No. 54 Fintech
Steven Song on Reinventing Underwriting with AI Decision Intelligence
The founder of Diald on why real estate underwriting needs intelligence, how specialized AI agents mimic human teams, and what investors still underestimate about AI in the field.
Conducted in writing · Answers published as received
Underwriting is where commercial real estate decisions are made and where judgment most often breaks down. Former architect turned investor Steven Song built Diald to fix this cognitive bottleneck.
This interview was conducted in writing. Answers are published as received.
It became obvious after years of watching deals succeed or fail for reasons that never showed up in a spreadsheet. The data was never the problem. Humans simply can’t process the sheer complexity of real-world conditions around a site — politics, safety, sentiment, zoning nuance — at any meaningful scale.
Once you see that, the conclusion is unavoidable: CRE doesn’t need more data. It needs intelligence. So I built Diald to turn chaotic, qualitative noise into structured, reliable reasoning. If you can solve that, you unlock an entirely new era of decision-making.
At my real estate investment firm Axle Companies, underwriting felt like flying with outdated instruments. Too much manual interpretation. Too much variability across analysts. And far too slow.
We didn’t need more inputs — we needed a system that could think. A system that could take thousands of weak signals and synthesize them into a coherent view. That’s the missing layer in CRE. Once you realize that, everything else is downstream.
Underwriting is where conviction is created. It’s the core of the entire CRE engine. And ironically, it’s the part everyone ignored because it’s the hardest.
Most proptech went after the surface layer — sourcing, CRM — because it’s easier. But if you want to transform the industry, you go after the bottleneck: the decision itself. That’s underwriting. That’s where real leverage is.
Three fundamental ones:
- Cognitive overload — humans aren’t built to process thousands of unstructured signals per site.
- Inconsistency — the same deal gets different answers depending on who looks at it.
- Speed — underwriting cycles are far too slow for a modern market.
Diald removes those constraints.
A single general model is like asking one person to master every discipline in real estate. Not realistic.
We built Diald with specialized agents — zoning, sentiment, safety, market context — each optimized for its domain. They report their reasoning, and then a coordinator agent synthesizes it. It works exactly like a high-performing analyst team.
Transparency isn’t an afterthought; it’s built into the architecture.
Soft signals are noisy. They shift. They contradict each other. But they’re also where most deals are won or lost.
The challenge is teaching AI to think probabilistically — to weigh patterns rather than fixate on single datapoints. We solved it by training agents to benchmark conditions against thousands of comparable environments. Relative reasoning, not absolute claims. That’s how experienced humans think.
Both — but the endgame is infrastructure.
Right now, it’s the most powerful analyst you can hire. Over time, it becomes the intelligence layer that sits underneath acquisitions, lending, asset management — everything. Underwriting is the foundation. If you control that layer, you influence the entire CRE workflow stack.
AI will compress the industry’s cognitive hierarchy.
- Analysts will spend less time gathering facts and more time validating strategy.
- Brokers will be armed with instant, data-driven narratives instead of intuition.
- Developers and GPs will differentiate on decision speed and clarity, not headcount.
- LPs will expect AI-driven transparency as the baseline.
The net effect: smaller teams, faster cycles, higher-quality decisions.
Most people underestimate how much underwriting is pattern recognition, not math. They also underestimate how inconsistent human judgment is — and how quickly AI can normalize that.
AI isn’t just automating tasks; it’s upgrading the fundamental decision mechanism. Once people see that, the adoption curve steepens dramatically.
Depth beats volume.
It’s far more valuable for the system to deeply understand decisions and outcomes than to skim tens of thousands of shallow examples. But the real power comes from combining both: deep reasoning across a large surface area.
That’s when the intelligence curve becomes exponential.
Diald becomes the default qualitative underwriting engine for global CRE. Not a tool — an expectation. Underwriting becomes faster, more accurate, and less dependent on who happens to be in the room.
If a small team in a secondary market can underwrite with the same clarity as a major GP in New York or Seoul, we’ve succeeded. We’re democratizing high-level judgment.
They start with data instead of physics — the underlying mechanics of the workflow.
Most founders build dashboards. They chase APIs. They try to “add AI.” That doesn’t work. You have to rebuild the workflow from first principles and then place AI where human cognition breaks down.
Don’t build AI to show off AI. Build AI that makes the fundamental bottleneck disappear.