Last week, I decided we weren’t moving fast enough to keep up with progress on the AI front. At Pensford, I formed an internal AI team. At LoanBoss, I shut the engineering and product team down for the month of March (which I am calling March Madness) to do nothing but learn Claude Code.
The reason? The pace of change in AI is crazy and I don’t want to get left behind.
The catalyst? Anthropic’s Claude Opus 4.6.
I’ve talked a bit in here about how we use AI for abstracting loan docs very successfully. But the leap forward with the most recent Claude model is mindblowing.
- We are abstracting Freddie and Fannie loans in under 3 minutes with 100% accuracy.
- We are abstracting the nastiest construction loans in under 15 minutes with 90%+ accuracy.
That accuracy was more like 15% a year ago.
This isn’t for an executive summary, this is up to 400 fields per loan, down to stuff like daycount conventions. Here’s an image from our demo account of the sort of stuff we capture.
Not just data points, either, but the tedious stuff like Insurance.
For the first time, AI is truly transformational for our business.
Until now, abstracting loan documents has been our biggest bottleneck. That same nasty construction loan might take our abstract team 4 hours. Now it’s 15 minutes?
And I can spin up as many AWS servers as needed, so theoretically we could abstract thousands of Freddie and Fannie loans in 3 minutes. We could abstract every construction loan originated this year in 15 minutes.
Remember, we have our own AI model via our relationship with AWS. Like all relationships in real estate, the more you spend, the more “special” the relationship. We. Spend. A. Lot. So we get our own model.
This model is walled off from the Claude you and I interact with via a browser, so customer data isn’t used in their training. But we also aren’t limited by the typical token limitation. I just have to be willing to pay for it.
And before you freak out, we still have two independent rounds of QA by trained human loan specialists. But now our highly skilled abstractors can focus on the hardest, most
complex loans rather than the easy ones. As a business owner, they are focusing their time in the most impactful places. Isn’t that what we want for all of our teammates?
Think about why we need to abstract loan docs in the first place: turn unstructured data into structured data. That’s it. We need it in a structured format if we want to turn it into a report. Or set up a notification. Or…layer AI on top…
The bottleneck for all of us is the time/effort to do it.
We run a mental calculation: is it worth my time to turn this unstructured data into structured data?
The implication is: what benefit will I have when I’m done?
For us, a core value prop of LoanBoss is turning unstructured data into structured data that you can access through those pages I included above. To achieve this, I had to build an in-house team to abstract loan docs. We have to capture every edge case across all borrowers and lenders. I can’t opt out of abstracting loan data because then I can’t charge you. I’m nothing if not a capitalist.
You may have reached a different conclusion. You may have decided the value prop drops off after the fifty most important things abstracted. Or 10. Or maybe you just don’t do it at all. If there’s no benefit on the far side of the effort, why bother?
We get requests all the time to abstract things like JV Agreements, mortgage statements, Term Sheets, etc. I had to say no because the effort required from our team didn’t justify the benefit. I decided abstracting term sheets wasn’t important enough to distract us from abstracting loan docs. My math was the same as yours, just at a different part of the debt supply chain.
But now?
If we can abstract JV Agreements in under 3 minutes, why wouldn’t we? Turning unstructured data into structured data.
If we can abstract mortgage statements in under 30 seconds, why wouldn’t we? Turning unstructured data into structured data.
And Term Sheets. Brokers have thousands of term sheets just sitting somewhere in a folder. Their clients have upcoming extension tests that they would love to know about,
but they didn’t spend the time at closing to turn unstructured data into structured data. They ran the math and concluded it wasn’t worth the effort. Of course they want the extension details three years after closing, but have concluded the effort required to do so is not worth it.
At its core, the recent jump in AI model performance allows us (and you) to turn unstructured data into structured data with much less effort.
I don’t know what awesome things we are going to want to do with AI, but I know it always begins with clean, structured data. The effort to convert unstructured data into structured data is plummeting.
A few weeks ago, I shot down the SaaSpocalypse selloff. I believe it was overdone more now than ever. The experience of working with a SaaS company like LoanBoss might be evolving, but the core value is the same: turning your unstructured data into structured data.
When I say “software” you think of the screens you interact with, but those are just the delivery mechanism for the real value: structured data.
You still need structured data, but now you’ll pay less or get more or get it faster or more accurately etc.
One prospect this week said, “This is a data project that comes with free software.” Even if no one at the company ever logs into LoanBoss, he felt like this set them up to take advantage of AI in a way they can’t currently. Their property data has been structured for 15 years, but their loan data is trapped inside of PDFs.
Turning unstructured data into structured data. That’s it.
Next week, I’ll go into more detail about how I decided to basically shut the companies down for the month of March so every single teammate begins working exclusively out of Claude. It’s a huge (read: expensive) gamble, but I genuinely believe that we need to get our house in order ASAP.
One last thing before you go…
AI Survey
The guys at CRE Analyst are some of my favorite in the business. We send about 10 teammates a year through their training courses and I can’t recommend them highly enough.
They are conducting a survey on how CRE folks are actually using AI. The results are anonymous and I promise they aren’t going to cold call you. But I think this benefits all of us, so I am hoping you will take two minutes to fill it out: AI Survey
