Not Everything Should Be AI. Most of It Should Be Wired.
AI is not the answer, but it is part of the solution.
The most important thing about AI is not what it does. It is what it enables. Most work sold as "AI work" does not require AI. It requires SQL, Python, and ETL. Words most teams heard once and nobody explained, now repackaged as "AI" and "agents."
Which parts of the work do not need AI?
The technical word for it is deterministic: exactly one right answer, reached the same way every time. Call it wired. Here is where wired work belongs.
Extracting recurring financial reporting (Python script). Your property manager sends the same operating statement, same format, same day every month. A script pulls it, maps the accounts to your chart of accounts, and loads it before anyone opens a laptop. A person builds that crosswalk once. After that the script applies it, and kicks out any account it has not seen before instead of guessing.
Populating numbers in reports (SQL stored procedure). The quarterly investor letter, the lender package, the Monday pipeline sheet. Same fields, same sources, every time. At an institutional pref shop I worked at, those numbers came out of a stored procedure inside the database I architected for the acquisitions team, with the report bound straight to it. It was not an analyst's Tuesday.
Extracting data from your financial models (VBA or an Office.js add-in). Purchase price, going-in cap, exit cap, DSCR, five-year IRR. Those cells already exist. Name them once, the same way in every file, and an add-in reads them and pushes them to the database with one click. The naming is the work. The click is free. That is how a pipeline builds itself, out of work the team already did.
Why don't these require AI?
Because the answer already exists. It sits in a cell or a file that arrives in the same shape every month. Nothing is being decided. The number only has to move from point A to point B without a human retyping it.
What happens when you point AI at wired work?
You still get an answer. Ask for NOI off a T12 and it hands you one, wrong because a partial-year line got read as annualized. It didn't hallucinate something absurd. Absurd is easy to catch. It produced something plausible.
Close is worse than wrong. A wrong NOI clearly sticks out. A plausible one survives the meeting and lands in the lender's inbox. It passes a gut check because nothing about it looks wrong. Then the lender's analyst asks what is in that expense number. Accuracy secures capital, and you cannot defend a number you cannot trace to a source. It is a guess, and what it costs you when someone scrutinizes it is not the number. It is your credibility.
The T12 lines are not up for debate. What each account reported is what it reported, and that had to move from a PDF into a cell, unchanged. Which of those lines belong in NOI is a judgment call. It should be your rule, applied the same way on every deal, not a model's read on this one.
Where does AI actually earn its place?
The other half is where I spend my own money:
Pulling structured data out of unstructured deal documents. Rent rolls, T12s, leases. No two of them arrive in the same shape: merged cells, footnotes, four different names for "concessions," a unit mix only its author understands. There is no rule to write because there is no rule being followed. This is where an LLM comes in, and it is what we built: a rent roll extraction tool that reads the file an analyst would otherwise retype by hand. The extraction is AI. The destination is not. A person checks it against the source before anyone underwrites off it.
Market research. What is delivering in the submarket is a data pull, not AI. Which economic drivers will justify strong rent growth is a question you can hand to AI.
Drafting a summary on top of numbers a database already computed. The math is wired. The paragraph explaining the math is not.
First-pass written work a human edits. The investor update. The market section of a memo.
How do you tell which list the work belongs on?
Here is the test. If someone at your firm could walk to a file and point at the answer, do not ask a model to generate it.
People are looking at AI adoption backwards. They are hand-keying the work that has one right answer, and asking a model to produce the numbers that have to survive an IC.
Why does everything get sold as AI now?
AI gets thrown around as the solution because it simplifies technical tools that already existed, that nobody explained, and that operators never understood. Everything on that first list was buildable before ChatGPT was a chatbot. None of it existed at most shops because the people who knew exactly what they needed could not describe it in a language that would build it, and the people who could build it did not know the difference between in-place, stabilized, and exit NOI.
That is the gap AI closed. Not capability. The ask.
And once you can ask for anything in plain English, you stop noticing which asks were supposed to go to a script.
Knowing which list the work belongs on is what makes the numbers hold up in front of an IC, a lender, and an LP.
If you want to see the wired half running under a real pipeline, I walk through one in 15 minutes. No data required.