The first version of any professional AI tool extracts data. It reads a document and pulls out the fields: tenant name, lease term, base rent, expiration date. This is real work — the extraction has to be accurate, the fields have to be identified correctly, the output has to be structured. But it’s not what the professional needs. It’s the input to what they need.

What the professional needs is the calculation that comes after the extraction.

An underwriter reviewing a lease doesn’t need a list of expiration dates. They need to know what percentage of their total rent expires in each 90-day window over the next year, so they can assess rollover risk. An analyst reviewing a rent roll doesn’t need to know the in-place rent on each unit. They need to know the gap between in-place rent and market rent, aggregated across all below-market units, so they can quantify the upside they’re underwriting. A financial reviewer doesn’t need the raw expense line items. They need the expense ratio, the comparison to market benchmarks, and the flagged items that look like missing costs or one-time adjustments.

The extraction is the prerequisite. The calculation is the product.

This distinction matters a lot for how you build. A tool that extracts data and presents it as structured output is genuinely useful — it saves time and reduces transcription errors. But a tool that extracts data and then performs the downstream calculations that the professional would otherwise do manually transforms the workflow rather than just accelerating one step in it. These are different things, and professionals can immediately tell which one they’re looking at.

The calculation gap is also where domain knowledge becomes irreducibly necessary. Identifying that a lease has an expiration date is a reading task. Building the 12-month rollover schedule from a set of expiration dates requires knowing what rollover risk is, why 25% in a 90-day window is a flag, and how this number gets used in the underwriting model downstream. The calculation encodes the professional’s reasoning process, not just their data.

Tools that only extract are replicable — the extraction logic follows the document structure. Tools that calculate are harder to replicate because the calculations encode domain understanding. A competitor can read your extraction schema. They can’t easily reverse-engineer the judgment embedded in which calculations you chose to perform and what thresholds you chose to flag.

The extraction is the foundation. The calculation is the moat.

+++