Episode 18 of Multipliers: JEV Is a Bigger Deal for CRE Than LLMs

Spencer posted about JEV on LinkedIn a few days before this episode and very few people from our industry reacted to it. He understands why. This is boring engineer stuff. The end user of an AI system doesn’t care how the micro-decisions inside that system get made. They care that the output is better, faster, and cheaper than what they had before.

System one models like JEV make that possible in a way that LLMs can’t

JEV is a specialized model developed released Sept 15th 2026 by a company called TypeSafe. Rather than producing words, summarizing a document, or generating a paragraph, it takes unstructured information and returns probabilities across a defined set of answers: yes or no, true or false, or a selection from a list.

Simply put, it’s an AI-powered if statement.

It runs at roughly 400 times the speed of a frontier large language model, at a fraction of the cost. And for an industry whose processes are built almost entirely on micro-decisions, it arguably is a bigger deal than the invention of LLMs.

Or at least, that’s the statement Spencer makes in this episode as Spencer and Sam Carlson discuss a variety of topics from family, to AI and real estate financial modeling proficiency, to system one models.

You might also enjoy: Episode 17 of Multipliers: The AI Tech Stack Every CRE Professional Needs to Understand, where Michael and Spencer dig into the five-layer framework that sits underneath everything in this conversation.

Related: If you want to go deeper on the intelligence layer and how Spencer thinks about model cost and capability, we’ve been tracking that at AI.Edge through our LLM leaderboard and monthly intelligence briefs.

---

Episode 18 of Multipliers: JEV Is a Bigger Deal for CRE Than LLMs

Michael was in the dentist chair this week. Sam was back after a couple of episodes on grandpa duty. Before getting into JEV, the conversation turned to what this is all for.

Sam has a grandson named Bear. Sam is a young grandpa, but grandpa nonetheless. He described what it felt like to be left alone with the baby for a couple of hours, walking around the block with a stroller, watching old Donald Duck cartoons from 1938. The same unconditional care as raising your own kids, but without the weight of building a career underneath it.

Spencer had a parallel moment. His college daughter’s credit card got compromised and her phone locked at the same time, which left her stranded. She sent a message to Spencer from her boyfriend’s phone asking for grocery money. Spencer described it as the best text he had gotten in years!

Both stories connect to a question Spencer had been turning over in the steam room that morning: what is all of this for? The answer he kept coming back to was moments like those.

The conversation then turned nerdy with the discussion about JEV, Sam’s son trying to buy a used 4Runner, a CREW Miami workshop Spencer was teaching that afternoon, and why the bottleneck to AI adoption in this industry right now isn’t the technology.

Spencer also showed Sam the workshop presentation he had built the night before using Opus 5.5 and Claude Design, which was a good example of a different kind of multiplier before the group got into the main topic.

Listen on: Apple Podcasts | Spotify

Episode Highlights

Here are the themes that stood out.

1. What JEV Is and Why It Matters for CRE

Spencer started where he always starts when explaining something technical: with something everyone in this audience already knows.

The Excel if statement.

If you have ever built a financial model, you have written dozens of them. If this is true, do that. Otherwise, do something else. The logic that drives a predictive cash flow model is, at its core, a series of if statements. Boolean logic. Everything is either true or false, and something different happens depending on which it is.

JEV, developed by a company called TypeSafe, applies that same logic to unstructured data.

You hand it a PDF, an email, a web page, a lease, an offering memorandum. You ask it a question with a defined set of possible answers. It reads the input and returns a probability distribution across those answers: 97% chance this is yes, 3% chance this is no.

What makes this different from asking a large language model the same question is threefold.

According to the benchmarks Spencer discussed, JEV runs at approximately 400 times the speed and 40 times lower cost than a frontier model. For binary and classification tasks, those benchmarks also show significantly higher accuracy at comparable cost.

TypeSafe describes JEV as having zero hallucinations because the model is bound by the answer choices you give it. It can’t invent an answer outside those choices. It can still return the wrong answer from the valid set, which is why accuracy and confidence still matter. The natural question is why this matters more for CRE than large language models.

Because the vast majority of the processes in the industry are built on series of micro-decisions.

Think about an invoice arriving in a property management team’s inbox. From the moment it lands until it appears as a reconciled line item in the quarterly investor report, it passes through dozens of micro-decisions. Is there an invoice in this email? Does the amount match the purchase order? Does this invoice belong to this property or that one?

Every one of those is a yes or no question.

Right now, a large language model handling those questions is expensive and slow relative to what the task requires. A human doing it is expensive, slow, and subject to fatigue. According to the benchmarks Spencer discussed, JEV handles them at approximately 400 times the speed and 40 times lower cost than an LLM alternative, with comparable or better accuracy.

Multiply that across an average property with hundreds of invoices per month, across a portfolio of assets, and the economics of running AI-native property operations shift materially.

And property accounting is just one example. Deal screening, lease abstraction, estoppel reconciliation, investor reporting, compliance review: all of them are built on series of micro-decisions. JEV is built exactly for this.

2. How CRE Agents Has Already Implemented JEV

The most concrete evidence that JEV is not theoretical is what the team has already done with it at CRE Agents.

CRE Agents has implemented JEV across seven different internal processes. Spencer described one in enough detail to be useful.

CRE Agents uses a capability called task calling, where the system hands an AI the specific task it needs to perform, along with all the relevant instructions and data. The way it previously worked required two round trips. The AI would ask CRE Agents what it should do. CRE Agents would return three or four or five options. The AI would pick one. Then CRE Agents would send back all the details needed to perform it.

Two conversations where one should be enough.

With JEV making the task selection decision, it now happens in a single response. JEV reads the context, identifies the right task, and returns that task along with the full instructions. One round trip instead of two.

The result: what used to take 10 seconds now takes 2 seconds. What used to cost roughly 15 cents now has effectively zero marginal model cost.

That’s a structural shift in the economics of AI-powered workflows, not a marginal improvement. And it compounds across every task call, in every session, for every user.

CRE Agents uses a 95% probability threshold on JEV’s output. When it returns a result above that threshold, CRE Agents uses it and communicates the probability as part of the result. When it returns below that threshold, the system defaults back to an LLM and flags to the user that confidence is lower. That threshold logic is what makes a probabilistic system trustworthy in a professional context where errors have consequences.

3. Sam’s 4Runner Story: A Single-Use AI Agent in Practice

Before getting into JEV, Sam shared a story about his son Cooper that is one of the clearest illustrations of what AI-native research actually looks like day-to-day.

Cooper wanted to buy a Toyota 4Runner listed in Salt Lake City, five hours from their home in Boise. Sam’s wife Morgan had already made up her mind: it was a jalopy, and she wanted Sam to tell Cooper as much. Sam is not a car person, so instead of just passing along an opinion, he went to Manus.

He asked it to research the vehicle by VIN and model. Manus did a full discovery on the VIN: the car had previously been in Iraq, which was not disclosed in the listing, though Sam noted it had not affected the title. The bulk of its two hundred thousand miles had been spent on the East Coast, which raised the immediate concern of corrosion and rust. Manus also found the car was overpriced by $4,000 and gave all the reasons why.

Sam then asked Manus to search within a hundred-mile radius around Boise for comparable vehicles at better prices. It found seven or eight options, including one with upgraded wheels that caught Cooper’s eye. The whole exercise took fifteen minutes and about thirty-five hundred tokens.

Cooper did not have to drive five hours. Sam got to be a useful dad.

Spencer framed this as a single-use AI agent: a task with a clear objective, the right data readily available, and a concrete outcome. What made it high feasibility in the Multiplier Framework sense is that car VIN history, pricing data, and regional listings are all publicly accessible. The challenge in real estate is that most of the data CRE professionals need sits behind paywalls, inside unstructured drives, in email threads, or in proprietary systems that were never designed for AI access. JEV helps enormously on the decision layer. Getting the data layer right is still the harder problem.

4. The Presentation Multiplier: Opus 5.5 and Claude Design

Spencer was teaching a three-hour workshop for CREW Miami that afternoon and had built the presentation the night before using Opus 5.5 and Claude Design.

A bit of context on how he got there. Spencer had been using Manus for presentations for about a year. When Manus experienced server disruptions, Spencer switched to Claude Design. It was not as good at first, but Opus 5.5 came out the day before the workshop and the quality had improved enough to change the equation.

The multiplier that matters:

Spencer used to spend roughly half his time on a presentation dealing with formatting, fonts, layout, and design. That half now takes about five percent. The other ninety-five percent goes to content: what is on each slide, why it matters, and what the audience should do with it.

The output is better because the cognitive energy that used to go into formatting now goes into meaning.

The CREW presentation included animated slides, embedded voting functionality, a live countdown timer, and a competition where the best agent built during the three-hour workshop wins a sailboat in a bottle. All of that was built from a Word document outline with minimal revision.

Sam made a good observation about one thing AI still does not do well: choosing the right story image for a conceptual point. The horse harness photograph in the presentation, which Spencer uses to explain what an AI harness is, came from the internet rather than from an AI image generator. AI image generation is getting better, but it still can’t reliably produce the specific visual you need for a specific conceptual moment. That is one of the remaining gaps.

5. The Real Bottleneck Is Retooling and Reskilling

Sam asked a good question near the end: where will the industry be a year from now?

The bottleneck right now is not the technology.

JEV is available. The models are capable. The harnesses are good enough to get started. For many of the use cases CRE professionals are exploring today, the technology is no longer the primary blocker.

What is blocking faster adoption is retooling and reskilling: the organizational and individual work of changing how people operate, what systems they use, and what skills they need. All of that has to happen while everyone still has actual work to do. Portfolios to manage. Transactions to execute. Investors to serve. You cannot stop the machine to overhaul it.

Spencer draws the analogy to the transition from paper desks to computer desks in the early days of the personal computer. When a desk was covered in papers, you worked one way. When a computer arrived on that desk, you worked a different way. Getting from one to the other required learning new skills and replacing old tools, and it took years. That transition is underway in CRE right now.

Firms that start retooling and reskilling now will have more time to learn what works, build internal capability, and adapt as the technology evolves.

The A.CRE Accelerator is designed to build the foundational CRE and financial modeling skills that make everything else meaningful. AI.Edge is where professionals develop the AI fluency to put those skills to work. And CRE Agents is what it looks like when the knowledge layer, the tools, and JEV are already assembled and pointed at commercial real estate.

---

The Bigger Idea: Underwrite Your AI the Way You Underwrite a Deal

The invoice-to-investor-report pipeline is the right mental model for understanding what JEV makes possible in CRE.

Every workflow in this industry that moves information from one place to another, through a series of decisions, in a specific sequence, is a candidate for this kind of improvement. Many of the micro-decisions in those workflows are potential candidates for a system like JEV, particularly when the decision can be reduced to a defined set of possible answers.

The math on the task calling example Spencer described is worth sitting with. Effectively zero marginal model cost and two seconds versus roughly fifteen cents and ten seconds, across every task call and every user session. That compounds into a real economic difference.

Most CRE professionals already know how to think about this. The industry underwrites investments for a living, comparing incremental cost against incremental value all day.

That same thinking applies to AI infrastructure.

If an AI workflow costs more to operate than the value it produces, don’t build it. If a frontier model reduces errors on a high-stakes underwriting task enough to justify its cost, pay for it. If a lightweight model can handle fifty thousand binary classifications at one-tenth the cost with no meaningful degradation in quality, use that instead.

It’s about assembling the lowest-cost stack that reliably produces an output you can act on, not simply using the cheapest AI available.

JEV is one of the most significant developments Spencer has seen toward making that possible across the kinds of workflows that define our industry. It is boring from the outside. The engineers building AI agents will be the ones who actually notice it. But the CRE professionals who end up with faster, cheaper, more accurate operations will benefit from it whether they know what JEV is or not.

---

Frequently Asked Questions about Episode 18 of Multipliers: JEV Is a Bigger Deal for CRE Than LLMs