Ask like a founder. Answer like an auditor.

The plan fails closed. Ambiguous geography or unsupported paths return a clear refusal instead of a guessed number, and deterministic fallbacks cover AR, AP, cash, balance sheet and P&L.

No. It emits a semantic plan — measures, dimensions, filters, period — as JSON. core-api validates every field against the governed model and Cube executes it. Zero lines of model-generated SQL run in production.

Cash and runway, revenue by customer, product or geography, spend by supplier group, AR and AP ageing, month-over-month P&L lines, payout and fee breakdowns, and the personal equivalents. 65 of the 86 questions in our catalog are certified answerable; the rest fail closed.

Every answer returns the plan, the rows and the number, in that order. Drill from any figure to the ledger entries behind it; the same drill works from a dashboard tile.

Yes — Personal mode has its own semantic tables (15) and the same planner, isolated by tenant and product mode.

Report previews run in well under a second because no model is involved. AI answers take a few seconds: the model plans, core-api validates the plan against the semantic layer, then runs it on your own hardware.

No. The question, the plan and the data stay on the Finanly host. The AI gateway is internal-only and talks to Finanly's own 8B finance model running on your hardware; there is no cloud fallback.

Yes. Saved questions re-run every 15 minutes by a scheduled runner and can feed dashboard tiles, so 'what's our cash position' is always current when you open the page.

Three tools are exposed to the model: the chart of accounts, the cube measures and a period resolver. Vocabulary such as supplier groups, customer geography and product custom fields is read from the ledger, not guessed.

It is conversational, but it isn't free-form generation over your data. Every turn compiles to a validated plan; the model can explain and refine, but it cannot fabricate a figure the ledger doesn't hold.