My AI Loan Servicing Journey – AI SECRETS REVEALED!!!

How a commercial-loan servicer uses Google NotebookLM to tame the paper-mountain, set up new loans swiftly and unlock ongoing value

I’ve always believed that good servicing is more than cash-collection and spreadsheets. For a commercial lender or owner of a servicing business, it’s about precision, transparency, responsiveness—and yes, a touch of magic when you get it right. As a guy who’s been in the servicing trenches, with portfolios that leverage scale, legacy systems, and myriad documents, I’m here to pull back the curtain. In this first edition of My AI Loan Servicing Journey, AI SECRETS REVEALED!!!, I’ll Walk you through how we at GOLDERSUN leaned into Google NotebookLM, loaded our document sets, and began to extract value—not just at loan setup, but across the lifecycle.

Why modernization matters

Let’s start with a truth most of us know but few talk about: servicing commercial real-estate debt is hard work. The documents are thick, the exceptions many, the processes often manual. Legacy servicing platforms were never built for the scale and the complexity we face today. At GOLDERSUN we had a big question: how do we keep up — and even get ahead — with fewer errors, faster turnarounds, more transparency?

Enter artificial intelligence. Not sci-fi “robots take over” stuff, but practical tools that help human beings do their jobs better, faster, smarter. My thesis: the companies that learn how to use AI as a “co-pilot” will win. The rest may still service, but they’ll be on the back foot.

In other words: the question isn’t if you’ll use AI—it’s how fast. As the Lewis Carroll-loving folks among us know, “Here we must run as fast as we can, just to stay in place. And if you wish to go anywhere, you must run twice as fast.”

Why NotebookLM got into the toolkit

There are plenty of AI tools out there: generative AI, large-language models, workflow bots, robotic process automation (RPA). But we selected Google NotebookLM for a specific use case: a document-heavy, data-rich, messy environment. Here’s why:

• We had sets of loan documentation in the hundreds, thousands even—especially for workouts, refinancings, takeovers. Loading them manually into a servicing system, abstracting data one line at a time, was slow, expensive, error-prone.

• NotebookLM lets you load large volumes of documents (PDFs, Word files, scanned exhibits) into a notebook environment, and then ask it questions (“What’s the interest rate?”, “What is the reserve account structure?”, “When is the next reset date and who triggers it?”). Google then uses its LLM backend to index, embed, and retrieve answers. That gives you speed + context.

• Because you can cite back to exact document pages, you get auditability. You’re not just trusting the machine—you’re verifying the machine. As the good ol’ Reagan rule applies: “Trust, but verify.”

• The platform integrates relatively easily with other tools. We built a lightweight middleware layer so the extracted data could be pushed into our servicing database, dashboards, exception workflows. No need to swap out everything.

So in short: NotebookLM became our secret weapon for transforming “tonnage of documents + manual abstraction” into “input documents → structured data → useful insights”.

How we used it for new-loan setup

Let’s walk through a typical flow for a new commercial loan we took over or originated, and how NotebookLM shortened the setup process and raised quality.

Step 1: Collect and load the documents.From the borrower, lender or prior servicer we pull: the loan agreement, promissory note, mortgage/deed of trust, guarantee, intercreditor agreement (if applicable), reserve account agreements, cash-management instructions, lockbox agreements, UCC filings, special provisions, subordination/ranking documentation, property management contracts, and any amendments. That easily runs 400 – 1,000 documents. For one deal we recently inherited, there were 1,030 documents. (Yup.)We upload them into NotebookLM, and the program reads the identifying features of the documents (borrower name, property location, effective date, etc.). We are trying to figure out how to integrate this functionality with our document storage system. Like AI overall, - so little time, so much cool stuff to do.

Step 2: Train the context.Once loaded, we create a notebook label for the loan (e.g., “Loan X123 – ABC Industrial”), or the program will name the collection of documents by itself and the ambitious robots start summarizing the deal to give a preview of what will be found in the documents. With this very small amount of setup, we can begin prompting NotebookLM for answers: “Give me a summary of this loan: maturity date, interest rate (initial / resets), amortization, prepayment terms, reserves, cash-management structure, other triggers.” Within seconds we get a summary. These answers are stored as “notes” for quick reference. Then we go through verifying critical data. NotebookLM returns page references and actual snippets of the documents, effectively bookmarking the data sources for our QC processes.

Step 3: Abstract the data.We built an extraction template, a series of “prompts” in AI LLM terminology that extract data that we need for setting up a loan: borrower, guarantor(s), principal amount, interest rate type (fixed/floating), margin/coupon, reset schedule, amort schedule, maturity, outstanding amount, reserve types and amounts, collateral description, property address, borrower cash-flow triggers, covenants, defaults/accelerations, investors pay-through structure, servicing fees, etc., you get the idea. NotebookLM finds the answer, cites the page, provides a bookmark to the source document and we paste or ingest into our servicing setup system. This moves us from days of manual work into hours (or less) for complete, and more thorough, setup. Rightfully so, the servicing system we use has series of logic checks before accepting data. Critical servicing data field inputs require our intervention. With these tools, we can take considerable time out of processing but the processes are not fully automated for the sake of security.   

Step 4: Validation and exception-flagging.Just because the machine churned the data doesn’t mean we’re hands-off. We spot-check flagged items, note unusual clauses, ambiguous language and contradictory documents requiring manual override. The system is surprisingly good at noting amendments and explaining how Amendment 1 change paragraph 4.4 of the original loan agreement.  We resolve exceptions, update the setup, and now the loan is ready in our servicing database with structured metadata, tags, dashboards and alerts configured.

Ongoing value: beyond setup

But the real magic isn’t just getting the loan set up. It’s what we do after. Once you’ve got a system like NotebookLM feeding your ecosystem, you unlock valuable ongoing queries, insights, risk-flags, and efficiencies.

1. Ad-hoc document queries.Borrower calls and asks: “What does the note say about subordination if we do a mezzanine piece?”. Instead of pulling 600 pages and reading, our servicing analyst asks the notebook: “In the Loan Agreement, what happens if there’s a mezzanine loan?” The answer is returned in seconds with page refs. That responsiveness = client trust + internal efficiency. A BIG HOWEVER, even if the answer given by the AI program is technically accurate, we do not give legal advice and would recommend a Borrower seek legal counsel.

2. Monitoring triggers and covenants.We set periodic notebooks or workflows: “List all upcoming interest reset dates within 90 days.” or “Find all reserve account sweep conditions.” The system had already abstracted the metadata so we can run these queries programmatically. When an item appears, we flag servicing or asset-management teams. This proactive monitoring cuts our risk and provides data for our workflow monitoring systems.

3. Document audit-trail & compliance.One of the longest-standing headaches in servicing: when an investor or auditor asks “show me the document that supports Clause 9.4 about the sweep account”. NotebookLM’s citation system means we can instantly point to document X, page Y. That helps us reduce audit risk and defend our work-product.

4. Efficiency gains = cost savings.When abstraction goes from manual (analyst reads 100+ pages, enters data) to semi-automated (notebook reads + returns snippet) we reduce time spent, reduce error, and free up staff to do value-added work (analysis, client communication, problem-solving). That cost advantage is real, especially at scale.

Lessons learned & cautions

Of course, no tool is perfect. Here are some practical lessons from our deployment:

• Document cleanliness matters. If you upload 300-page scanned images with poor OCR, the notebook struggles. Pre-processing (OCR cleanup, file naming, folder structure) is still critical.

• Model limitations = human oversight required. The notebook may mis-interpret ambiguous language or fail on very custom/legal‐drafting. Our rule: machine suggests, human verifies. The co-pilot, not the driver.

• Data ingestion matters. Extracted metadata must be loaded into our servicing system, dashboards, and workflows. Without integration, NotebookLM becomes a standalone silo—not full value.

• Change management: Some staff were skeptical (“AI will replace me”). We emphasized “AI frees you from grunt work; you still make the calls, build the relationships, add the judgment”. Human + machine.

What this means for the servicing business

If you’re in servicing—either as a sub-servicer, owner, or internal servicing team—this approach offers a blueprint:

• Speed = competitive advantage. Faster setups, faster delivery, fewer drop-outs in the onboarding process.

• Quality + transparency = trust. Investors, lenders and owners demand data, clarity, auditability. Structured metadata plus document-citations win.

• Scalability. Manual abstraction doesn’t scale. With AI-assisted workflows you build for growth.

• Proactive vs reactive servicing. When you stop chasing the documents and the exceptions, you get ahead of the risk.

• Human capital reallocation. Freed from repetitive tasks, your team adds more value: borrower relations, asset management, investor reporting, strategic analytics.

The big picture: what’s next

Where do we go from here? Having launched the initial NotebookLM workflows, at GOLDERSUN we’re starting to explore the next frontier:

• Perfect Loan Set-up integration. Build API’s into servicing system. Program not only key loan set-up components but go deeper into populating the “Rules Engine” that triggers downstream compliance processes.

• More Robust Reporting. Using internal data and external market data, maps, photos, etc. we will be able to produce higher production value reporting for our clients.

• Investor / borrower communication. We do not have current plans to automate any communication directly from the machines to third parties – No ChatBots. However, using AI to prepare communication is encouraged. After all, everyone that has played with AI recognizes that the robots are just better writers. I know this is true for at least myself. Alright, I admit, that the 25-year anniversary poem “I” wrote for my wife was actually the product of ChatGPT.

• Continuous learning loops. The more data we feed and the more prompts we refine, the smarter our system gets. Over time we hope to move from “setup aid” to “servicing assistant”. 

Final thoughts: trust the machines—but keep your human super-powers

If you’re reading this and thinking “AI is too risky / too early / too hype” — fair. But here’s the truth: the loans are coming, the demands are growing, the portfolio complexity is rising. If you don’t find a way to work smarter, someone else will.

We are not even close to experts in NotebookLM or other AI applications. Get a mentor to guide you through this rapidly-evolving technology. 

Ai use, only do there is, not try. Scary but exciting it may be. Yes, hrrrm.

For us at GOLDERSUN, NotebookLM isn’t a flashy novelty—it’s a workhorse. It helps us deal with scale, ambiguity, volume—and gives us back time for the real value: borrower relationships, lender service, risk management, transparent reporting.

Let me leave you with a takeaway: AI doesn’t replace the human heart of servicing. It amplifies it. The machines don’t negotiate, don’t build trust, don’t decide workouts. But they sure can give you the data, the speed, the clarity so you can focus on what humans are best at.

Call to Action

So, how are you using AI?

Have you found ways to make it more than just a back-office helper?

Drop me a note. I’d love to trade stories, compare battle scars, and maybe share a laugh about the beautiful chaos we’re all stumbling through together.