insights Engine answers plain-English questions, but it works only from a pack of tables that the dashboard prepares for each question. Its instructions tell it not to invent figures, causes or outside facts, and to say so when the data cannot answer a question. A narrow scope is what makes an AI assistant usable in finance.
3. Always show the working
Every answer carries a "Data used" section with the tables behind it. A finance reader should never have to take a number on trust. If you would ask an analyst for the source, ask the same of the tool.
4. Keep Excel in the process
Finance people audit in Excel, so the What-If builder exports a formula-driven workbook. Every figure below the inputs is a live formula, inputs are marked, and check rows should read zero. People can then rebuild the logic, challenge it and take ownership of it.
5. Prototype on synthetic data
All the data in the demo is generated by a seeded model, so no real company figure is involved. That let me test freely and share the result openly. Before any real data goes near an AI tool, the access and confidentiality questions need clear answers.
The biggest change for me was speed. Once the logic was clear, I could go from a finance idea to something people can click through far faster than a traditional build. The finance thinking stayed the slow and valuable part, and I think that is the right way round.
Try the dashboard demo. Written Insights Engine answers need Claude, so on this site it shows the data tables behind each question.