AI for Real Estate: Why Humans Win

Real estate businesses carry complex transactions, shifting regulations and property-level facts that no model holds. Software still does a great deal of the work, and Parikh Financial builds the pipelines that do it. The question worth answering is which decisions those pipelines are allowed to make on their own.

The Risk in AI for Real Estate Is Unreviewed Output

A categorization rule that quietly misfires does not announce itself. It produces a clean-looking ledger, a plausible margin and a return that reconciles to nothing, and the first person to notice is usually an examiner. The failure mode is rarely a wrong answer on screen; it is a reasonable-looking answer nobody was assigned to check.

Getting the automated half right is a bookkeeping problem before it is an AI problem, which is what our bookkeeping service is built around.

Where the Forecast Needs a Person

A forecast extrapolates from what has already happened, which makes it reliable right up until something discontinuous occurs. In real estate the discontinuities are the whole story: a rate move, a rezoning, an insurance carrier leaving the state, a large employer closing. Three things the model cannot supply on its own:

• What is about to change: A model reads the comparables and projects the trend in them. It does not read the council agenda, the permit filings on the next block, or the letter from the insurer. Those arrive before the comparables move, which is the whole point of knowing them.
• Local Conditions: Zoning restrictions, permitting timelines and tenant demographics are local facts a national model does not hold, and they change what a property actually costs to run.
• Downside cases: A single projection is a guess with a decimal point on it. What a lender or an investor actually wants is the same model run at a higher vacancy, a higher insurance premium and a refinance at today’s rate, with the covenant headroom shown in each.

On building the forecast itself, see our financial forecasting service.

Where the Judgment Calls Actually Are

Categorizing a transaction is a matching problem and software is good at it. The calls below turn on facts and intent that sit outside the ledger, which is why they stay with a person:

• Capitalize or expense: The same $14,000 spent on a roof is a repair or an improvement depending on what it restored and what condition the building was in beforehand. Nothing in the bank feed records that.
• The cost segregation split: Deciding which components of a build come out of the 27.5 or 39-year bucket and into shorter-lived classes is an engineering judgment documented in a report, and it has to hold up years later. Our note on when a study earns its fee covers the arithmetic.
• Entity-level allocations: An inter-entity transfer looks like income to one LLC and an expense to another unless someone tells the system it is neither.
• The passive loss position: Whether a loss lands this year turns on participation hours and, for short-term rentals, on average guest stay. Those are facts about how the owner spent the year, not about the property.

Our piece on what an owner-operator actually has to decide each year walks through the same calls against a calendar.

Where the Line Falls

The workable arrangement is the one the tools were designed for: automate the repetitive, rules-based work, and keep a person on anything that turns on intent, on local facts, or on a position that has to be defended later. Both halves are our work: the pipelines that move the data and the review that decides what it means arrive in the same engagement.

If you want to see where that line sits in your own books, schedule a call with us.