Notes from building a sovereign AI CFO: control loops, local models, cash forecasting, covenants, master data, EU rules on AI and OSINT for finance teams.
An AI PC with 128 GB, a desk-side DGX Spark or a European AI factory: where a controlling model should run, and why a night-shift agent needs more than a laptop.
Whether a company can afford a machine, a hire or a new site usually takes a week of spreadsheet work. With a standing model of cash, commitments and scenarios it is a query.
Revenue per customer is in every report. Cost to serve is scattered across returns, terms, late payments and support time, so the best customer may be the least profitable.
A demo forecast always looks good. A backtest on your own closed periods, against a naive baseline, shows whether it is any better. How to run one, step by step.
Loan covenants are tested quarterly on trailing twelve months, after the close, when it is too late to act. How to forecast covenant headroom months ahead instead.
Ask a model why revenue missed plan and you get a fluent story. Unless price, volume, mix and FX are separated first, that story is often wrong. How to fix the order.
Duplicate customers, drifting cost centres and free-text fields cap what any model can tell you. The master-data failures that matter most, and a test to run today.
Most 13-week cash forecasts are built from due dates and AR ageing, but cash moves on payment behaviour. Where the gap comes from, and a test to run this week.
An AI CFO is not a chart that answers questions. It is a closed control loop that notices drift, decides what to do about it, and shows its reasoning.
Documentation, logging, accuracy evidence and human oversight: the habits worth building into a finance model as EU rules on AI take shape.
Sending ledger lines to a hosted model is a business decision, not a technical one. What running controlling models on your own hardware actually costs and buys.
Your forecast is built from internal history. The disruptions that break it are visible outside the company first — in public sources, weeks earlier.
The monthly close is a batch process inherited from paper. What replaces it is not a faster close — it is a control loop that never stops running.
