How Credit Instruments Automate Their Own Failure

Written By Christoph Hinske

Context note: This post presents conceptual research reflections. It does not describe, assess, or report on any specific organization. All examples are synthetic or composite and non-attributable.

In a prior note, I argued that measurement blindness is not a missing indicator but a structural condition: frameworks calibrated to the wrong unit of analysis cannot register the field that shapes an organization before it acts (Calibrated to the Wrong Unit). This note addresses the question the argument leaves open. If the blindness is structural, why does it persist even where it is fully known?

The persistence is the puzzle, because the knowledge is old. Research on relationship lending established decades ago that soft information, relational and contextual, and hard to verify, carries predictive weight in small-firm finance (Petersen & Rajan, 1994; Berger & Udell, 1995). Stein (2002) even specified the organizational answer: keep allocation authority close to the person who holds the soft information, because that information does not travel. Liberti and Petersen (2019) further sharpen the definition: soft information is precisely the information that cannot be separated from the person who collects it. The gap, in other words, is one of the most explicitly acknowledged open gaps in SME finance. And it stays open.

The admission is easy.

A recurring pattern in conversations across different banking contexts: risk professionals state, without hesitation, that the quality of a firm's relationships is among the most relevant indicators of its resilience, and in the same breath, that their instruments do not capture it. Sometimes a third statement follows: that capturing it is unnecessary.

Two justifications alternate. One is economic: too costly and not scalable; the process must stay lean. The other is epistemic: not objectifiable, not verifiable, an invitation to arbitrariness. Both have long genealogies. Automated credit scoring entered small-firm finance as a cost-and-scale solution (Berger, Frame, & Miller, 2005), and mechanical objectivity has historically been the means by which institutions have addressed distrust of personal judgment (Porter, 1995).

What makes the configuration robust is not either justification but their alternation, often within the same person, depending on where the conversation stands. Challenge the cost argument, and the objectivity argument takes over; challenge objectivity, and cost returns. Argyris (1990) describes such moves as defensive routines: skilled, automatic, and invisible to the one performing them. The gap itself is discussable. The alternation is not.

Sincere, and without a port

Nothing here requires anyone to be lying. The professionals who name the gap are sincere, and many are frustrated by it. Some carry their observations upward, where they are not rejected but absorbed: the response stays within the prevailing logic of digitalization rather than reopening it. The talk that results may look ceremonial in Meyer and Rowan's (1977) sense, yet the sincerity behind it is real.

That sincere admission and intact practice can coexist has a structural explanation: the knowledge is person-bound, while the decision is instrument-bound. There is no port between the two. This extends an argument developed earlier for coordination roles (person-bound legitimacy). Here, it is not legitimacy that adheres to persons but the risk knowledge itself, which does not survive translation into formal instruments.

Automation does not create this condition; it expresses it. What is already measurable gets automated because it is measurable and cheap. Each round of automation devalues what resists standardization, making the next round more likely. Below certain decision sizes, the loop closes completely: no human judgment remains through which relational signals could leak into the decision. Stein's solution is not being rejected; it is being optimized away. And the segment below that threshold is precisely where most smaller firms live.

How failure confirms the instrument

If the configuration were merely stable, awareness might eventually erode it. The harder claim is that it tightens. Consider, as a structural reconstruction, the trajectory of a firm's failure through this configuration.

The relational precursors of the failure were never recorded, so within the system's own data, the event appears unforeseeable. The review that follows asks whether the process was followed, not whether the instrument could see. The event is filed in the only category the instrument recognizes, a financial default, so relational explanations never enter the competition. And by the time a response is due, nothing that remains of the event can implicate the instrument. Tightening control over the measurable is the only reaction that can be defended; loosening never is.

Each layer rests on the one before it. The failure that reaches the decision is no longer the failure that occurred. The system corrects inside its governing variables rather than examining them, single-loop learning in Argyris and Schön's (1978) terms, under conditions that narrow information processing exactly when it would need to widen (Staw, Sandelands, & Dutton, 1981). More control over the measurable means less room for human interaction, less capture of relational signals, and a next failure that will look just as unforeseeable as the last.

What this resolves, and what it opens

The question this note set out to answer: why does an explicitly acknowledged gap stay open for decades? Because the configuration does not persist despite awareness. It persists with it. Awareness travels through persons; decisions travel through instruments; and failures arrive at decisions already stripped of everything that could implicate the instrument.

This reframes what it would take to close the gap. Not more indicators; the prior note ruled that out. Not more awareness; the system has plenty. What is missing is a port: a way for relational, person-bound knowledge to enter decisions in a form that survives the objectivity test rather than triggering it, evidence rather than anecdote. What such a port would have to look like, and why it cannot be built by softening the instrument's standards but only by extending what counts as observable, might be the subject of the next note.

Download post as visual tool


References

Christoph Hinske
Previous

Why Your System Won't Move, No Matter Who Pushes
Next

Why I do not ride an e-bike or my use of Claude.ai