Why Credit Analysts Need Rules, Not Robots

“That’s why I’m still employed and AI hasn’t taken over just yet — because we have to look at it on a case-by-case basis.”

That’s a credit underwriter we spoke to recently. She wasn’t being defensive. She was stating a fact about how good lending decisions actually get made.

Her team doesn’t use a checklist. They don’t run applications through a scoring matrix and accept whatever comes out. Every deal is manually underwritten. They look at each one individually, because the numbers alone don’t tell the full story.

Seeking the Story

What stood out from the conversation was this: even when the data looks like a clear decline, her team goes back to understand the context before making a decision.

Missed payments? Maybe the client’s accountant was off sick for two months. Over the overdraft limit? Could be a seasonal business that always dips in winter. A farmer with 12 months of statements showing no income from November to February isn’t a bad credit risk — that’s just how farming works.

She called it “seeking the story.” The raw transaction data is the starting point, not the answer. The analyst’s job is to understand what happened and why, then make a judgement call that a simple pass/fail system never could.

The Problem with Automated Decisioning

There’s a reason experienced underwriters are sceptical of anything that sounds like automated decisioning. They’ve seen what happens when you reduce complex business situations to a score.

A rules engine that auto-declines based on missed direct debits would reject the seasonal business that’s been profitable for 15 years. An AI model trained on historical decisions would encode every bias in the training data. A rigid affordability check would miss the fact that last quarter’s low revenue was a one-off, not a trend.

For lenders who compete on judgement — who win business precisely because they look at deals that others auto-decline — an automated decision system doesn’t just miss nuance. It actively undermines their competitive advantage.

What a Rules Engine Should Actually Do

This is where the distinction matters. A decision engine for credit analysts isn’t about replacing judgement. It’s about making that judgement faster and better informed.

Think about what an underwriter actually does with a stack of bank statements. They’re scanning for patterns: are direct debits being returned? Is the account regularly over its limit? Are there gambling transactions? Is income consistent or volatile?

These are the decline drivers that come up in every conversation we have with lenders — affordability and payment history. Rejected direct debits, unpaid standing orders, returned cheques. The analyst needs to find these signals in pages of raw data before they can even start making a decision.

A rules-based engine does this extraction work. It flags the patterns. It surfaces the evidence. Then the analyst reads the evidence, considers the context, and makes the call.

• Affordability flags — income vs. expenditure trends, irregular revenue patterns, overdraft usage
• Payment history — returned direct debits, missed payments, bounced transactions
• Risk indicators — gambling spend, high-interest lending, county court judgement payments
• Evidence packaging — structured data the analyst can review, interrogate, and reference in their decision

The output isn’t a decision. It’s structured evidence for the analyst who makes the decision.

Human Accountability, Machine Efficiency

The underwriters we work with don’t want a system that tells them what to do. They want a system that saves them the two hours of manually reading through statements so they can spend that time actually thinking about the deal.

There’s a meaningful difference between “the system declined this application” and “the system found three returned direct debits and a declining income trend — here’s the data, what do you think?” The first removes accountability. The second enhances it.

When a lender’s competitive edge is their underwriting judgement, the technology should amplify that judgement, not replace it. The analyst who understands that a farmer has no winter income isn’t making an exception to a rule — they’re demonstrating exactly the kind of expertise that makes manual underwriting valuable.

Rules That Serve the Analyst

This is the approach we’ve taken with ExactSum. The rules engine extracts structured data from bank statements, applies configurable checks, and presents the results as evidence. Pass, refer, or fail — with the supporting data visible and auditable.

The “refer” category is where it matters most. These are the deals that need human eyes. The ones where the data alone says one thing, but the context might say another. A good system doesn’t try to resolve that ambiguity. It identifies it, packages the relevant evidence, and puts it in front of someone qualified to make the call.

Because the underwriter who looks at a struggling account and asks “what’s the story here?” before deciding — that’s not inefficiency. That’s the whole point.