The AI composes a custom configuration for every account — picking the right processor, tuning its parameters, linking it to the rest of the model — and refines until the numbers hold together. You override anything you want.
An LLM picks the strategy per account. The math is in Rust. NZ defaults are built in — GST, IRD payment dates, provisional tax, FBT exposure, working-capital filters. Every cell carries a calculation trace; every setting has a manual control.
Multi-tenant OAuth. Tokens encrypted at rest. Rate-limited per tenant and refreshed automatically. Pull live, or drop in last year’s PDFs and let the agent reconstruct the history.
For every account the agent chooses one of 23 processors and configures its options — drivers, lags, payment profiles, GST basis, tax-payment dates, linked accounts. Most material accounts end up on bespoke deterministic processors, not generic forecasters.
Click any account to open its deep editor: swap the processor, tune the sub-options, link it to other accounts. Or click a cell for a one-off Set / Add / Scale on top. Every figure carries a calculation trace.
The AI never invents numbers. It changes the model that produces them.— a working principle, not a marketing line
Run a forecast and the agent walks your Chart of Accounts. For each account it picks the right processor, tunes its sub-options, links it to drivers and other accounts. By the time it stops, every cell has a calculation trace you can open. Every assumption has a control you can flip.
Manual adjustments stack on top of the underlying processor. Click any forecast cell, choose Set, Add/Subtract, or Scale, and apply to this period only, the rest of the year, or every period. Edited cells render with a blue box. The agent sees them and refits around them.
For clients without clean Xero history. Drop in last year’s accounts as PDFs, Word docs, or images. The agent extracts every account and anchors them in time. The engine interpolates between snapshots to monthly history that feeds directly into the forecast.
If you cannot explain a number, fix it before completing.— the agent’s final quality gate
The same balanced-statement core powers a four-stage GST sweep and a full workpapers review. They share data, tools, and the audit trail. They never share invented numbers.
Sanity Check → P&L Sweep → BS Sweep → GST Wash-up. Inspectors pick up entertainment splits, shareholder accounts, depreciation, inter-entity balances, prepayments. Re-runs are change-tracked: New / AmountChanged / Resolved.
Across every process of the statements: prior-year column, compilation report, P&L, balance sheet, equity movements, shareholder current accounts, depreciation, the notes, and director, shareholder and trust resolutions. Where a check demands evidence, its finding must quote the exact line from the statements — a flag raised without that quote drops to manual review.
Every figure has a primary source. Workpaper findings quote the exact line of the financial statements that raised them. Forecast cells cite the ledger line or the override that produced them. Re-runs are change-tracked. Partner sign-off locks the audit trail.
Companies Act 1993
FRA 2013
Charities Act 2005
SES-2 Compilation · SPFR for FPE
NZBN · IRD · ATO
In the real product, you type in plain English and the agent rebuilds the Q3 model around your change. Here, four canned prompts show what that feels like — including the Numerical Impact card that surfaces in chat after every save.
| Account | Jul | Aug | Sep |
|---|---|---|---|
| Sales | $50,000 | $52,000 | $55,000 |
| Cost of sales | -$20,000 | -$21,000 | -$22,000 |
| Marketing | -$5,000 | -$5,000 | -$5,000 |
| Salaries | -$15,000 | -$15,000 | -$15,000 |
| Net profit | $10,000 | $11,000 | $13,000 |
In the live product you type a change in plain English. Here, pick one of these to see the same flow:
Pick a preset, reorder its pages, edit the prompt for any page, hit generate. The agent renders every page in parallel; a review agent reads them all and pulls in the assumptions you’d otherwise have to copy across. Then it’s yours to edit.
Formal, table-heavy, conservative palette. Minimum 12 pages. No charts. Full firm branding.
Paired tables and narrative. Critical Assumptions, Risk Factors, Conclusion & Outlook.
Concise KPI dashboard. Vibrant palette. Charts and action items, no boilerplate.
EBITDA continues to expand on the back of repeat commercial joinery work and the new cabinet-grade plywood supplier. Cash floor in August holds above the $45k covenant minimum under the base case; a 15% revenue stress reduces the floor to $19k.
Working-capital pressure eases through Q3 as the AR auto-OLS fit re-anchors against the commercial split. The model recommends holding the credit-control intervention paused until the FY27 review.
Connect a Xero tenant, pick a horizon, watch the agent compose the model. First draft to partner sign-off, typically under two hours.