You are here: / / / Episode 17 of Multipliers: The AI Tech Stack Every CRE Professional Needs...
A.CRE
, , ,

Episode 17 of Multipliers: The AI Tech Stack Every CRE Professional Needs to Master

More commercial real estate professionals are using AI every day. But using AI and understanding how to engineer an AI output are two very different things.

And that distinction is going to matter a lot.

We’re moving into a world where the question is no longer simply, “Which AI model is the smartest?” The better question is: What combination of AI technology will reliably produce the output I need, at a cost that makes economic sense?

That’s the reason we’ve started thinking encouraging you to think of AI in terms of the AI tech stack.

For the CRE professional, the stack consists of four primary technology layers: the harness, the intelligence, the tools, and the knowledge. Sitting above all four is the human who applies judgment to the output.

Get the stack wrong and you can spend a lot of money producing mediocre work. Get it right and a less expensive model, paired with the right harness, tools, and domain knowledge, can outperform a more expensive model operating without the rest of the stack.

And that is ultimately the point. The goal isn’t to use the most powerful AI available. The goal is to assemble the right AI tech stack to produce a usable output at a cost that delivers an ROI.

In this episode of the Multipliers podcast, Michael Belasco and I dig into exactly how I think about that stack, why each layer matters, and why understanding how the pieces fit together will increasingly be an important skill for commercial real estate professionals.

Episode 17 of Multipliers: The AI Tech Stack Every CRE Professional Needs to Understand

Sam was on grandpa duty this week, so Michael and I had the podcast to ourselves.

We started with the less glamorous side of commercial real estate. Michael has been dealing with septic issues, Wi-Fi problems, landscapers, telecom contracts, and the normal whack-a-mole that comes with operating real estate.

That led to an interesting discussion about asset management and value creation before we moved into the primary topic for the episode: the AI tech stack.

I’ve been spending a lot of time lately training CRE professionals and organizations on how to apply AI to their work. And one thing has become increasingly clear to me. As this technology begins to mature, there are certain frameworks that I think will endure even as the individual models, applications, and vendors change.

The Multiplier Framework is one of those. Audit the work you do, determine where automation would have the greatest impact, assess how feasible automation is, and then attack the highest-impact, highest-feasibility tasks first.

The AI tech stack is another.

If the Multiplier Framework helps answer “Where should I use AI?”, the AI tech stack helps answer “How do I assemble AI to actually do the work?”

Listen on: Apple Podcasts | Spotify

Why This Episode, Why Now

There is an enormous amount of attention being paid to AI adoption right now. Firms are buying ChatGPT licenses. They’re buying Claude licenses. Teams are experimenting with Copilot, Gemini, Grok, and dozens of other applications.

But buying an AI subscription is not the same thing as implementing AI.

A firm can give every employee access to the most intelligent model in the world and still produce very little measurable ROI from it.

Why?

Because intelligence is only one ingredient in a useful AI output.

Think about a CRE analyst. You wouldn’t hire the smartest analyst you could find, sit that person at an empty desk, give them no computer, no Excel, no access to your files, no market data, and no understanding of how your firm underwrites deals, and then expect great work.

Yet that’s effectively what we’re doing when we hand a general-purpose AI model a prompt and expect it to perform specialized commercial real estate work.

To produce a useful output, the AI needs an environment in which to operate. It needs enough intelligence for the task. It needs the tools required to complete the work. And it needs the data and methodology that tell it how a commercial real estate professional would actually approach the task.

Then, after all of that, a human still needs to determine whether the output makes sense.

That combination is the AI tech stack.

And increasingly, I think the CRE professional who understands how to assemble and manage that stack will have an enormous advantage.

Episode Highlights

Here are a few of the themes that stood out from the conversation.

1. The Four-Hour Phone Call and Why ROI Is the Entire Point

Before getting into AI, Michael shared a good example of value creation at one of his RV parks.

He inherited a telecom contract from the previous owner that nobody had spent much time challenging. So Michael got on the phone with the provider.

Four hours later, he had reduced that expense by roughly $20,000 per year.

It wasn’t glamorous. It wasn’t some brilliant acquisition strategy. It was an asset manager grinding through an operating expense line item that didn’t make sense.

But take roughly $20,000 of additional NOI and capitalize it at an 8% cap rate, and you’ve created roughly a quarter-million dollars of real estate value.

Michael was also rebidding landscaping, where an expense of nearly $2,000 per month was being reduced to roughly $300 per month.

I like this example because it gets at the heart of why the economics of AI matter.

Commercial real estate professionals already think this way with physical assets. Spend a dollar here, save five dollars there, increase NOI, improve value.

We need to start thinking similarly about our AI infrastructure.

If an AI workflow saves an analyst 100 hours per year but costs $50,000 to operate, that’s probably not a particularly good automation.

If another version produces essentially the same work for $500, that’s a very different economic proposition.

The output matters. But the cost of producing the output matters too.

And as firms begin running thousands or millions of AI tasks, those differences will compound.

2. The Harness Layer: Where the AI Actually Works

The first technology layer I think about is the harness.

ChatGPT is probably the most familiar example.

The underlying technology is the large language model. But historically, interacting directly with a model was a technical exercise. OpenAI wrapped the model inside a simple user interface, gave normal people a text box, and ChatGPT was born.

That interface is the harness.

I sometimes use a horse analogy. The model is the raw horsepower. The harness is what allows you to direct that horsepower toward something useful.

Today there are hundreds of AI harnesses.

ChatGPT is a harness. Claude Chat is a harness. Claude Code and Codex are coding harnesses. Claude in Excel, ChatGPT in Excel, and other AI spreadsheet products are harnesses built for spreadsheet work. Perplexity is a harness optimized around research. Products such as Grok Bot, Goose, OpenWork, and others are all different environments for putting AI models to work.

This distinction matters because the harness changes what the underlying intelligence can actually do.

A model sitting inside a basic chat interface has one set of capabilities. Put that same model inside a coding harness with access to a terminal, files, and software development tools, and suddenly it can perform entirely different work.

Put it in an Excel harness and it can interact with a financial model.

Give it an agentic harness connected to email, calendar, and internal systems and it can begin performing multi-step business processes.

So when I hear someone say, “We use Claude” or “Our firm uses ChatGPT,” I increasingly think that answer is incomplete.

Which harness are you using? What can it do? Which tools can it access? Which models can it run? How secure is it? How much control do you have over the environment?

The right answer depends entirely on the task.

If you’re trying to better understand the growing set of options available today, we’ve been maintaining a broader list in our AI Tools for Commercial Real Estate guide.

3. The Intelligence Layer: Don’t Pay for Intelligence You Don’t Need

The next layer is the model itself, or what I call the intelligence layer.

This is GPT, Claude, Grok, Gemini, Kimi, GLM, and the growing list of other models available today.

Different models have different capabilities. Some reason better. Some code better. Some are faster. Some are dramatically cheaper.

And this is where the economics start to get interesting.

During a recent AI.Edge lesson, I took a messy multifamily rent roll and ran essentially the same task through three different models. The job was to parse the rent roll, clean up the tenant information, organize the data, and roll it into a unit mix table.

All three models got the primary answer correct.

But their cost was dramatically different.

The most expensive run cost approximately $2.70. Another cost roughly $1.30. The cheapest was about four cents.

Think about that spread.

We’re not talking about a 10% or 20% difference in cost. We’re talking about orders of magnitude.

Now, the four-cent model wasn’t my preferred model for that particular task. It introduced a few semantic and classification errors that made me uncomfortable. The $1.30 model produced the best combination of accuracy and cost, so that was the winner.

But the lesson wasn’t that one particular model is better than another.

The lesson is that the optimal model depends on the task.

For one task, I may happily pay $2.70 because the incremental intelligence is worth it. For another, a four-cent model may be more than capable.

That distinction barely matters if you run two prompts per week.

It matters enormously if you’re an enterprise running millions of AI actions.

I think there will increasingly be a role inside CRE organizations for someone who understands this cost-capability matrix. That person will know which tasks require expensive frontier intelligence and which can be routed to cheaper models without degrading the output.

That’s real value creation.

If you can deliver the same usable output while materially reducing the cost of producing it, you’ve effectively done the AI equivalent of Michael’s four-hour telecom call.

4. The Tools Layer: Give Your AI the Ability to Do Something

A harness and a model still aren’t enough.

Your AI also needs tools.

Imagine hiring a new analyst and asking her to parse a rent roll. Then imagine giving her no computer, no spreadsheet software, no PDF reader, and no access to the file.

She can’t do the work.

The difference between a human analyst and an AI model is that the analyst will probably stop and tell you what she’s missing. An AI may attempt the task anyway and confidently give you something that looks like work product.

This is why the tools layer matters.

Web search is a tool. Code execution is a tool. Excel is a tool. Gmail is a tool. Your calendar is a tool. Your document repository is a tool. A proprietary CRE database can be a tool.

Increasingly, those tools are being connected to AI harnesses through MCP, or Model Context Protocol.

You don’t need to become a software engineer to understand the important part.

MCP is effectively a standard way for an AI environment to discover and interact with outside tools and data sources.

This opens up all sorts of possibilities in commercial real estate.

Michael, for example, has collected hundreds of offering memoranda over his career. Imagine organizing those documents into a proprietary database and then giving an AI agent a tool that allows it to query that database.

Now the agent doesn’t just know what the internet knows.

It can access Michael’s accumulated experience.

Or perhaps your firm connects an agent to its property management system, CRM, data warehouse, market data providers, email, and document storage.

Now the agent has the tools necessary to actually perform work across the organization.

This is also why simply comparing models on a leaderboard misses a lot of what matters. A slightly less intelligent model with access to the right tools may be far more useful than the world’s smartest model sitting alone in a chat window.

5. The Knowledge Layer: Where CRE Expertise Actually Enters the Stack

The fourth technology layer is the one I think is both the most important and the hardest to build.

Knowledge.

I break the knowledge layer into two components: data and methodology.

Data is the information required to perform the work.

If you’re underwriting a multifamily acquisition, that may include the rent roll, T12, property tax records, market rents, demographics, interest rates, supply data, and internal assumptions.

Some of that data is external. Some of it may be proprietary to your firm.

The second component is methodology.

This is how you actually do the work.

How does your firm normalize a T12? How do you determine which expenses are recurring? How do you underwrite property taxes? How do you screen a deal against your buy box? How do you construct an investment committee memo? How do you perform a particular waterfall calculation?

The model has general intelligence. It doesn’t automatically have your firm’s methodology.

In an AI system, one increasingly common way to encode that methodology is through an Agent Skill.

A Skill is essentially a structured set of instructions, examples, reference materials, and resources that teaches the AI how to perform a particular task.

I think of it like having an experienced CRE professional sitting on the AI’s shoulder, whispering instructions while the model works.

“Here’s how we handle the T12.”

“Here’s how we think about taxes.”

“Here’s where you should get this piece of data.”

“Here’s the Excel model we use.”

“Here’s what you should check before you call the work complete.”

That’s the knowledge layer.

And this is where I think a lot of CRE firms will discover that implementing AI is more work than simply buying licenses.

The firm’s institutional knowledge has to be captured. Its data has to be made accessible. Its methodologies need to be encoded. And all of that has to be maintained.

There are really two approaches to that problem: build it or buy it.

You can build your own knowledge layer internally. In many cases, I think firms absolutely should, particularly where the methodology or data represents proprietary intellectual property.

But there are also enormous amounts of CRE knowledge that are common across firms.

That’s part of why we’ve built CRE Agents.

CRE Agents is, at its core, a commercial real estate knowledge layer. It packages CRE-specific methodology into Agent Skills and pairs those skills with real estate data so that general-purpose AI can perform specialized CRE tasks.

Rather than every organization independently teaching its AI how to abstract a lease, analyze a rent roll, size debt, research a property, evaluate comps, or work through dozens of other recurring CRE tasks, that domain context can be made available to the AI and then supplemented with the firm’s proprietary knowledge.

For those learning how this works, we’ve taken a similar approach with the A.CRE Intelligence Hub. The Hub is designed as a teaching environment where CRE professionals can work with Agent Skills and primary-source real estate data inside the AI tools they already use.

This distinction between general intelligence and domain knowledge is important.

The models will continue improving. But nearly everyone will have access to those same models.

Your proprietary data and methodology are different.

That’s where durable differentiation can live.

6. The Next Generation CRE Analyst Will Understand How to Engineer an Output

Michael is putting this into practice right now.

He’s building an investor reporting agent named Warren.

Warren is connected into Slack and Gmail. Michael is working through Drive, GitHub, and the other systems Warren will need. And he’s intentionally moving slowly because his goal isn’t just to get Warren working.

His goal is to understand why Warren works.

I think that’s exactly the right approach.

Over the next few years, I expect we’ll go through an enormous reskilling of the commercial real estate workforce.

And I don’t think the people best positioned to lead that transition are necessarily traditional software engineers.

They need to understand the work.

The person architecting an acquisition underwriting agent needs to understand acquisition underwriting. The person building an asset management workflow needs to understand asset management. The person automating investor reporting needs to understand what investors actually need to know.

That’s why I keep coming back to the Excel nerd.

The best Excel modelers already think like systems engineers.

They take inputs. They apply a methodology. They create a structured process. They check for errors. And ultimately they produce an output that someone else can use to make a decision.

AI engineering is surprisingly similar.

Instead of only asking, “What formula should go in this cell?”, the next generation analyst will ask:

  • Which harness is right for this workflow?
  • How much intelligence does this task actually require?
  • Which model gives me the best cost-to-capability tradeoff?
  • Which tools does the AI need to perform the work?
  • What data does it need?
  • What methodology needs to be encoded into a Skill?
  • How do I validate the output before someone relies on it?

That’s a very different skill set from simply being “good at prompting.”

And I think it’s a much more durable one.

If you’re looking to develop that capability, this is also the philosophy behind AI.Edge. The goal isn’t to teach a collection of AI tricks. It’s to help CRE professionals develop the fluency to understand and apply this technology as it changes.

The Bigger Idea: Build for the Output and Underwrite the Cost

If there is one idea I hope people take away from this episode, it’s this:

Don’t start with the model.

Start with the output.

What exactly are you trying to produce?

Maybe it’s an underwritten acquisition model. Maybe it’s an investor report. Maybe it’s a lease abstract, a market study, a comp set, an IC memo, an expense variance analysis, or a first draft of a broker opinion of value.

Then work backward.

What knowledge is required to produce that output correctly?

What tools are required to gather the inputs and create the deliverable?

How much intelligence is actually necessary?

Which harness gives that intelligence the environment it needs to perform the task?

And finally, what does the whole thing cost?

This last question is going to become increasingly important.

Today, most individual AI users don’t think much about inference cost because their subscription hides it from them. You pay $20, $100, or some enterprise fee and use the product.

But underneath that subscription, every AI action has an economic cost.

As firms move from occasional chat usage to agents running hundreds or thousands of workflows every day, the cost of intelligence becomes a real operating expense.

The difference between a $2.70 task and a $0.04 task doesn’t sound like much when you do it once.

Do it a million times and the difference is $2.66 million.

Of course, that’s an intentionally simple example. The cheaper model is worthless if its error rate makes the output unusable.

And that’s precisely the point.

We’re not optimizing for the cheapest AI.

We’re optimizing for the lowest-cost stack that reliably produces an output we can use.

That’s an ROI exercise.

Commercial real estate professionals should be uniquely comfortable with that way of thinking. We underwrite investments for a living. We compare incremental cost against incremental value all day long.

We should underwrite our AI the same way.

Maybe a frontier model costs ten times as much but materially reduces errors on a high-stakes underwriting task. Great. Pay for it.

Maybe a lightweight model can classify 50,000 documents at one-tenth the cost with no meaningful degradation in the result. Great. Use that instead.

Maybe an expensive AI workflow only saves an employee ten minutes per month. Don’t build it.

Maybe another workflow eliminates hundreds of hours of repetitive work, improves consistency, and allows a deal team to evaluate twice as many opportunities. That’s where you invest.

This is where the Multiplier Framework and the AI tech stack come together.

First identify the work where automation creates the most value.

Then architect the stack that produces that work reliably and economically.

Harness. Intelligence. Tools. Knowledge.

And finally, human judgment.

I don’t think CRE professionals need to become computer scientists. I do think we need to understand these layers well enough to ask the right questions and make informed decisions.

Because the winners won’t necessarily be the firms using the smartest model.

They’ll be the firms that figure out how to assemble the right combination of technology and domain expertise to consistently produce better work, faster and at a cost that delivers a return.

Frequently Asked Questions about Episode 17 of Multipliers: The AI Tech Stack Every CRE Professional Needs to Understand

I think about the technology portion of the AI stack in four primary layers: harness, intelligence, tools, and knowledge. The harness is the environment where the AI works. The intelligence layer is the underlying AI model. Tools give the AI the ability to interact with applications, files, software, and outside systems. Knowledge consists primarily of data and methodology, including Agent Skills that teach the AI how to perform specialized CRE work. A human then sits above that technology stack and applies judgment to the output.

Because the quality and economics of an AI output depend on far more than which model you use. A powerful model without the right tools, data, or CRE methodology may produce an unusable result. A less expensive model operating inside a well-designed stack may perform the same task extremely well. Understanding the stack allows a CRE professional to work backward from the desired output and assemble the technology needed to produce it reliably and economically.

An AI harness is the environment or application that wraps around an AI model and gives it useful capabilities. ChatGPT and Claude Chat are familiar chat harnesses. Claude Code and Codex are examples of coding harnesses. There are also harnesses designed for spreadsheets, research, autonomous agents, and other types of work. The harness matters because it determines how the model can interact with files, tools, software, and the user.

Because more intelligence generally comes with a higher cost, and many tasks don’t require the most capable model available. In the rent roll exercise discussed in this episode, three models produced the correct primary result while costing approximately $2.70, $1.30, and $0.04 per run. The cheapest model introduced several issues that made it less appropriate for that specific task, so the middle-cost model offered the best combination of accuracy and price. The objective is not to choose the cheapest model or the smartest model. It is to use the least expensive model capable of reliably producing the output you need.

MCP stands for Model Context Protocol. It is a standard that allows AI applications to connect with outside tools and data sources. For a CRE professional, that might mean giving an AI access to email, spreadsheets, a property management system, internal databases, market data, property records, demographic information, or other applications needed to complete a task. Tools are what allow AI to move beyond producing text and begin doing useful work across the systems a CRE professional uses.

The knowledge layer is the domain context that turns general-purpose artificial intelligence into something useful for specialized CRE work. I divide it primarily into data and methodology. Data is the information required to perform the task, including both external data and proprietary internal data. Methodology is the process for doing the work, which can increasingly be encoded into Agent Skills. The model provides general intelligence. The knowledge layer teaches that intelligence how your particular commercial real estate task should be performed.

CRE Agents is a commercial real estate knowledge layer. It packages CRE-specific methodology into Agent Skills and connects AI to the real estate data needed to perform specialized work. The idea is that firms shouldn’t necessarily have to rebuild common CRE knowledge from scratch. They can use an existing knowledge layer for common methodologies and data, and then supplement it with the proprietary knowledge that makes their organization unique.

Start with the task and the value of automating it. Then determine the lowest-cost combination of harness, model, tools, and knowledge that can reliably produce a usable output. A cheap AI workflow that produces bad work has no value. An extremely expensive workflow that only saves a few minutes may also have little value. The objective is to create enough productivity, revenue improvement, expense savings, risk reduction, or quality improvement to justify the total cost of producing the AI output.

I think the next generation analyst or associate will combine CRE subject matter expertise with the ability to architect AI workflows. The natural candidates are often the same people who have historically become strong Excel modelers. They already think in terms of inputs, methodology, checks, and outputs. The new skill is learning how to apply that systems thinking across AI harnesses, models, tools, data, and Agent Skills.

Don’t begin by asking which AI model is best. Begin by defining the output you need and then work backward. Determine what knowledge the AI needs, what tools are required, how much intelligence the task demands, and which harness is best suited to the workflow. Then compare the cost of producing that output with the value it creates. The goal is not simply to use AI. The goal is to build an AI tech stack that reliably produces usable commercial real estate work at a cost that delivers an ROI.

by
Real Estate Financial Modeling Accelerator (Updated September 2026)