KeystoneUnderwriting agent A week to a day
Rebuilding a sponsor's model used to take an analyst a week, and now it is done the day the package arrives.
The ask
Galapagos Capital's US real estate desk in Miami judges every deal through its house view: decades of how this desk underwrites, carried in a few analysts' heads and the spreadsheets they rebuilt by hand. Every sponsor sends a different model, built a different way, and an analyst spent a week or more rebuilding each one to reach the case the house believes. That judgment did not scale, and it walked out the door every night.
Four weeks in, the real problem was not the documents.
What we came to build
We came to read sponsor packages. The hard part sat somewhere else: capturing the house view itself, the instinct that tells an analyst a number is wrong, and giving it to something that runs on every deal.
“Auditing a sponsor's spreadsheet was so complex. We were fighting to take a course at NASA just to understand it.”
The models are enormous
A sponsor's model is a web of linked formulas, not a clean table, and it hides the purchase price inside a calculation or points at a sheet that never arrived. The agent rebuilds it, proves which lines actually produce the return, and lists what it cannot resolve instead of guessing.
“Take whatever Excel comes in, however it comes, break it apart and rebuild it in a way that is more comparable with what we have. This is exactly it.”
The agent carries the judgment, not the arithmetic.
The house view, encoded
The desk's own way of underwriting was rebuilt alongside the desk. The agent names the cell, the direction and the reason an analyst would give; exact calculation does the math, never the AI. Every figure it shows is checked on its own, or held back as unknown.
The judgment that lived in their heads now runs on every deal.
Why a person still signs
The desk's own research found that asking an AI to grade its own numbers caught fewer than one bad answer in four, and confident writing raised the score on wrong math. So the AI never decides a figure. It carries the reasoning and the source behind every input; the desk signs.
The agent found what the humans could not.
What it caught
A sponsor's model is written to sell the deal. The agent reads it against what the building really spent over the last twelve months, and on one deal it surfaced what the desk's own finished model had carried through: four expense lines under real spend, insurance by a factor of two and a half. Read against the real numbers, the return fell by more than half.
Inside the formulas
Then it audits the arithmetic. It opens every formula in the workbook and finds the total that does not tie, the number typed inside a formula, the rule buried in a rent line. It writes down the sponsor's calculation, why it is wrong and the fix, and a person approves it. No analyst opens every cell on every deal. The agent does.
“Using the same assumptions, the sponsor's spreadsheet and the system match. That's it. Great. Excellent.”
What the desk decided
The desk used to audit each sponsor by rebuilding the model in its own spreadsheet. Once the agent runs in full, that spreadsheet goes away, and the agent audits every sponsor that arrives.
“This platform is far more advanced, and I don't think it needs any input from our scorecard. Quite the opposite.”
End to end
A sponsor package enters on the left. A committee memo leaves on the right.
Intake
- Offering memoranda, sponsor models, market reports
- File type read from its bytes
- Stored by SHA-256 fingerprint, never overwritten
Act one
The sponsor
- Reader and locator: meaning and cell address
- Source check: code reads the original cell
- Gap reviewer
- Input investigator: units, period, scenario
Act two
The house
- Market reader: the desk's own reports and public data
- Reconciliation by precedence: contract, history, measurement, sponsor
- The math by exact calculation: NOI, IRR, cash flows, limits
Review
- Analyst, risk, committee, admin
- Signing needs rank at or above the owner
- A correction becomes a cited fact and a new snapshot
Delivery
- Underwriting
- Sponsor against house, line by line
- Workbook in the house format
- Committee memo
Every call to the AI goes through one guarded door, with personal data stripped first. Nothing else can reach it.
One citation per fact, held by the database.
Every figure audited by code, or blocked as unknown.
To rebuild a sponsor's model and reach the case the house believes, on every deal.
The judgment that took a week in one analyst's head reaches every package the day it arrives. The house view stops walking out the door at night, and the analyst goes back to the call only a person can make.
Book a callGrowthOS
The Commercial Operating System for Law Firms
Next file