Financial Modeling in the Era of AI – Watch Me Build
In a recent post, I made the case that AI and financial modeling follow the same logic as long division and calculators. Learn the mechanics first. Once you’ve got them, the tool multiplies your output. The calculator only helps if you understand what it’s computing.
That post was the argument. This one is the demo.
What you’re about to watch is me loading a real offering memorandum, a trailing twelve-month income statement, and a rent roll into an AI Skill built for the A.CRE Apartment Acquisition Model, then watching it work through first-pass underwriting on a 134-unit Phoenix apartment deal. The whole thing takes about 20 minutes, and most of that time I’m talking through what it’s doing rather than waiting on it.
- You might also enjoy: The argument for learning fundamentals before using AI tools in CRE: Financial Modeling in the Era of AI
- Related resource: The AI Skill used in this demo: Apartment Acquisition Model AI Skill
Financial Modeling in the Era of AI – Watch Me Build
A bit of background on why this matters
Michael Belasco and I have been building real estate financial models for about 13 years. The medium has changed, but the logic hasn’t. You start with inputs, run them through a series of calculations, and produce outputs. That process plays out in a spreadsheet the same way it played out on paper thirty years ago. The math doesn’t change because the tool does.
What changed with templates was speed. Rather than building from scratch every time, you populate a pre-built structure. That’s useful. But it doesn’t excuse you from understanding what the template is doing. If you can’t build the model yourself, you shouldn’t trust the output. Your boss probably shouldn’t either.
AI shifts the calculation again. For the last few years, every time a new large language model came out I’d test it on a financial modeling task. The pattern was consistent: it would produce errors faster than I could catch them, and fixing the output took longer than just doing the work myself. That’s not a useful tool.
That has changed. With the right instructions and the right model, AI can now get through a first-pass underwriting faster than I can, and the time required to validate the output is less than the time it would take me to build it. At current benchmarks, I’d call it roughly a 2x speed improvement on standard modeling tasks. That’s worth paying attention to.
What’s in the skill
The skill is a folder of structured instructions that tells the AI how to use the model. Not just “fill in the inputs” but the full methodology: what the inputs are, how they relate to outputs, what order to work in, and why. The skill also includes our modeling conventions (blue cells are inputs, red flags a changed value, green indicates a cross-sheet link), a full input map, an output reference, and a role framework.
That last piece is worth explaining. Who you are when you’re underwriting a deal changes what you’re looking for. An acquisitions analyst at a sponsor sees the model differently than an LP re-underwriting the same deal, or a lender sizing debt coverage. The skill understands those roles and adjusts its focus accordingly, including how aggressively it interprets assumptions. A broker representing a seller is trying to tell the rosiest story the market will accept. A lender’s upside is capped, so conservatism matters more. The skill accounts for that.
The skills themselves are built using a separate tool we developed: an Excel Model Skill Generator that our team uses to translate a model’s logic into AI-readable instructions. Building skills has become its own discipline, and it’s one we’re training our A.CRE Accelerator members to handle.
What happened in the demo
The deal is Avenue 19, a 134-unit apartment property in Phoenix, Arizona. I pulled the skill into Claude in Excel, then dragged in three files: the offering memorandum, the T12, and the rent roll. No other setup.
From there, the AI read the skill, read the deal documents, and started asking clarifying questions before touching anything. What role am I in? How should it approach rents: in-place, market, or a blend? How should it size debt? I answered those, including a direction to use the A.CRE Intelligence Hub for live agency debt pricing rather than the OM’s assumed terms.
The AI queried the Hub, pulled a current Freddie Mac rate sheet, and sized the loan accordingly. Then it began populating inputs: purchase price, loan amount, occupancy, hold period, rent roll, expenses. Each cell it changed got a brief note explaining the source and logic. “65% LTV per Hub agency multifamily rate sheet, Freddie K SB deal, May 13th.” That kind of specificity is what makes validation faster. You’re not reverse-engineering where a number came from.
When it finished the first pass, it didn’t just report the returns. It flagged what was broken. The deal doesn’t clear an institutional return threshold at the asking price. More specifically, the OM’s headline cap rate used a year-one proforma NOI that was roughly 50% above in-place actuals, and excluded the management fee. Running actual T12 expenses against in-place rents produces a going-in cap of about 4.54%, not the 5.47% the OM advertised. That gap does a lot of work. The AI caught it and said so.
That’s the output of a first-pass underwrite. Not a final answer, but a clean starting point that surfaces the right questions.
What this isn’t
AI responses contain errors. That’s not a disclaimer, it’s a fact about the current state of the technology. The same expectation you’d bring to a junior analyst’s first pass applies here: verify before acting on anything.
More to the point: this skill is not a shortcut for people who haven’t learned to model. If you can’t build the Apartment Acquisition Model yourself, if you can’t look at the output and know whether the inputs are rational, you shouldn’t be using the skill. The long division analogy from the companion post is not rhetorical. Mrs. Jackson didn’t hand out calculators until the class had proven it could do the work by hand. The same rule applies here.
The Accelerator exists because that foundation matters. Learn to build first. Then use the skill as your calculator.
Where this is going
The skill architecture has a longer arc than a single demo. Right now, the person who builds the model is probably the person who builds the skill. That’s a specialized role, and it’s genuinely valuable in any shop that does volume underwriting. At some point, every organization that processes deals will either have internal skills or will buy them. We’re in the early window of that transition, and the people building skills right now are the ones setting the standard.
We’re rolling out skills across the A.CRE model library and embedding them into the Accelerator curriculum. The training path will follow the same arc described in the companion post: learn the fundamentals, build the case study by hand, watch it done by hand, then work alongside an AI that knows the procedure.
The calculator has arrived. It’s more useful than the templates were at their best. And just like the templates, it still requires someone who understands what it’s computing.
If you have questions, comments, or want to say hello, feel free to reach out via email or LinkedIn.
Frequently Asked Questions about Financial Modeling in the Era of AI: Watch Me Build
What is an AI Skill for a financial model?
An AI Skill is a folder of structured instructions that teaches an AI how to use a specific financial model. It includes the model’s input map, output logic, modeling conventions, and a methodology guide so the AI understands not just what to fill in, but how each input affects the outputs. Think of it as the operating manual the AI reads before touching the spreadsheet.
Which models does Claude use in the demo?
The demo runs inside Claude in Excel, using the A.CRE Apartment Acquisition Model and its corresponding AI Skill. The A.CRE Intelligence Hub is connected via MCP to pull live agency debt pricing from a current Freddie Mac rate sheet. The same workflow works in Claude desktop, Claude Cowork, ChatGPT, Gemini, or Copilot, since all of those now support skills and MCP connectors.
What does the AI actually do with the OM and rent roll?
The AI reads those documents and uses them to populate the model inputs. It parses the T12 for in-place income and expenses, reads the rent roll for unit mix and current rents, and cross-references the OM for deal-specific context like purchase price and occupancy history. Each input it changes gets a note explaining where the number came from.
How long does the first-pass underwrite take?
In the demo, the AI completes a first-pass underwrite of a 134-unit Phoenix apartment deal in roughly 20 minutes, and most of that time is conversation rather than waiting. Spencer estimates this roughly doubles the speed of a standard modeling task compared to doing it manually, while still producing output that requires validation before acting on it.
What does the role framework do?
The role framework tells the AI whose perspective to apply. An acquisitions sponsor might run conservative assumptions to stress-test a deal. A broker representing a seller is looking for the most supportable aggressive case. A lender focuses on debt coverage and downside scenarios. The AI adjusts its assumption-setting and output emphasis based on the selected role.
Can the AI use live market data?
Yes. In the demo, the AI connects to the A.CRE Intelligence Hub via MCP to pull a current Freddie Mac agency rate sheet for debt sizing. The Hub updates weekly with actual priced deals, so the debt assumptions reflect current market pricing rather than figures from the OM.
How did the AI catch the cap rate discrepancy?
After completing the first-pass inputs, the AI ran actual T12 expenses against in-place rents and compared the result to the OM headline. The OM used a year-one proforma NOI that was roughly 50% above actual in-place income and excluded the management fee, which inflated the advertised cap rate from approximately 4.54% to 5.47%. The AI flagged this discrepancy and explained what was driving it.
Who should use this skill?
The skill is built for real estate professionals who already understand financial modeling. At minimum, you should be able to build the Apartment Acquisition Model from scratch, understand how the inputs produce the outputs, and evaluate whether the AI’s first pass is rational. If you can’t do that, the skill is not a shortcut. Learn the model first, then use the skill as your calculator.
Will AI replace real estate financial modelers?
No. The AI handles first-pass input population and initial analysis, but validating the output, refining assumptions, and making judgment calls still requires someone who understands the model and the deal. What changes is the volume of deals a skilled modeler can process, and the baseline quality of the starting point they hand to a reviewer. The fundamentals matter more now, not less.
How do I get started?
If you already know how to build the A.CRE Apartment Acquisition Model, you can get the AI Skill directly at adventuresincre.com. If you want to learn the modeling fundamentals first, the A.CRE Accelerator teaches you to build models from scratch before introducing AI tools. More skills for other models in the library are rolling out over the coming months via AI.Edge.
