To populate an Excel underwriting model with AI safely, the system must understand both sides of the handoff: what the source document means and what the target cell expects. Copying a plausible number into a nearby blank cell is not automation; it is an untraceable model risk.
Map the model before mapping the deal
Document input sheets, named ranges, units, date periods, sign conventions, scenario toggles, formula regions, and required fields. A cell labeled taxes might expect the latest historical total, a per-unit amount, or a forward reassessed value. The label alone is insufficient.
Treat the workbook as the firm’s decision language. Its structure encodes how the team defines revenue, normalizes operations, sizes debt, and presents returns. The point is to preserve that language while removing repetitive entry.
Create evidence-aware source bindings
For each input, define the preferred source, acceptable alternatives, required unit and period, and whether the field is historical or forward-looking. Bindings should carry the source reference and any transformation, such as annualizing a monthly value or converting an expense to a positive model input.
- Value and normalized unit
- Reporting period and as-of date
- Original sign and target sign
- Source document and location
- Transformation or assumption reasoning
Handle missing history and future assumptions differently
Do not invent an absent historical fact. Leave it visible and request the evidence. A required future input can be estimated when the context supports it, but the estimate should be labeled, explained, and easy to replace.
This boundary matters because a complete-looking workbook can still be misleading. Coverage is useful only when a reviewer can tell what was observed, what was inferred, and what remains unresolved.
Recalculate and validate the completed workbook
Check that formulas remain formulas, input cells receive the expected types, totals reconcile, and key outputs respond to scenario changes. Reopen and recalculate the file in the supported spreadsheet environment when native behavior matters.
Review should be exception-led. Focus on conflicts, unmapped fields, estimates, high-impact assumptions, and unusual formula behavior instead of spending equal time on every clean populated cell.
Want to see this on your own pipeline?
Send us a real OM and we’ll show you the underwrite, no sales pitch required.
Book a demo