You are here: / / / Episode 14 of Multipliers: Unpacking the Layers of AI that Make it Useful...
A.CRE
, ,

Episode 14 of Multipliers: Unpacking the Layers of AI that Make it Useful for CRE

Most people in commercial real estate are using AI in some form right now. Very few of them understand what they are actually using. That gap matters, because the professionals who understand the structure of the technology make better decisions about where to invest their time, which tools to build on, and what is actually solid ground versus what is still shifting sand.

Spencer Burton has been in the thick of this for four years, building real estate intelligence for AI agents through every wave of the technology’s evolution. Michael Belasco has been watching from close range, close enough to have real concerns about whether the whole thing was built on a foundation that could shift underneath it. Last week, for the first time, something clicked for Michael. He sat down with Spencer, connected CRE Agents to his preferred AI tools through a natural conversation, watched two agents land in his Slack, and realized the dream he had been waiting for had quietly arrived.

This episode is the conversation that followed. Spencer uses it to lay out the four layers of AI, including the AI capabilities layer that makes the technology useful for CRE work, and Michael uses it to describe what finally changed when all four of those layers came together in a single working experience.

In this episode of the Multipliers podcast, Spencer Burton and Michael Belasco break down the full AI stack, from compute to harness to capabilities layer, explain what CRE Work and Excelente are, why A.CRE is building them open source, and why the AI capabilities layer may become the most important advantage for CRE firms needs to treat this as a four-alarm fire right now, whether they feel it yet or not.

Episode 14 of Multipliers: Unpacking the Layers of AI That Make It Useful for CRE

Sam Carlson is off this week. Spencer and Michael are on their own for an episode that started as a debrief from a meeting the week before and turned into one of the more technically detailed conversations the podcast has produced. Michael’s opening is worth quoting directly: last week, for the first time, he felt the power of what Spencer has been building. Not abstractly. In practice, through a natural conversation, with agents showing up in his Slack and doing real work without any of the clunky friction that had characterized every previous version.

Spencer teaches real estate professionals the technical skills to success here at A.CRE while he also trains AI to work with real estate professionals via CRE Agents. Michael leads an RV park investment platform called RV@ and has begun experimenting with autonomous agents to take on daily tasks.

The episode connects back to Episode 13, where David Conroy talked about the tenant P&L framework and the mentors who accelerated his career. This one is more technical and more urgent, a conversation about the infrastructure that makes AI actually useful and the window that is open right now for firms willing to move.

Listen on: Apple Podcasts | Spotify

Why This Episode, Why Now

Michael’s fear going into this episode was one he had carried for years: that everyone building on AI was building on sand, because the technology was moving too fast for any foundation to solidify. He had that fear about CRE Agents specifically. Spencer kept building anyway. And last week, the thing Michael had been waiting for arrived: a working system that felt like a natural extension of how he already communicated, rather than a tool he had to learn to use.

Spencer’s framing of the moment is precise: it was not that the technology suddenly got better. It was that four years of building through every layer, compute, model, harness, capabilities, had finally produced something where all four layers were working together well enough to deliver a coherent experience. The technology did not change overnight. The accumulation crossed a threshold.

The urgency of the episode comes from Spencer’s observation that most CRE firms are still on the wrong side of that threshold, not because the technology is not ready for them, but because they have not invested in the capabilities layer that would make it work for their specific context. The big institutions are talking about AI. They are running Copilot licenses. They are not building the thing that actually matters, which is the layer that teaches the AI to do the specific kind of work their firm does. And the window for small and mid-sized firms to outmaneuver them on this is real, finite, and closing.

Episode Highlights

Here are the themes that stood out.

1. The Four Layers of AI: A Framework for CRE Professionals

Spencer’s four-layer framework is the most useful organizing structure the podcast has offered for thinking about where AI is headed and where professionals can actually play.

The first layer is compute: the data centers, the inference infrastructure, the raw processing power that makes large language models run. This layer is real estate in a very literal sense. Data centers are being built at extraordinary scale. But it is not a layer where most CRE professionals can participate directly, unless they are in the business of developing or financing the facilities themselves.

The second layer is the model: the large language models themselves, the Claudes and GPT-5s and Geminis of the world, along with the open-weight models like DeepSeek, GLM, and Kimi that are approaching frontier capability at a fraction of the cost. Model training is technically demanding and capital-intensive. Most CRE professionals cannot and should not try to play here.

The third layer is the harness: the interface, the system prompts, the tools, and the conventions that wrap around a model and make it useful. ChatGPT was the first major consumer harness. Claude Code is, in Spencer’s view, currently the best harness in the world. The harness is where most professionals interact with AI, and it is the layer where quality differences are most immediately felt.

The fourth layer is the capabilities layer: the domain-specific knowledge, methodology, and skills that make a general-purpose AI useful for a specific type of work. This is the layer where CRE professionals can build real, lasting advantages. And it is the layer that most firms have not started building yet.

2. The Harness Analogy: Saddle, Buggy, and Why It Matters More Than the Model

Michael asked Spencer to explain what a harness actually is, and the answer Spencer gave is worth dwelling on because it clarifies something that most AI conversations skip over.

Spencer’s analogy: a horse has raw power, but raw power alone is not useful for riding or for pulling a buggy. The saddle makes the horse rideable. The harness connects the horse to the buggy and makes it useful for transport. The better the saddle and harness, the more effectively the horse’s power can be directed toward a purpose.

The practical implication of this analogy showed up in a real comparison from this episode. GPT-5.5 was, on benchmark measures, a more capable model than Opus 4.7. And yet Claude was producing better outputs for most CRE tasks, because the Claude harness, specifically Claude Code and the surrounding ecosystem, was significantly better at directing the model’s capability toward useful work. A weaker model with a better harness outperformed a stronger model with a weaker harness.

Spencer also noted that harnesses are becoming more accessible to build. The AI-assisted coding tools that exist today lower the barrier enough that a subject matter expert, not just an engineer, can develop a harness for their specific domain. That is what CRE Work is: a harness built by people who understand commercial real estate, for the specific ways that professionals in the industry need to interact with AI.

3. What AI Output Quality Actually Depends On

Spencer’s breakdown of the three factors that determine AI output quality is the most directly actionable section of the episode for anyone trying to get better results from the tools they already have.

The first factor is instructions: the quality of the prompt, the specificity of the request, the context provided. This is what most people focus on when they think about using AI well. It matters, but it is only one of three factors.

The second factor is knowledge: the combination of data and methodology that the AI has access to. Data is the harder piece in commercial real estate, where the most valuable information is proprietary and not in any training set. Methodology is where skills and agent skills come in: you encode how your firm does something, and the AI uses that encoding to do it the same way.

The third factor is tools: the external capabilities the AI can invoke to perform work. Spencer’s morning example is concrete: the autonomous agents running for Michael needed access to Google Drive to organize files, Google Calendar to manage scheduling, and Gmail to communicate. Those are tools. Without them, the agent could think but could not act. With them, it could complete actual work.

The quality of a harness is largely a function of what instructions, knowledge, and tools it makes available by default. That is what makes Claude Code different from a raw API call to the same model.

4. CRE Work and Excelente: Two Open Source Harnesses A.CRE Is Building

Spencer announced two harnesses A.CRE is building and will give away for free, both of which address real gaps in the current ecosystem.

Excelente is the in-Excel harness, model-agnostic, that lets users bring their own API key and choose from over 600 available models including free options. The barrier to release has been Microsoft’s approval process for the Office add-in marketplace, which Spencer described with barely concealed frustration: denied twice, both times over something like a missing comma in the company address. It is coming. When it does, it will bring AI-native capability into the spreadsheet environment where most CRE financial modeling already lives.

CRE Work is the more ambitious of the two. It is an open source harness, installable on a personal computer or an organization’s own server, that can connect to any large language model including US-based open weight models like NVIDIA’s Nemotron. The key capability is on-premises deployment: an organization that does not want its data leaving its own infrastructure can install CRE Work on an internal server, connect it to a self-hosted model, and have a fully functional AI environment that never touches an external cloud. For institutional investors and large operators who have been hesitant about AI because of data security concerns, this is a meaningful development. Spencer noted that A.CRE is building it first for its own needs, including protecting what may eventually be a best-in-industry proprietary database, and will share it with the AI.Edge community once it reaches a usable stage.

5. The Capabilities Layer: What Every CRE Firm Will Have to Build or Rent

Spencer’s central argument in this episode is that the capabilities layer is the layer that matters most for CRE firms over the next decade, and it is the layer that almost no one has started building.

The capabilities layer is the encoding of domain expertise into a format that AI can use. At CRE Agents, this means 570 tasks, specific things AI can be taught to do in the context of commercial real estate, packaged and delivered through a connector that works with any harness. When Michael asks his AI to build a stabilized pro forma from an OM and a T12, a figure called Vic sits metaphorically on the AI’s shoulder and whispers in its ear: this is how you handle a T12, this is how you think about non-recurring items, this is how you size property tax, here is the actual property tax record for this asset. The AI, which has a PhD in everything but no street smarts, suddenly has the street smarts it needs.

Spencer’s prediction is direct: every firm will have to either build its capabilities layer or rent it from someone who has. The firms that build it themselves will encode their own methodology and gain a proprietary advantage. The firms that rent it will gain access to a shared methodology that is better than nothing but less differentiated. And the firms that do neither will be operating at an increasing disadvantage as the gap between AI-enabled and AI-naive operations widens.

The AI engineering insight is important here. Spencer asked Claude to analyze the last ten episodes of the All-In podcast and identify the three careers that would outperform all others over the next forty years. The number one answer was something like an AI engineer, defined not as a software developer but as someone who can take vertical domain expertise and encode it in a format that AI can use. In a CRE context, that might mean working with a forty-year accountant to transpose his knowledge into something an AI agent can apply to the books every month. That skill, sitting at the intersection of domain knowledge and AI fluency, is genuinely valuable right now.

6. G Brain: Memory That Actually Works

Spencer introduced G Brain as an open-source tool that has transformed how memory works for the autonomous agents running in Michael’s operation. The concept is sometimes called an LLM wiki: a structured external memory that agents can read from and write to, separate from the context window of any individual conversation.

The problem G Brain solves is one that has come up across several Multipliers episodes: the context window is finite. A long conversation degrades as the AI loses access to information from early in the thread. Across sessions, the AI starts fresh unless memory is explicitly managed. G Brain provides a persistent, structured memory layer that agents can query when they need context from previous sessions or from other agents. Spencer described the quality of memory it provides as meaningfully better than what he gets from Claude or ChatGPT’s built-in memory features.

For Michael’s operation, this means the acquisition agent that runs every morning has access to a growing body of context about what Michael has seen, what he has passed on, and why. The accounting agent that handles the books monthly can refer to previous months and carry forward the relevant context. The agents are not starting from scratch every time. They are building on accumulated experience, which is exactly the living skill concept Spencer described in earlier episodes.

7. The Four Alarm Fire: Big Firms Are Floundering, Small Firms Have a Window

The most pointed section of the episode is Spencer’s assessment of where most CRE firms actually are relative to where the technology is.

His observation about big institutions is specific and unflattering: the senior people do not fully understand what is available, the junior people who do understand it are not being unleashed, and the incentive structure at large organizations actively discourages the kind of bold moves this moment requires. If you are a frontline person at a pension or endowment and you push hard on AI adoption and it does not work, you hurt your bonus and possibly your job. If you push and it works, you get a pat on the back. The asymmetry produces caution.

The result is a lot of Copilot licenses and a lot of conversation, and very little actual capability being built. Spencer’s frustration is real: firms he spoke to a year ago are in the same place today. AI has crossed the threshold. The tools are ready. The constraint is no longer the technology.

For small firms, the opportunity is exactly the inverse. There is no bureaucracy to navigate, no senior skeptics to convince, no incentive structure rewarding caution. A small firm that commits a talented person to building the capabilities layer, or hires someone who already knows how, can be operating at a level of AI-native efficiency that a firm ten times its size cannot replicate, because the large firm’s advantages in capital and credibility do not translate into advantages in this particular race.

Michael’s closing observation is the most concrete version of this: he has two to three AI agents doing tasks that used to take his team hours. He can prompt them in natural language. He does not need to understand the technical infrastructure. He just needs to communicate clearly, the same way he would with a new employee, and the work gets done.

The Bigger Idea

Spencer’s four-layer framework is worth keeping because it gives CRE professionals a map for where to focus. The compute layer and the model layer are real but largely inaccessible to most practitioners. The harness layer is where you interact with AI daily, and the quality of the harness shapes the quality of every interaction. The capabilities layer is where competitive advantages will be built and sustained over the next decade, and it is the layer that almost no one has started building.

The saddle analogy does real work here. A weaker horse with a better saddle outperforms a stronger horse without one. The same is true for AI: a weaker model with a well-built capabilities layer, specific to commercial real estate, will outperform a frontier model pointed at the same task with no domain context at all. That is what Michael experienced last week. Not magic. The right layers, working together, crossing a threshold.

The G Brain observation is the thread to watch most carefully going forward. Persistent, structured memory is the thing that turns an AI agent from a very good assistant into something that behaves more like a developing employee, one that gets better over time because it can refer to what it has learned. Combined with the living skill concept from earlier episodes, where skills update at the end of every session to incorporate new experience, the trajectory is toward agents that genuinely accumulate domain expertise rather than starting from scratch in every conversation.

Michael’s closing is the one to carry: today is the day. Not tomorrow. The technology that was almost there a month ago is here now. The firms and professionals who shift their attention to it today will find, in twelve to eighteen months, that they have built something their competitors cannot quickly replicate. The ones who wait will look around and wonder what happened. The A.CRE Accelerator builds the foundation. AI.Edge keeps you current. And CRE Agents is the capabilities layer that makes the whole stack actually useful for commercial real estate.

Frequently Asked Questions about Episode 14 of Multipliers: Unpacking the Layers of AI That Make It Useful for CRE

Spencer described four layers: compute (the data centers and inference infrastructure), the model (the large language models like Claude and GPT-5), the harness (the interface and tools that make the model useful), and the capabilities layer (the domain-specific knowledge and methodology that makes the AI useful for a specific type of work). Most CRE professionals cannot meaningfully participate in the compute or model layers. The harness layer is where daily interaction happens. The capabilities layer is where lasting competitive advantages will be built, and it is the layer that is most accessible to professionals who understand their domain deeply.

A harness wraps around a large language model and makes it useful: the interface, system prompts, built-in tools, and conventions that direct the model toward specific kinds of work. Spencer used the saddle analogy: a horse has raw power, but you need a saddle to ride it and a harness to connect it to a buggy. The quality of the harness determines how effectively the model’s power can be directed. GPT-5.5 was a stronger model than Opus 4.7 by benchmark measures, and yet Claude produced better outputs for most CRE tasks because the Claude harness, particularly Claude Code, was significantly better at directing the model toward useful work.

Spencer identified three factors: instructions (the quality and specificity of the prompt), knowledge (the combination of data and methodology the AI has access to, including agent skills that encode domain expertise), and tools (the external capabilities the AI can invoke to perform real work, such as access to Google Drive, email, or a property tax database). Most people focus only on instructions. The harness determines what knowledge and tools are available by default, which is why a well-built harness can make a weaker model outperform a stronger one with fewer built-in capabilities.

CRE Work is an open source harness A.CRE is building and will give away free. It can be installed on a personal computer or an organization’s own server and connected to any large language model, including US-based open weight models like NVIDIA Nemotron. The key capability is on-premises deployment: organizations that do not want their data leaving their own infrastructure can run a fully functional AI environment that never touches an external cloud. This is particularly relevant for institutional investors and large operators who have been hesitant about AI because of data security concerns. A.CRE plans to share it with the AI.Edge community once it reaches a usable stage.

Excelente is a free, model-agnostic Excel add-in A.CRE has built that lets users bring their own API key and choose from over 600 available models, including free options from providers releasing models for testing and training. The barrier to release has been Microsoft’s approval process for the Office add-in marketplace, which has denied the application twice over administrative issues. The release is coming. When it arrives, it will bring AI-native capability directly into the spreadsheet environment where most CRE financial modeling already happens.

The capabilities layer is the encoding of domain expertise into a format that AI can use. Without it, AI has general knowledge but no domain-specific methodology: a PhD in everything but no street smarts. CRE Agents delivers 570 tasks as capabilities that teach AI how to do specific commercial real estate work, from building a stabilized pro forma to sizing property tax to handling non-recurring T12 items. Spencer argued that over the next decade, every firm will have to either build its own capabilities layer or rent one from a provider. Firms that build their own will encode proprietary methodology. Firms that do neither will fall behind as the gap between AI-enabled and AI-naive operations widens.

G Brain is an open source tool that functions as an external memory layer for AI agents, sometimes called an LLM wiki. It provides structured, persistent memory that agents can read from and write to across sessions, separate from any individual context window. Spencer described the quality of memory it provides as meaningfully better than what Claude or ChatGPT offer through their built-in memory features. For Michael’s autonomous agents, this means the acquisition agent running every morning has access to a growing body of context about deals reviewed, passed on, and why, rather than starting from scratch each time.

Spencer’s argument is that AI has crossed the threshold where the constraint is no longer the technology. The tools are ready. What is missing is the investment in the capabilities layer that would make them work for specific firms. Most large institutions are still at the talking stage: Copilot licenses, pilot programs, conversations at conferences. They are not building the thing that actually matters. Meanwhile, smaller firms that move now can build AI-native operations that their larger competitors cannot quickly replicate. The window is open today. Spencer believes it will not stay open indefinitely, and the firms that wait will find the gap has widened considerably by the time they start.

Spencer used the term AI champion to describe the person inside an organization who has both the technical fluency and the domain knowledge to bridge the gap between what AI can do and what the firm actually needs it to do. For large firms, this is probably someone already on the team who has been showing an unusual amount of interest in the tools. Spencer argued those people should be pulled off all other responsibilities and focused entirely on this. For smaller firms, it might mean hiring someone with this specific combination of skills. Every firm needs that person right now, either developing internally or brought in from outside.

Michael put it simply: today is the day. The technology that felt almost there a month ago is here now. The starting point is not a technical course or a research project. It is finding your AI champion, whether that is you or someone you hire, and giving them the mandate to start building the capabilities layer for your specific operation. The A.CRE Accelerator builds the foundational skills that make the capabilities layer meaningful. AI.Edge keeps you current on the tools. And CRE Agents is the capabilities layer you can start using today while you build your own.

by
The AI.Edge Q2 2026 Build Competition: Winners Announced!