A young field examiner opens a completed receivables analysis. An AI system has already sorted the aging, compared the borrowing base certificate with supporting records and flagged a handful of unusual entries. The examiner reviews the exceptions, approves the work and moves to the next borrower. The team’s output rises.
Six months later, that examiner encounters a borrower whose records are internally consistent but economically misleading. Every number reconciles. The problem is the business explanation behind the numbers. Does the examiner know which question to ask?
That is the workforce challenge asset-based lenders need to put beside their AI productivity plans. The immediate test is whether automated analysis helps an experienced professional find a problem. The longer-term test is whether the next generation will learn to recognize the problem without being handed a red flag.
ABF Journal has already addressed the first question. In their Oct. 8 article, “AI Can Flag the Exception. Only an Examiner Can Explain It,” Donald F. Clarke and Dwight Clarke of Asset Based Lending Consultants compared the value of automated sorting with the examiner’s responsibility to interpret what a flag means. That distinction matters. But it leads to an uncomfortable follow-up: How does someone become the examiner capable of supplying that interpretation when much of the messy, repetitive work that once taught it is disappearing?
The Disappearing Apprenticeship
Professional judgment often develops through contact with ordinary cases before a truly difficult one arrives. A junior examiner who manually traces invoices, wrestles with customer reconciliations and follows a discrepancy into the underlying records begins to learn the difference between a clerical error, a pattern and a plausible business explanation. Much of that labor is slow. Some of it is exactly what technology should eliminate.
The training value of the labor, however, is easy to overlook. The novice is building an internal sense of what typical records look like, which departures matter and where the documentary trail stops being trustworthy. When automated systems present a polished exception list, the worker may never have to form those expectations independently.
Consider a hypothetical distributor whose receivables aging appears clean. An automated review correctly finds no unusual dates or arithmetic errors. Yet a large customer’s purchasing terms have changed, and credits are being negotiated informally. A skilled examiner may ask whether apparent collectability reflects contractual and commercial reality. A novice trained mainly to confirm dashboard flags might see nothing to investigate.
The problem is not an AI failure in that illustration. It is a human capability gap created when the institution mistakes efficient task completion for the development of expertise.
Why a Second Scorecard Matters
That conversation is taking place amid real developments in asset-based finance. ABF Journal’s September feature by Lon M. Singer of Riemer & Braunstein describes how collateral structures and data-driven tools have expanded beyond familiar receivables-and-inventory formulas. More kinds of information create more chances to identify risk, and more opportunities to miss its meaning.
In a December 2025 roundtable, ABF Journal’s technology leaders described improved document extraction, underwriting and monitoring while warning about domain-specific weaknesses and inappropriate reliance on AI. Those observations support investment in better tools. They also suggest that software performance and staff competence need separate evidence.
A lender can count minutes saved per borrowing base review, the proportion of exceptions flagged and turnaround time. These are useful productivity measures. They do not show whether junior staff are becoming better at independently evaluating collateral, challenging borrower explanations or recognizing when a technically complete file does not answer the credit question.
I would add a second scorecard: the ability of each experience level to detect, explain and escalate material exceptions before seeing an AI recommendation.
That measure need not mean subjecting every live credit file to duplicate manual work. A structured training set can contain realistic, anonymized or synthetic scenarios drawn from the types of issues the lender actually handles. Some cases should contain obvious discrepancies. Others should include plausible documents with a hidden business assumption, contradictory explanations or a legitimate exception that should be accepted rather than rejected.
Give the trainee the underlying file first. Ask for the two or three questions that would change the credit conclusion, the evidence needed to answer them and the point at which the employee would escalate. Only after that exercise should the trainee compare an AI-generated analysis and explain any disagreement.
That distinction is crucial. The learning objective is judgment, not guessing which flags the software will raise.
A Decision Trail Instead of a Checkbox
One inexpensive change would be to replace a generic “human reviewed” attestation with a short decision trail for consequential exceptions. The examiner would record what the system suggested, which source documents were independently checked, why the issue mattered to availability or collateral quality and what would justify reversing the initial conclusion.
An approval checkbox proves that somebody touched the workflow. A decision trail reveals whether the examiner exercised independent judgment.
That approach also improves the institution’s ability to learn from near misses. If an exception later proves significant, leaders can examine the reasoning rather than merely assigning blame to the person who approved it. If the AI incorrectly escalates a benign pattern, the file captures what the professional understood that the system did not.
For a lending operation, the questions might be straightforward: Is the apparent change in receivables aging explained by customer behavior, billing timing or altered terms? Does inventory eligibility depend on a fact outside the structured report? Have the relevant contractual or collateral assumptions changed since the last examination? What evidence would make the examiner reject the initial interpretation?
The answers will differ by portfolio. What matters is the habit of making the reasoning inspectable.
Effective Challenge Must Have Apprentices
There is an instructive parallel in banking supervision, with an important boundary. The Federal Reserve, the Office of the Comptroller of the Currency and the FDIC issued revised model risk management guidance on April 17, 2026. It emphasizes risk-appropriate controls, model limitations and “effective challenge” by people with the expertise and independence to question models. The guidance explicitly places generative and agentic AI outside its defined scope, so it would be incorrect to present it as a new regulatory requirement for generative AI systems. Its discussion of informed challenge remains a useful organizational analogy.
Where, then, will the effective challengers come from? Organizations cannot preserve expertise indefinitely by assigning every difficult judgment to the same handful of senior people. Those experts retire, move firms and face growing workloads. New analysts have to build the pattern recognition that makes an exception meaningful.
The Government Accountability Office’s 2025 review of AI in financial services likewise weighs the efficiency gains of AI against risks that include biased lending decisions, data quality problems, privacy concerns and new cybersecurity threats. That evidence does not establish that any particular asset-based lender has a skills crisis. It shows why an institution should test the quality of oversight rather than assume it from a job title.
A practical apprenticeship can combine three stages.
First, ask juniors to work a small sample of raw files without machine recommendations and record their questions before viewing the model’s output. Keep the task limited and purposeful so it develops judgment without re-creating entire manual workflows.
Second, have experienced examiners discuss why they questioned a borrower explanation, accepted a seemingly unusual transaction or refused to accept a comfortable answer. A narrative about the decision is more instructive than simply showing the correct red flag.
Third, give juniors graduated authority. Start with low-consequence reviews, measure how they handle disagreements, then expand responsibility as their independent decisions improve. The measurement is whether they can explain sound judgments across varied cases, including cases the AI gets wrong.
Senior examiners need training too. They must learn to recognize when an apparently compelling automated summary has omitted an important question. Experience with paper-heavy processes is valuable, but it does not automatically transfer to scrutinizing AI outputs.
Make Capacity Transfer Part of the Business Case
There is a legitimate counterargument: Asking employees to re-create work already done by software wastes resources. A lender facing competitive pricing and demanding turnaround times cannot afford educational exercises on every transaction. That is true. The answer is to reserve deliberate practice for a small, carefully chosen part of the workflow and demonstrate its value.
A pilot might compare two groups of newer examiners. Both use the same AI tools in production. One receives ordinary workflow training; the other also practices on unseen exception cases with structured explanations from senior staff. After an appropriate training period, both groups receive fresh cases that test recognition, explanation and escalation. Record accuracy, false positives, material misses and confidence calibrated against the evidence. Avoid turning a tiny pilot into a sweeping statistical claim.
The test could reveal that additional practice adds little. It could reveal a substantial gap. Either result would be more useful than assuming that greater document throughput means the future workforce is getting stronger.
A lender could also track its concentration of critical judgment: Which decisions require the same two senior specialists? Which tasks can juniors perform independently after three months? What happens when a familiar borrower presents an unfamiliar exception? Those are operational resilience questions with implications for succession, staff retention and credit quality.
A Stronger Argument for AI
The best case for AI in asset-based lending is that it gives examiners more time to investigate questions worth asking. Faster document processing and broader monitoring are important achievements. They become more durable when the people using the tools continue to grow in expertise.
Leaders should therefore make two promises when they introduce automation: better credit work now and better credit professionals later. Measure both.
The next examiner who spots an unflagged problem will demonstrate the value of that second promise. The institution should be able to explain how that employee learned to see it.
Gleb Tsipursky is CEO of Disaster Avoidance Experts, a behavioral scientist and author of “The Psychology of AI Adoption at Work: From Resistance to Results” (Georgetown University Press, 2026). Learn more at disasteravoidanceexperts.com/aibook.