A.CRE
, , , ,

Episode 10 of Multipliers: Modeling Like a Deal Seasoned Pro in the Era of AI

For years, AI real estate financial modeling was a promise that kept arriving too early. Every new model, every new tool, produced more errors than it saved time. You would spend longer validating the output than you would have spent building the thing yourself. That is no longer true.

Something shifted in the last few months. The tools are now fast enough, accurate enough, and connected to enough real-time data that a trained analyst can run a first-pass underwriting on a live deal, in a real model, in under 20 minutes, while having a conversation with two colleagues. That is what happened on this episode. Spencer Burton did it live, on screen, with an actual offering memorandum and a real T12, in front of Michael Belasco and Sam Carlson, and the output was not just directionally correct. It caught a specific problem with the broker’s underwriting that most buyers would have missed.

The caveat matters as much as the capability. The skill works because Spencer built it. The output is trustworthy because Spencer can evaluate it. The calculator is only useful to the person who already knows long division.

In this episode of the Multipliers podcast, Spencer walks through what A.CRE has built, why it took thirteen years of modeling to make it possible, and what the arrival of AI real estate financial modeling means for the professionals who are ready for it and the ones who are not.

Episode 10 of Multipliers: Modeling Like a Deal Seasoned Pro in the Era of AI

All three co-founders are back this week. Michael joins from Washington, where he spent time at Olympic RV Park getting operations primed for peak season: 70 trees going in, e-bikes on order for the trail adjacent to the property, and a music festival sponsorship locked in for the next two years. Sam is in his usual place. Spencer is fresh off a ULI Nashville council session where he spent ninety minutes teaching a room of real estate professionals to build a live data app with AI, and he arrives at this episode with something he has been working toward for over a month: a financial modeling skill he wants to demo on air. The episode connects back to Episode 9, where Michael’s RV park served as the case study for what AI-powered operations look like when someone does the work. This one brings that same principle into the discipline that sits at the core of what A.CRE has always been: real estate financial modeling. Spencer runs CRE Agents and teaches at UNC Kenan-Flagler. The demo he runs on this episode is not a proof of concept. It is a working tool that A.CRE Accelerator members are already receiving.

Why This Episode, Why Now

Spencer has been asking the same question every time a new large language model releases for the past several years: is it here yet? By “here,” he means a specific threshold: can AI help with real estate financial modeling in a way that saves more time than it costs to validate? For years, the answer was no. The models would hallucinate inputs, produce errors at a rate that made review more expensive than just doing it manually, and generally fail at the kind of methodical, structured numerical work that underwriting requires.

That changed. Spencer is careful about when he says this kind of thing, because the history of AI announcements is littered with premature declarations. But on this episode he is direct: for standard, repeatable modeling tasks, AI now roughly doubles the speed of underwriting. The first-pass work, parsing the OM, reading the T12, populating the rent roll inputs, sizing debt against current market rates, can be done in parallel across multiple deals simultaneously while the analyst does other work. The analyst’s job shifts from data entry to validation, refinement, and judgment.

The timing of the episode also connects to something Spencer has been building for a month alongside the A.CRE team. They have been systematically creating AI skills for models in the A.CRE library, starting with the simpler ones and working toward more complex structures. The apartment acquisition model he demos here is an early iteration, and the lesson from building it is as important as the demo itself: creating a skill that actually works requires encoding not just the inputs and outputs of a model, but the full methodology behind it. That is not a task you can outsource to someone who does not understand the model. It is a task that requires the person who built the model to teach the AI how to use it.

The ULI Nashville workshop, where Spencer built a live FEMA data app with a room full of real estate professionals in ninety minutes, is the other piece of context. The exercise was designed to give people the experience of seeing what is possible with AI when used correctly. Michael had a version of that unlock earlier when Spencer helped him build the backend for the RV park platform. The live demo in this episode is an attempt to give the same unlock to everyone watching.

Episode Highlights

Here are the themes that stood out.

1. Olympic RV Park: Prepping for Peak Season

Michael spent the week before this episode in Washington at Olympic, which approaches its one-year anniversary beating 2026 projections. Prepping for peak season means a lot of things at once: seventy trees going in to address the bare landscape that comes with a brand-new development, work underway in the clubhouse, a two-year sponsorship of a stage at the Wanda Fuka Music Festival that will bring thousands of people to the area, and a fleet of e-bikes on order to capitalize on the Olympic Discovery Trail that runs directly past the property.

The e-bike decision is a good illustration of how Michael thinks about the park as a business. The trail goes the length of the northern peninsula. A guest with an e-bike can reach Port Angeles and back without a car. The bikes are a rentable amenity, priced separately from the site, and other parks have reported strong ROI on the investment. It is a revenue line that also deepens the guest experience and creates a reason to book the park over a competitor without the same access.

The broader lesson here is one that runs through every episode that touches the RV park: Michael is operating this asset the way a founder operates a startup, not the way a passive investor manages a property. Every amenity decision is evaluated for ROI, every marketing tactic is tested and measured, and the asset is being actively improved based on what was learned in year one. That orientation is what makes the AI tools useful rather than decorative.

2. The ULI Nashville Workshop: Building a Live App in 90 Minutes

Spencer’s ULI council session was a 90-minute workshop rather than a traditional speaking engagement. The exercise: build a data application with AI that pulls in FEMA flood risk data to help underwrite an insurance line item, useful either as part of a quarterly valuation, a listing package, or a standard underwriting workflow. The room ranged from people who had never opened ChatGPT to professionals already running multiple Claude Code instances simultaneously.

The setup Spencer used is the same two-AI workflow he has described across several episodes: Claude in chat mode as the prompt engineer and architect, and Replit as the coding agent and deployment environment. The key insight he emphasized to the room is why those two roles need to stay separate. The short version is both cost and context quality. Replit prices on token usage, and every prompt in an agent context triggers the system to read all related files, which adds up quickly. More importantly, the coding agent produces better output when its context window is clean and focused exclusively on the code, not cluttered with the exploratory brainstorming that preceded it.

What Spencer was really trying to give the room was not a tutorial. It was the experience of the unlock, the moment when someone who has been abstractly aware of AI’s potential actually sees it do something real in a domain they understand. Michael had that moment when they built the RV platform backend together. Spencer has been chasing ways to give it to others ever since, because he believes that experience, more than any explanation, is what converts skeptics into builders.

3. Why You Need a Separate Prompt Engineer

The context window question came up multiple times at ULI and again during the episode. Michael’s admission that he is a “habitual context clutterer” made it concrete in a way that will resonate with most people who have spent time working with AI.

Spencer’s one-piece-of-paper analogy is the clearest way to understand what a context window actually is. The AI knows only what is on that piece of paper at any given moment. Everything it knows about you, your project, your preferences, was written on that paper at the start of the conversation. As the thread progresses, more gets added. At some point, the paper fills up, and in basic implementations, the oldest content starts getting erased. If the early content included the critical objective of the thread, you now have an AI that has forgotten what it was doing.

The practical consequence is that any content you add to a thread that is unrelated to the ultimate objective of that thread reduces the quality of every subsequent output. The AI has to carry the irrelevant material alongside the relevant material, which crowds out the context it actually needs. Spencer’s solution is structural: keep the exploratory, brainstorming, and prompt-engineering work in a separate Claude chat thread, and bring only the clean, finalized instructions into the coding agent. One role thinks. The other builds. Neither is burdened by the other’s work.

Michael asked what most people working with AI are quietly dealing with: what do you do when you need a quick tangent in the middle of a build? Spencer acknowledged it is a real friction point and noted that ChatGPT’s ability to fork a conversation into a separate thread without polluting the main context is something Claude does not yet do. For now, the discipline has to be manual: recognize when you are about to go off-topic, open a new thread, and come back.

4. Living Skills: Giving AI the Experience It Needs

One of the more conceptually rich moments in this episode is Spencer’s description of what he calls a living skill. The basic concept of a skill is already established in the Multipliers series: a folder of files that encodes a methodology or body of knowledge, injected into an AI’s context when you need it to perform a specific task. A living skill adds a layer that changes the nature of the tool over time.

The addition is a log file, a living document within the skill that gets updated at the end of every session. Spencer builds the instruction directly into the skill.md file: at the end of every discussion, update the skill based on what we did. The AI prompts for confirmation, and when you say yes, it appends to the log and optionally updates the relevant reference files, so that the next time you use the skill, it has context from every previous session. The skill accumulates experience.

Michael made the connection that made this tangible: it is the digital version of experience. When you start with a skill, you encode what you know at that moment. But you always leave out nuance, edge cases, the specific situations you have not yet encountered. As you use the skill and run into those situations, the log fills them in. The methodology improves because the person using it is improving, and the skill reflects that in real time.

Spencer extended this to his daughter Lexi, who is using a skill to prep for the SAT. At the end of every session, the log records what she worked on. When she takes a practice test, the results get added as a reference file. The AI can access her test history and adjust the difficulty and focus of the next session accordingly. It is personalized tutoring that gets more personalized the more she uses it, because the skill itself learns from the sessions.

5. The Live Demo: AI Underwrites an Apartment Deal

The live demo is the centerpiece of the episode and worth understanding in detail, because what Spencer does in under 20 minutes represents something genuinely new in AI real estate financial modeling.

The model is the apartment acquisition model from the early A.CRE library: annual periods, three primary input tabs (property summary, rent roll, expenses), an investor returns tab, and output tabs. Simple enough to be a clean test case, complex enough to have a meaningful number of inputs that require judgment. Spencer injects the financial modeling skill into Claude in Excel, then uploads three documents: the offering memorandum, the T12, and the rent roll for an actual deal in Phoenix (134 units, Avenue 19).

The AI reads the skill, reads the documents, and begins asking clarifying questions: what role is the user playing, how should rents be modeled, should debt be sized from the OM or from market data? Spencer instructs it to use the A.CRE Intelligence Hub (available to AI.Edge members) to size the loan, which triggers the system to query a live Freddie Mac rate sheet for current agency multifamily pricing. This is not a static assumption. It is a real-time data pull from actual deal pricing, incorporated into the underwriting automatically.

The inputs start populating, with the AI adding comments to cells as it goes: “65% LTV per hub agency multifamily rate sheet, Freddie KSB deal from May 13th.” “Per OM T12 occupancy trended 90 to 93%.” It is documenting its reasoning in the model, the way a good analyst would.

The conclusion is instructive. At the asking price of 18 million, the deal does not clear an institutional return threshold. The AI identifies why: the broker’s headline 5.47% cap rate uses a year-one pro forma NOI nearly 50% above the actual T12 in-place figures, with the management fee excluded. The AI caught a specific and consequential piece of broker underwriting that the market was being asked to accept at face value. Spencer confirmed he remembered the deal and the analysis was correct.

The caveat Spencer opens with is worth restating: the skill is built for real estate professionals who already have a deep understanding of financial modeling. If you cannot build the model yourself, you should not be using the skill to underwrite deals. The output requires someone who can validate it, catch errors, and make judgment calls when inputs are ambiguous. The skill is the calculator. The Accelerator is still the long division.

6. The Skill Builder Window

Michael made an observation late in the episode that deserves more attention than it got in the moment. There is a window right now, probably two to three years, where the person building AI skills inside a CRE firm is in the driver’s seat. Every firm will eventually have internal skills for their models, their underwriting workflows, their asset management processes. The skills that will be most valuable are the ones built by the people who understand the underlying work deeply enough to encode the methodology correctly.

Right now, most firms do not have those skills. The person who created the template model is the natural person to create the skill for it. That person will also manage it going forward, because managing the skill means managing what the AI knows about how the firm does its work. That is a meaningful role, and it requires understanding both the modeling and the AI tools well enough to hold both simultaneously.

Spencer connected this back to the Accelerator. Part of what A.CRE is now building into the curriculum is an Excel model skill builder: a skill for building skills, that teaches Accelerator members to take the models they have learned to build from scratch and create the AI skills that make those models actionable for their teams. Learn to build the model. Learn to build the skill. Use both to do the work at a level of speed and scale that was not possible before.

Sam was direct: the opportunity is there for anyone willing to take it, but it will not stay open indefinitely. The firms and professionals who move now accumulate experience, data, and institutional knowledge inside their skills that will be genuinely hard to replicate later. The ones who wait will find the playbook more crowded and the advantage compressed.

7. Human Times the Tool

Sam’s closing observation is the editorial summary of the episode. What he watched Spencer do on screen was impressive. But what made it impressive was not the AI. It was the combination: thirteen years of financial modeling expertise, a month of skill-building work by Spencer and the A.CRE team, and a set of tools that are finally good enough to execute the methodology reliably.

His formulation: human times the tool. Both factors have to be present. Both have to be at their best version for the opportunity to materialize. A great tool in the hands of someone who does not understand the underlying work produces unvalidated output that creates more risk than it saves time. Deep expertise without the tool produces the same work at the same speed, with no multiplier. The combination is where the value lives.

This is also why the fear response to AI, the idea that it will replace the analysts who do this work, misreads what is actually happening. What the demo showed is not that AI replaces the analyst. It is that AI handles the mechanical parsing and input population work, which frees the analyst to do the validation, refinement, and judgment work that requires genuine understanding. That is a better use of an analyst’s time. The analyst who understands that and builds the skills to take advantage of it becomes more valuable. The one who does not becomes less so.

The Bigger Idea

The arrival of AI real estate financial modeling is a meaningful moment for the industry, and Spencer treated it that way on this episode. He was careful not to overstate it. The skill works for standard, repeatable modeling tasks. More complex, nuanced underwriting still requires more human involvement. The output needs validation every time. But the threshold has been crossed. The tool now saves more time than it costs to check, which is the only measure that matters.

What makes the demo historically significant for A.CRE is what it represents in terms of the long arc. Spencer and Michael started sharing template models on a blog over a decade ago because they wanted jobs and used the models as a portfolio. The models grew into a library, the library into an Accelerator, the Accelerator into one of the most widely used real estate financial modeling programs in the country. The skills being built now are the next layer: not replacing the models or the training, but connecting both to AI in a way that multiplies what a trained professional can do with them.

The living skill concept is the thread worth pulling on. Every skill you build today is a starting point. Every session you run adds to the log. Every edge case you encounter refines the methodology. The skill becomes more valuable the more you use it, because it accumulates the experience of the person using it. That is how expertise works. What Spencer has built is a way to encode expertise and have it compound over time, which is a different kind of tool than anything the industry has had before.

Sam’s human-times-the-tool framing is the right way to think about what the next few years look like for CRE professionals navigating AI. The question is not whether to use the tools. It is whether you have built the foundation that makes the tools worth using. The A.CRE Accelerator builds that foundation. AI.Edge helps professionals stay current as the tools keep evolving. And platforms like CRE Agents bring AI capability into real estate workflows in a way that produces outputs a trained professional can trust and act on.

The era has arrived. The professionals who are ready for it are the ones who did the long division.

Frequently Asked Questions about Episode 10 of Multipliers: Modeling Like a Deal Seasoned Pro in the Era of AI

An AI skill is a folder of files that encodes the methodology behind a specific model: the inputs, the outputs, the modeling conventions, and the logic that connects them. When you inject the skill into an AI session alongside the model file, the AI reads the methodology and uses it to guide how it populates inputs, validates assumptions, and interprets data from documents like an OM or T12. Spencer demonstrated this live in the episode, uploading an actual offering memorandum, rent roll, and T12 for a Phoenix apartment deal and watching Claude populate a full first-pass underwriting in under 20 minutes.

The skill produces output that requires a trained eye to validate. If an input is wrong, an assumption is misread from the OM, or the model produces an error, the person using the skill needs to catch it. That requires understanding what the output should look like and why each number is what it is. Spencer was explicit: the skill note reads that it is built for professionals who already have a deep understanding of financial modeling, ideally graduates of the A.CRE Accelerator. The skill is the calculator. You still need to know long division first.

Every AI conversation exists within a context window: everything the AI knows in that session lives on a single metaphorical piece of paper. When you add content unrelated to the main objective of the thread, you pollute the context, which reduces the quality of every subsequent output because the AI has to carry irrelevant material alongside the relevant. Spencer recommends keeping exploratory and prompt-engineering work in a separate Claude chat thread, and bringing only clean, finalized instructions into the coding agent or builder tool.

Two reasons: cost and context quality. Tools like Replit price on token usage, and every prompt triggers the system to read all related files, which adds up quickly during exploratory brainstorming. More importantly, the coding agent produces better output when its context window contains only what is relevant to the build. Using Claude in chat mode as the prompt engineer keeps the brainstorming separate, produces a clean brief, and hands that brief to the coding agent to execute. One role thinks. The other builds.

A living skill is a standard AI skill with an added log file that gets updated at the end of every session. Spencer builds the instruction directly into the skill.md: at the end of every discussion, update the skill based on what we did. The AI prompts for confirmation, then appends the session to the log and optionally updates the relevant reference files. The next time you use the skill, it has context from every previous session. Michael described it as the digital version of experience: the skill accumulates the nuance and edge cases you encounter over time, the same way a person does.

Spencer uploaded an OM, T12, and rent roll for a real 134-unit Phoenix apartment deal into Claude in Excel alongside the apartment acquisition model skill. The AI queried live Freddie Mac rate data from the A.CRE Intelligence Hub to size the debt, populated inputs with documented reasoning in cell comments, and completed a first-pass underwriting in under 20 minutes. It identified that the deal did not clear an institutional return threshold at the 18 million asking price, and pinpointed why: the broker was using a year-one pro forma NOI nearly 50% above in-place figures while excluding the management fee.

The A.CRE Intelligence Hub is a data connector available to AI.Edge members that brings live CRE data into any AI environment that supports MCP. In this episode, Spencer used it to pull a current Freddie Mac agency multifamily rate sheet, which the AI used to size the debt on the Phoenix deal in real time rather than relying on a static assumption. The Hub connects to Claude in Excel, Claude desktop, ChatGPT, Gemini, and any other tool that supports MCP connectors.

Michael described a window of roughly two to three years where the professionals building AI skills inside CRE firms have a structural advantage. Every firm will eventually have internal skills for their models and workflows, but most do not have them yet. The person who builds those skills encodes the firm’s methodology into the AI and manages it going forward. That role requires deep understanding of both the underlying work and the AI tools. The professionals who build this capability now accumulate experience and institutional knowledge that will be genuinely hard to replicate later.

The skill includes different roles: acquisitions analyst, broker representing a seller, lender, LP capital allocator, and others. The role changes both the narrative the AI brings to the underwriting and which outputs it prioritizes. A broker representing a seller will model a rosier story, push assumptions toward the optimistic end, and focus on price rather than IRR. A lender will be conservative, cap the upside, and focus on debt service coverage. The AI understands these orientations and adjusts its approach accordingly, including which inputs to emphasize and how to frame the output.

Sam put it cleanly: human times the tool. Both have to be present and both have to be at their best for the opportunity to materialize. The AI skill handles the mechanical parsing and input population work, freeing the analyst to focus on validation, refinement, and judgment. But that shift only creates value if the analyst understands the model well enough to catch errors and make good decisions with the output. The A.CRE Accelerator builds the foundation. The skill is the calculator that goes with it. Communities like AI.Edge help CRE professionals stay current as the tools keep evolving.

by