Money needs an engine.
A convincing explanation of a number is not enough. I want to know how the number was calculated.
Separate reading from calculating.
A model can help locate a rent figure in a document or explain what a lease provision means in context. The extracted value still needs a source and a check. Once it becomes an input to a financial model, the calculation should follow explicit rules.
Rent growth, recoveries, leasing costs, and discounting belong in an engine that can be tested. The same inputs should produce the same outputs.
Expose the assumptions.
In my Rust valuation prototype, the rent roll and annual projection are visible together with the valuation assumptions. Changing an input changes the result through the engine. That makes it possible to inspect which assumption moved the answer.
The prototype is deliberately limited to an annual model. Naming that limitation is more useful than presenting a precise-looking output without explaining what it leaves out.
Give AI a useful job.
There is plenty for a model to do around the calculation: organize a source document, flag a missing input, or explain the effect of an assumption. Keeping arithmetic separate gives those tasks a firmer foundation.