Inside the underwriting engine: how the platform prices a building

Automated underwriting is a pipeline, not a magic number. Data is pulled from sources of record, cross-checked across independent passes, and priced inside a deterministic financial engine that sits outside the language model. Every output carries a confidence score, and anything below threshold is routed to a human before it reaches a decision. The result is an estimate with its evidence attached, not a verdict.

The pipeline starts with data, and it starts strict. Every input that feeds a valuation, whether it is a rent roll, a comparable transaction, an operating statement, a market report or a planning record, is pulled from the original of record. Nothing is recalled from a model’s training data, and nothing is re-imagined from memory. If a number cannot be anchored to a source, it does not enter the model.

This matters more than any modelling choice downstream. A valuation built on recalled or approximated inputs can be internally consistent and still wrong, because the foundation drifted before the arithmetic began.

General-purpose AI has a variance problem. Ask a standard model the same question about the same deal twice and you can get materially different answers, because the model re-samples the world on every run. An underwriting system cannot work that way.

The engine runs the same scenario through parallel, independent passes. Where the outputs agree within tolerance, the result stands. Where they do not, the disagreement is flagged and the number is re-derived from primary data rather than re-prompted. Repetition is treated as a test that the number is real, not as a formality.

The financial mathematics itself, cash flows, return measures, leverage effects and sensitivities, runs inside a deterministic financial engine, outside the language model. The model proposes; the engine computes. Re-running the same deal with the same data produces the same answer, which is the property that makes an output auditable at all.

Every line of the output is scored on data quality, comparable depth and assumption defensibility. Numbers that fall below threshold are surfaced for human attention before they reach anything that looks like an investment decision. An analyst does not re-derive the whole model; they interrogate the specific points the system itself has flagged as weakest.

The division of labour is deliberate. The machine covers breadth and consistency, applying the same rigour to the five-hundredth asset as to the first. The human covers judgment: whether an assumption is reasonable is a question about the market, and it stays with people.

Illustrative. Not a Smart Bricks transaction, and deliberately abstract. A residential unit is listed in a market the platform covers. Ingestion pulls the listing itself, registry records for the building, recent transactions for comparable units, and current rental evidence, each from its source of record. Reconciliation runs the pricing question through parallel passes; two agree closely and one drifts, so the outlier is traced back to a stale comparable and discarded. The deterministic engine then prices the unit off the surviving evidence and attaches a confidence score. Because comparable depth for that building type is thin, the score falls below threshold and the estimate is routed to an analyst with the weak points already marked.

Is the estimate just an AI’s opinion?

No. The language model contributes reasoning and reading, but the number itself is computed in a deterministic engine from source-anchored inputs, and it is checked across independent passes before anything is shown. What you see is the output after interrogation, not a first draft.

Can the engine still be wrong?

Yes, and the design assumes it. That is why every output carries a confidence score, why weak numbers are routed to humans, and why every conclusion is traceable back to its source. A system that shows its evidence can be challenged in minutes; a black box cannot be challenged at all.

What happens when data is thin?

Thin data lowers the confidence score rather than silently widening the error. The estimate still exists, but it arrives flagged, with the gaps named, and it is treated as a starting point for human work rather than a finished answer.

Related reading: Glass-box AI: why every output traces back to a source, and How pricing models see a building.