Where does this thing go

Most writing about AI in real estate assumes the constraint is analysis. It never was. An analyst can underwrite a deal in a day, and a portfolio rent comp survey takes an afternoon. What eats the week is reconciliation. Why the property management system says one thing and the accounting export says another. Why a rent roll pulled Tuesday does not tie to a rent roll pulled Friday. Why an expense line moved, and whether the move is real or a coding error.

So the systems worth building are not the ones that produce answers faster. They are the ones that hold data consistent across platforms and surface a discrepancy before it reaches a decision. Less impressive to demo. Far more useful.

The second thing people get wrong is adoption. Something fails when nobody changes what they do on a Tuesday. The regional manager keeps her own spreadsheet. The weekly report still gets rebuilt by hand because someone does not trust the automated version. That is not a modeling problem. It is a question of whether the thing fits the shape of a job that already exists, and most of the people building these tools have never sat inside that job.

The hardest part of the design was deciding what the thing is not allowed to do. Everything that mattered turned out to be a refusal. It writes to input cells and never to a formula. It never asserts a number it did not read back out of a recalculated file. Every assumption that drives the return gets tagged for a person instead of chosen. Those rules exist because anything that quietly picks an exit cap has hidden the only decision that mattered, and it will be wrong in a way nobody catches until the disposition memo.

Which points at the real division of labor. Assumptions are judgment, and judgment is where returns come from. Exit cap. Rent growth. Downtime. The credit you actually give a renovation premium in a submarket where the last three comps leased soft. You can be told what those assumptions produce. You cannot be told whether to believe them. That gap does not close with better software. It closes when someone is accountable for being wrong.

My read is that this work belongs to people who can build the machinery and also defend the assumptions inside it. Not a technologist who hands off a tool. Not an analyst who consumes one.