How to evaluate AI underwriting software for commercial real estate

A commercial real estate underwrite begins with documents that disagree in format and purpose. The offering memorandum presents the seller’s case, the rent roll describes current leases, the T12 reports historical operations, and the buyer’s model expresses a future view. Useful software connects those layers instead of producing a detached chatbot answer.

At minimum, the system should identify periods and units, extract source facts, normalize them into the firm’s categories, flag conflicts, and place reviewed values into the model the team uses. The output should show which inputs came from documents, which are assumptions, and which remain unresolved.

AI can reduce transcription and make checks repeatable. It cannot decide how much conviction a team should place in a renovation premium, whether a market is becoming overbuilt, how a lender will react to a sponsor, or whether an execution plan fits the organization’s actual capability.

• Market and business-plan judgment
• Legal, physical, and environmental diligence
• Financing strategy and lender interpretation
• Materiality, recommendation, and final approval

Test each platform on the same completed deal and the exact workbook your team reviewed. Include messy account labels, missing months, ambiguous lease status, and conflicting OM claims. Clean samples prove little because the implementation challenge lives in exceptions.

Score source coverage, correction count, workbook preservation, time to a usable first pass, and whether another analyst can reproduce the result. Ask whether a model value traces to its source, whether missing evidence differs from a failed rule, and whether formulas and conventions survive the handoff.

Start with one asset class, one model family, and one clearly owned review process. Measure the time to a usable first pass, the number of corrections, the reasons for corrections, and whether principals trust the output enough to use it on live opportunities.

Expand only after the evidence loop is working. The goal is not to automate every edge case on day one. It is to create a reliable path from package to model to decision, then let firm-specific mappings and review rules improve with use.