I run finance operations for a commercial real estate portfolio: multi entity cash, month end close, reconciliations, AP, investor reporting. For the past year I have been rebuilding those workflows around AI and automation, and shipping the systems myself instead of waiting for a vendor to ship them.
The interesting question was never what a model can do. It is which decisions should stop being human, which have to stay human, and how you prove the difference to an auditor.
finance-ai-transformation is the portfolio: ten sanitized case studies from systems that run in production. Every one starts with a finance problem and ends with the control design, not with the model.
A few of them:
| Private Invoice Intelligence | Multimodal extraction plus RAG on private infrastructure, so invoice contents never reach a commercial AI API. The model recommends. It never posts. |
| Transaction Fraud Monitoring | Calibrated statistical detection, AI assisted interpretation, human decisions, and an audit trail behind all of it. |
| Deposit Reconciliation | Compares bank deposits against the ledger and sends only the exceptions somebody has to act on. |
The stack is whatever the finance problem actually needs: Google Apps Script and Workspace add ons, Cloud Run, bank data APIs, monday.com, Slack, Python, JavaScript, structured outputs, evaluation pipelines.
- Start with the financial control or the operating decision, not with the technology.
- Automate collection, comparison, routing and presentation before automating judgment.
- Keep consequential decisions with an accountable person, and design the failure states so the workflow stays safe when the model is wrong.
I wrote two books. The second, The Trap of Artificial Intelligence, is about where AI stops being useful and what that means for the people working next to it.
Most of what I publish goes on LinkedIn, where I write about finance operations, applied AI and the parts that did not work.