How to Use AI to Compare Two Versions of the Same Real Estate Feasibility Model Published

Comparing two versions of a real estate feasibility model is more complex than a simple cell-by-cell difference. Excel can identify which cells have changed, but it cannot explain why those changes led to variations in key metrics like internal rate of return (IRR) or equity multiple. Combining a deterministic cell diff with an AI explanation layer can help generate a more insightful comparison.

There are typically three types of changes in real estate models: intentional modifications to inputs, deliberate adjustments to formulas, and unintended alterations due to human error. A raw diff treats all changes equally, but reviewers should focus on which changes affected assumptions, altered logic, and shifted outputs. AI excels at this interpretation task, provided it operates on data extracted from the original files rather than the files themselves.

To begin, each version of the model should be extracted systematically, distinguishing between formulas, cached values, and cell inputs. The Python code snippet provided uses the openpyxl library to load workbooks both with and without data, generating a comprehensive list of cells and their corresponding information. However, cached values are only available if the workbook was last saved in Excel, so this should be verified before relying on the extracted data.

Once the cells are extracted, changes need to be classified into four categories: inputs, formulas, hard-coded values, and value changes. This classification allows for a clearer understanding of what has been altered. Notably, changes where a formula has been replaced by a hard-coded value are particularly significant as they can have substantial impacts on the model's outputs.

After categorizing the changes, it is essential to distinguish between root causes and downstream effects. Since changes in one input can ripple through the model, affecting numerous other cells, isolating the primary alterations from the resulting downstream impacts is crucial. The AI model is then tasked with explaining these changes, focusing on the most plausible drivers of shifts in key outputs like IRR and residual land value.

The AI should be provided with structured data, including cell addresses and labels, to ensure accurate and relevant explanations.

To validate the AI's output, two sanity checks should be performed: verifying the existence of cited cell references in the change list and confirming that the direction of each claimed effect aligns with the observed data. This step ensures that the explanation remains grounded in the evidence. Additionally, running a manual check by reapplying changes one at a time and recalculating can provide further insight into each change's contribution to the output metrics, despite the time-intensive nature of this approach.

This pipeline, while effective, does have limitations. It may struggle with named ranges, array formulas, data tables, external links, and inserted rows, as these elements can complicate the parsing process. Macros that dynamically write values may also not be captured by the extraction process. Lastly, AI models can sometimes misinterpret conventions, necessitating careful review of the generated explanations.