{ case study 05 }
AI Financial OS
Cash flow, runway, invoices and recurring bills over one ledger. The forecast beside them is plain arithmetic: it prints every assumption it made, every projected line in date order, and what it left out.
Demo login demo@example.com demo-password-123
The workspace is labelled demo data throughout.
{ the build }
A number
owes you its work.
Problem
Bookkeeping tools show you the past accurately. Forecasting tools show you a future you cannot interrogate: a smooth line to twelve months out, no statement of what it assumed, no way to find the one recurring item that is carrying the whole curve.
Put a language model on top and it gets worse in a specific way. Ask it for your runway and it will produce a number, formatted well, sourced from nothing you can check. For an operator deciding whether to make a hire, a plausible wrong number is more expensive than no number.
Approach
Split the arithmetic from the language and let neither do the other's job. Projections are computed deterministically from recurring items, open invoices and manual adjustments. No model touches them, and the forecast page says so on its face.
The assistant gets the narrower half: it may only speak about figures the system has already computed. It names the period behind each one, it is read-only against every record, and when the AI layer is off it degrades to verified summaries rather than going quiet. Every forecast ships with its assumptions listed, including the things it excluded, because a projection that hides its inputs is an opinion with a chart attached.
{ how it runs }
Ledger in,
decisions out.
Four stages. The model is only allowed anywhere near the last one.
One ledger
Transactions, invoices and recurring items live in a single workspace, filled by CSV import or entered directly. Everything downstream reads from here, so the dashboard and the spreadsheet cannot drift apart, because there is only one set of numbers.
Transactions Invoices · recurring CSV import
Derive, with the basis attached
Balance, runway, category breakdown, open receivables and collection rate are computed per period, and each states what it rests on. Runway is not "7.3 months" alone; it is 7.3 months on an average burn of $2,160 over 3 complete months, and it says so under the figure.
Period-scoped figures Stated basis Threshold alerts
Forecast, deterministically
Three scenarios with their rules written out, compared side by side on ending balance, lowest balance and whether either goes below zero. Underneath sits the assumptions list, including the exclusions, and a table of every projected inflow and outflow in date order with its source. No AI is used in this calculation.
Base · conservative · growth 30 / 60 / 90 day horizon Assumptions printed
Explain, read-only
The AI CFO answers "can I afford a $3,000 purchase next month" in plain language, using only figures the system calculated, naming the period and assumptions behind each. It cannot edit a record or move money, and with the AI layer off it returns verified summaries instead of silence.
Grounded in computed figures Read-only Deterministic fallback
{ what it produces }
Projections
that show their work.
Built as financial information and planning, and labelled that way on every screen. Not professional advice.
{ more work }