Available Data vs. AI Needs

Monitoring: Core ABL Concerns

AI Data vs. Boolean analysis

A lack of data keeps Boolean methods (IF, Then, Else, And, OR) valid in an AI obsessed world. AI needs data to dive deeper, but that data is often shallow, particularly with bank lenders.

If we believe collateral is the primary source of repayment, it should exist.

Items ship or services are rendered to create invoices
Inventory physically exists and is priced appropriately
Fixed assets are in place, in maintained working order and are not cannibalized for parts used in other machines.

Basically it exists, it has title transfer to a buyer, and it is unencumbered collateral with our liens on it before any other liens. However, we need to monitor this stuff, and that has changed over time due to product offerings and technology. We do get more data electronically now, but is it enough for AI Data vs. Boolean analysis?

Monitoring: The Old Way

For those of you old enough to remember that we originally did field exams with green or amber accounting paper and pencil and calculators. Way-back, those calculators were mechanical adding machines to lug about. We had deliveries by courier and eventually FedEx for daily sales and cash journals or monthly reporting. The Internet was not invented yet.

The data gathering was also limited because things like trend-cards were manual. This reliance on manual tools slowed the collection and verification of information during the early stages.

Along came the calculator and then the PC and eventually an affordable laptop like the Toshiba T1000 (circa 1988). This progression marks a shift from manual to digital aids in record keeping.

Hard drives eventually arrived and they got more capacity fast. But audit reports still had "This exam vs. last exam vs. FYE" and eventually those were in spreadsheets like VisiCalc, Lotus 123 and Excel.

Computer systems came along to monitor and track loan availability, interest calculations, covenant compliance, renewals, etc.

Risk Levels: Reporting and Data Depth

Banks tend to lend to more credit worthy customers at lower rates. That means less monitoring, perhaps just monthly reporting. The lender may require billing journals and cash receipts journals, daily sweep account deposits, or lock box direct receipts. Maybe that data comes in as emails, upload folders, or other formats. Maybe the lender does a roll forward of collateral based on billings and cash receipts, etc.

Compare the above to a private finance company that takes more risks, charges more interest and fees and gets daily reporting, perhaps recalculates eligible collateral with each borrowing.

Also consider that you don't write-off a ton of customers, you see deterioration and you increase monitoring in the hot spot companies (late reporting, losses, a lack of excess revolver availability, etc.). Those are the glowing risks. Again, banks and finance companies take up two different risk levels and reporting documents here.

Banks are not suddenly going to ask for, nor get, additional reporting on the regular accounts. Ongoing practices, the Loan & Security Agreement, the risk levels, etc. are limiting the data received. Problem accounts, forbearance agreements, etc. might change specific borrower's bank reporting, but the norm is lighter data loads for banks. Asking for more reporting is not likely to happen.

You have very different data depth between banks and private lenders.

Typical Modern Analysis

We'll start by noting that programing just works now and will still work in the future, it is still needed, pervasive and serves as the foundation of how we evaluate things. Even programmed AI results are often filtered with programming Boolean logic, etc.:

Programmers

Programmers use lots of Boolean (If, Then, Else, AND, OR) logic and most financial analysts and field examiners are familiar with IF statements, SUM, MAX, MIN, Average, SUMIF, AND, OR etc. calculations. Programmers work with subject matter experts (SME) such as Officers, Loan processors, Examiners, management, etc. to create input boxes for data that then analyze, summarize, process and output reports and information (screen or printed), based on available data.

Circuit Breakers

But those SMEs have years of experience to note critical concerns to trap, block, highlight, etc. It can be as simple as a line limit or highlighting any customer that is greater than $20,000.00 or any invoice greater than $20,000.00. These "circuit breakers" are as valid now as they were 40 years ago. Some great SMEs don't communicate well with programmers or they lack the time to get into more details. Revisions followed by revisions, and it evolves over time. There are many more complex things to evaluate and calculate, but once you set these up in a program, you get some hits and people take action to investigate... "Mr. Borrower, can we please get the P.O. and shipping documents for the following invoices?"

AI does not know much about fraud and your SMEs are likely more experienced about seeing issues and finding anomalies. Coding those concerns with Boolean logic has been done over and over again, yet AI is going to miss these things.

Downloading Data

If you are not downloading data for field exams and the back office, you are spending money on human resources instead. That payroll related cost is about 20X that of the download software. Sure, we produce AssetReader for downloads of Excel, PDF, Text, HTML, DOCX, delimited data, etc. and it is still surprising how many people still process collateral reports manually to this day. We have also seen outsource firms that don't download data efficiently and that makes the field examination cycle take longer, require more staff and ultimately increase the costs of the exams. Are you data download enabled as the norm?

Enter AI

In the 1986 Star Trek movie The Voyage Home, the chief engineer, Scottie, tries to talk to a McIntosh computer. "Computer?... Computer?" and he even talks into the mouse handed to him by "Bones" with "Hello Computer" (See it here). You can imagine how niece it is to ask for what you want and not type it.

With AI and data, you get to type (or dictate) human language prompts that lead to responses and you don't need a programmer. The age of end-user data queries is here now if the data is available to the user. Assume that you have a few thousand hours of prompt building experience and that your prompts work and are repeatable, then you have a fast non-programmer solution to some slick analysis. But there are a few catches and unfortunately we have seen people making bad mistakes with CoPilot, OpenAI, Claude, etc. Many use AI with untested assumptions.

AI Promises With Shallow Data

Repeatability

A lack of prompting experience is pervasive and people don't even know they are approaching the prompts in an unstructured way. "Blind prompts" are commonly used in web queries for things like "Who might win the world series this year?" But that can also give different answers at different times. We learned in grade school about sentence structure, paragraphs, writing styles, etc. Yet AI prompts have no standards [yet] to make them work repeatedly. There are ways to get structure and repeatability, but most users lack the knowledge. That means results can be hot and cold or pure hallucinations. We've spent days on prompts that don't work and when we find the driver that caused the issues it is almost always a lack of detail and a poor choice of words or failure to structure the prompt query with adequate details and order.

Prompt Testing

Running an AI query is exciting when you get numbers on the screen that summarize your data and seemingly answer the question. But the skeptical among us will test numbers with proofing models and repeat those tests with varying data to see if the AI can hold up to different circumstances. Those fast and simple non-programmer queries are not always right, and the more complex the query, the more likely you will generate flaws. Testing is necessary and it is often humbling. On-the-fly prompts are not tested and pose the greatest risk of errors.

AI Needs Data and You May Have Little

The big "BUT" is: Do you have the data for AI to work? Remember, conventional Boolean logic, SME expertise and circuit breakers work. Banks typically lack daily data reporting to make AI worthwhile. In fact, we've seen where Boolean logic is better than the AI models for limited data.

Example 1: Our AssetArchive program uses circuit breakers to note "dirty needles in the haystack of data" accumulated from detailed monthly aging reports. We ran AI Data vs. Boolean analysis on the same data; our results were faster and more cost effective.

Example 2: Our AssetWriter-AI product writes a field exam in about 3-5 minutes. It has tons of data from the field exam, footnotes, our libraries and it was programmed to identify what it doesn't know so that further investigation needs are flagged. It uses AI and Boolean and many years of knowledge, plus a few thousand hours of programming for just the AI part. Data matters in this case.

But feeding AI some things to look for is still a good idea. Financial statement intake, some report matching, etc. is typically done with scans or images. But AI is not going to know that the AR has no reserve for doubtful accounts or that the deposits are stuffed into Accrued-Other. But for routine reporting with limited data, is it better than Boolean logic, with limited data? Not likely.

On the other hand, if you have daily reporting (per borrowing reporting, etc.), and if you get more reports, then you have more data to analyze and AI can sift through a ton of data quickly, without the Boolean logic or in complement to the Boolean logic. Results can be off (see the above) without proper prompt structure and from a lack of testing. This blog is about data and AI, not necessarily using AI bots to do work like matching documents (banks generally don't match documents because they don't get documents).

Conclusion

AI results improve with deeper data. Additionally, your lending risk levels, reporting cycles, reporting depth, and monitoring cycles may lack data depth. However, Boolean logic is likely enough with limited data. In AI Data vs. Boolean analysis terms, AI makes more sense with detailed reports, but a lack of detail is normal for banks, and that bank-case is lacking support for a lot of the possible AI methods.

August 27, 2026
Joseph Caplan, CPA, Managing and Creative Director
Shark in Data Waters image generated with Microsoft Copilot (AI). Concept and design prompts by Joe Caplan, CPA.