COMMERCIAL REAL ESTATE · UNDERWRITING AND DILIGENCE
Automate underwriting and diligence in commercial real estate
A data room is two hundred files, and the pro forma always works because the assumptions were chosen so that it would. OutcomeCatalyst turns a scanned T-12 into general ledger detail, tests market rent against leases that actually signed rather than asking rents, checks each assumption against the deals you already closed, applies the gates your investment committee uses, and drafts the memo with every figure traced to the page behind it.
Underwriting and diligence, in plain terms
A data room is two hundred files and the answer is spread across a scanned T-12, a rent roll tab and sixty page leases. The broker's pro forma always works, because the assumptions were chosen so that it would. This tests each of those assumptions against a source document and produces the memo, not just a variance list.
From a data room to a tested underwriting, in four steps
Your model stays your model. This tests the inputs before they reach it.
Why underwriting capacity decides what you buy
A pro forma is a marketing document that has already survived its own review. The assumptions were selected to produce a number the seller wants, which is not dishonest so much as structural. Every OM in the market works on its own terms.
What is unusual is that the evidence to disprove it typically sits in the package the broker themselves provided. A T-12 scan supporting different expense numbers than the summary page is the normal case rather than the exception, because nobody expects the scan to be read line by line.
The real constraint is that underwriting properly takes weeks of analyst time, so it gets spent on the deals that already look best. That is circular. The deals that look best are the ones whose marketing worked, and the ones worth a closer look are often the ones nobody had hours for.
Consistency compounds the problem. When two analysts underwrite two deals in two different spreadsheets, the outputs are not comparable, and a committee ends up choosing between two arguments rather than two assets.
Agentic AI in commercial real estate, without the hand-waving
Three words get used interchangeably by vendors and they do not mean the same thing. The difference decides whether the work gets done or just gets read.
Most agentic AI pilots in underwriting stall for a reason that has little to do with the model. An agent asked whether a pro forma is credible needs the rent roll, the T-12, the leases, the comps and your own variance history on similar assets. That sits in a data room, an accounting system, a market subscription and somebody's spreadsheet. With no path between them the agent accepts the broker's numbers, which is the one thing it must never do.
The agents here are deliberately narrow. Each has one job, a defined set of sources it may read, a written standard to check against, and a person who approves before anything leaves the building. That is what makes them safe to run against real money, and it is why they survive an audit.
The data layer agentic AI actually needs
Every workflow above runs on one layer. Building it is most of the work, and it is the part nobody demos.
This is the part most vendors skip, because a layer does not demo well. It is also the reason one piece of infrastructure carries origination, underwriting and reporting at once, instead of three tools each rebuilding the same context badly and disagreeing with each other.
It compounds. The second workflow stands up faster than the first and the fifth faster still, because the entities, the ontology and the connectors already exist. Most of what a new workflow needs is already in the layer.
What else runs on the same layer
Once the layer exists these are weeks of work rather than months, because they read the same resolved entities.
Built for the stack a diligence team already runs
These are the sources referenced in the workflow above. Document formats vary by data room and are handled as they arrive.
Questions investment committees ask first
Does this replace ARGUS?
No, and it is not trying to. ARGUS models the assumptions you give it and does that well. The problem sits upstream, in whether those assumptions match the record. This tests each input against comps, trailing statements and public records before it reaches your model.
How reliable is extraction from a scanned T-12?
Confidence is scored per line item and anything low is surfaced for human confirmation rather than used silently. Every figure links back to the page it came from, so an analyst verifies in seconds instead of re-keying.
Does it make investment decisions?
No. It produces a tested underwriting and a drafted memo with sources attached. The committee decides. Every variance traces to a comp, a statement line or a filing, so the reasoning can be argued with.
What if our underwriting standards are not written down?
Then that is the first piece of work, and it is worth doing regardless. In practice most of the standard already exists in how the committee actually behaves, and the exercise is to make that explicit rather than invent it.
How is this different from an offshore analyst team?
An offshore team gives you more hours at the same consistency risk, because two people still read two leases two ways. The difference here is that the same standard applies to every file, and every figure keeps a link to its source.
Can we test it against a deal we already did?
That is the usual starting point and the one we recommend. Run it against a deal you underwrote by hand and compare. It is the fastest way to see where it agrees with you and where it does not.
What about the leases on assets we already own?
Same layer, different workflow. Every lease is read for expiry, escalation, option and recovery terms, so exposure is dated rather than discovered at renewal.
AI underwriting for commercial real estate: common questions
What does AI underwriting test?
Each assumption in the pro forma against a source document. Market rent goes against signed lease comps, operating expenses and economic vacancy against the trailing statements in the package, real estate taxes against county assessment, and the exit cap against recent trades in the same submarket. Every variance cites the page it came from.
Where does the data come from?
From the data room and the subscriptions you already pay for. The T-12, rent roll and leases from Box or SharePoint, signed comps from CompStak, trades from CoStar, taxes and liens from county records, and precedent from your own closed deals.
How is this different from ARGUS?
ARGUS models the assumptions you give it and does that well. The problem sits upstream, in whether those assumptions match the record. This tests each input before it reaches your model, so ARGUS is running on numbers the evidence supports rather than numbers the seller chose.
