RVP UnderwriterAI and underwriting
Can AI Underwrite a Real Estate Deal? Only If It Stays Out of the Math
Paste an offering memorandum into a chatbot and you will get a confident NOI in seconds. The question is not whether that answer sounds right. It is where the arithmetic happened, and whether you would get the same number if you asked again.
The short answer
Yes and no. AI is genuinely good at the reading: pulling income and expense lines out of a messy T-12, noticing that the rent roll and the OM disagree, flagging that a seller left out management. It is genuinely bad at the arithmetic, and an underwrite is an arithmetic task wearing a writing task's clothing. A language model predicts plausible text, and a plausible net operating income is not a correct one.
So the useful question about any AI underwriting tool, including a chat window, is never "does it use AI?" It is "where does the arithmetic happen, in the model or in code?"
How in-chat underwriting goes wrong
- The model does the math and does not tell you. When a figure is not clearly labeled in the documents, the model computes one itself and presents it with the same confidence as a figure it read. You cannot tell which is which.
- It is weakest exactly where it matters. In a peer-reviewed study of AI table extraction from 43 company annual reports, GPT-4 got 95.3% of inventory figures right but only 44.2% of EBIT figures (MDPI, Computers 13(10):257, 2024). EBIT is the closest public-company cousin of NOI. The authors concluded the accuracy was mostly not sufficient where high reliability is required.
- Same documents, different answer. Run the same package twice and the NOI can move. Setting temperature to zero does not fix it, because the way models are served (batching, floating-point order on the hardware) makes output vary from run to run.
- Formatting changes the answer. Every broker lays out a T-12 differently, and model accuracy is known to swing with formatting alone. In real estate the messy document is the normal case, not the edge case.
A newer model narrows some of these gaps. It does not change the architecture: a model generating a number is still generating it.
The fix the research points to
The same research literature also shows what works. In PAL, program-aided language models (Gao et al., arXiv 2211.10435), the model writes code and an interpreter runs it. That setup beat a far larger model reasoning step by step in its head by 15 points on grade-school math problems. A smaller model with a calculator beat a bigger model without one.
Engineers have a name for the alternative, sending a prompt and hoping the number is right: prompt-and-pray. The opposite is what we call Math-First AI Underwriting. Code does every calculation. The model is confined to reading and judgment, and every judgment is shown so you can overrule it. The model is allowed to have opinions. It is not allowed to do arithmetic.
Three tests for any AI underwriting tool
- Where does the arithmetic happen? In code, or in the model's text? Ask the vendor directly. A tool that only gives the model better instructions is still a prompt, just a tidier one.
- Does the same package give the same answer twice? Run it two or three times. If the NOI moves, the number is being generated, not computed.
- Can you see and argue with every judgment call? A management fee added, an expense the seller left out, a reserve, a reassessed tax. Each should come with a reason you can check and change, and the reasoning should export cleanly for a lender.
What we measured on our own product
We put RVP Underwriter through the second test. Before the underwriting rules moved out of the model and into code, ten fresh underwrites of one RV park's documents gave a normalized NOI anywhere from $138,200 to $169,400, which priced the park somewhere between $1.11 million and $1.91 million. That spread is wider than most negotiations. After the rules moved into code, five fresh underwrites of the same documents, with the property's research on file, matched to the dollar.
The lesson is not that AI is useless for underwriting. It is that the model was making the calls a careful buyer makes, like the right management fee or which season to annualize, differently every time. Once those rules lived in code and the model only read the documents, the answer held still.
How to use AI on your next deal
Use a chat assistant for what it is good at: explaining a term, drafting the list of documents to request from the seller, or pressure-testing your thinking. Do not take a value or an offer price from it. For the numbers, use tools where the math is computed: the free RV park valuation calculator, the NOI calculator and the cap rate calculator. And whichever tool you use, rebuild the seller's NOI yourself before you trust it, the way this guide walks through.
Put it to work
RVP Underwriter is built this way: code computes every number, the AI reads the documents and makes judgment calls you can see and overrule, and the same documents give you the same answer twice. Underwrite a deal free.
Questions
Can ChatGPT underwrite a commercial real estate deal?
It can read the documents and explain the method, which is useful. It should not be trusted for the numbers: models are least accurate on the figures closest to NOI, can compute an unlabeled figure without saying so, and can return a different NOI on the same documents. Use it to learn and to draft questions, and use a tool that computes the math in code for the value.
What is prompt-and-pray?
An AI engineering term for sending a model a prompt and hoping the output is right, with nothing checking it. In underwriting it means asking a chatbot for an NOI or a value and acting on whatever comes back.
What is Math-First AI Underwriting?
An approach where code does every calculation and the AI is confined to reading documents and making judgment calls, each shown with a reason so you can overrule it. The arithmetic is reproducible, so the same documents give the same answer.
Related guides
Part of the RVP Underwriter guide library: practical guides to underwriting and buying income property. Browse all guides, or read the complete free guide with a full worked example.