Where AI Actually Helps in Financial Modeling — And Where It Still Falls Short

Every underwriting team we talk to is running the same experiment right now: pointing a large language model at a rent roll, an offering memorandum, or a set of comps and asking what falls out. Some of what falls out is genuinely useful. A lot of it is noise dressed up as insight. The difference matters more in commercial real estate and financial modeling than almost anywhere else in business software, because a model that looks right but is subtly wrong doesn't just waste time — it produces a return projection someone might actually invest against.

What AI actually helps with today

Strip away the hype and there's a real, narrow set of tasks where current-generation AI earns its place in a financial model:

• Extraction, not analysis. Pulling structured data out of unstructured documents — rent rolls, T-12 statements, offering memoranda — is the single most reliable use case. It's pattern matching against known formats, not judgment.
• First-pass anomaly detection. Flagging an assumption that's an outlier relative to the comp set (a cap rate two points off market, a rent growth assumption well above trailing performance) is something a model can do quickly and tirelessly, even if a human still has to decide what to do about it.
• Scenario generation at scale. Once the underlying model logic is sound, generating and comparing dozens of sensitivity scenarios is grunt work AI does well — it's the same calculation repeated, not a new judgment each time.
• Drafting, not deciding. Turning a finished analysis into a memo, or a memo into a set of talking points, is a genuinely strong use case that has nothing to do with the underwriting judgment itself.

The moment an AI-generated number gets treated as ground truth instead of a first draft is the moment it becomes more dangerous than not using AI at all.

Where it still falls short

The failure mode isn't that the model is obviously wrong — it's that it's plausibly wrong. A rent growth assumption that ignores 400 units of new supply coming online next door. A cap rate pulled from a "comparable" that isn't actually comparable once you know the submarket. These are judgment calls that require context a general-purpose model doesn't have and, more importantly, doesn't know it's missing. It will answer confidently either way.

There's also a reliability problem underneath the judgment problem: language models generate plausible-sounding numbers by default, not necessarily correct ones. Without a human checking a generated assumption against the actual comp set or the actual lease, an error doesn't look like an error — it looks like the rest of the model.

Our take

This is exactly why we built Kelvo Education's underwriting platform the way we did: AI-assisted where it's genuinely reliable — extraction, anomaly flags, scenario math — and human-judgment-first everywhere an assumption actually drives the return. We'd rather ship a tool that's honest about that boundary than one that quietly blurs it. It's the same principle behind Kelvo Consulting: prove the judgment works before you scale the automation.