By Heorhi Tratsiak
I approved a loan once that I probably should not have approved.
The man sitting across from me was in his late forties, a construction contractor, and his financials were borderline. Not bad enough to reject outright, not clean enough to approve without thinking about it. His credit history had one gap I could not fully explain from the documentation. His stated income was plausible but hard to verify completely. By the scoring model, he landed in a zone where the answer was technically a judgment call.
What pushed me toward yes was something the scoring model did not record. The way he talked about the project he was borrowing for. The specific numbers he used when describing his materials costs, numbers that told me he had actually priced the job and not invented a figure for the application. The fact that he had brought more documentation than I asked for, not less. The particular kind of steadiness in how he answered my questions.
He repaid the loan on time, ahead of schedule.
I am not telling this story to suggest that loan officers are infallible or that instinct is a reliable substitute for analysis. I am telling it because it captures something real about how financial decisions actually get made, something that sits underneath all the infrastructure described in the previous chapter, and that determines, more than any technical system, whether financial institutions survive or fail.
That something is trust. And understanding what trust actually means in financial systems, where it comes from, how it behaves under pressure, and why it cannot be fully engineered away, turns out to be essential for understanding everything else about how money works.
The Paradox at the Heart of the System
Here is a fact that should stop you for a moment.
The global financial system, the most sophisticated information-processing infrastructure in human history, runs on something it cannot directly measure. Every bank, every transaction, every credit decision, every correspondent banking relationship ultimately depends on a judgment about the future behavior of a counterparty. And future behavior cannot be measured. It can only be estimated. The estimate is always, at some level, a bet on character.
Banks have developed elaborate tools for making that bet more systematic. Credit scores. Financial ratios. Cash flow analysis. Collateral requirements. Covenant structures. Stress testing. All of these are attempts to reduce the uncertainty around future behavior by measuring proxies for it: past behavior, current financial condition, the size of the cushion available if things go wrong.
But the tools are always proxies. The thing they are trying to estimate, the probability that a specific borrower will continue making payments when circumstances change, is not directly observable. And no model yet built can fully close the gap between the observable proxies and the underlying reality they are meant to capture.
This gap is where trust lives. It is not a soft concept or a philosophical abstraction. It is the operational space in which every real financial decision is made. And working inside a bank for seven years gave me a close view of how that space actually functions.
Two Layers in Every Decision
When you apply for a loan, what you experience is the formal layer of the decision process. You submit documents. The bank checks your credit history. It calculates ratios. It applies its risk policies. These steps are real, and they matter.
What you do not see is the second layer, which runs simultaneously and which a skilled credit officer never stops doing.
This layer is not about the documents. It is about the person.
How did they describe their situation in the initial conversation? Were they precise about numbers, or vague? Did they mention the risk in their plan before I asked about it, or did I have to pull it out of them? When I asked a question they had not prepared for, what happened in the room?
These are not soft impressions. They are data. They carry information about things the documents cannot carry: whether the person’s understanding of their own financial situation matches the numbers on the page, whether they are the kind of person who solves problems when they arise or avoids them, whether they have thought through what happens if the scenario they described does not play out as expected.
I noticed early in my career that the most experienced credit officers I worked alongside asked fewer questions than I did. Their conversations with clients were shorter. They got to a decision faster. At first I assumed this was confidence bordering on carelessness. Over time I realized it was pattern recognition. They had seen enough situations that they knew which signals mattered and which were noise. They knew where to look.
What they were reading was not the documents. The documents were a given. They were reading the person’s relationship to their own situation.
What the Scoring Model Cannot Capture
The modern credit scoring system has made consumer lending dramatically more efficient. Before standardized scoring models, lending decisions required significant manual review of every application. Scoring made it possible to process hundreds of thousands of applications with consistent criteria, dramatically reducing the time and cost involved.
But the efficiency gain came with a specific kind of loss.
A credit score is built from historical data. It measures what a person has done in the past, under the conditions that prevailed while the history was being built. It says almost nothing about what that person will do in a future that may not resemble the past.
This works reasonably well for large populations over normal economic conditions. The statistical relationship between past behavior and future behavior is real enough that models built on it produce useful predictions on average. The problem is that lending does not happen on average. It happens one decision at a time, in specific circumstances, with specific people whose individual situations may diverge significantly from the statistical average their scores represent.
I once reviewed an application from a woman who had almost no credit history. She had not borrowed money before, had never had a credit card, had lived for forty years in a financial world that operated mostly outside the formal credit system. Her score was low not because she was a bad credit risk but because the system had almost no information about her.
Her documentation, on the other hand, told a clear story. She had managed a household budget carefully for decades. She had savings. She had a specific, modest purpose for the loan. She understood what she was committing to. She could describe, in concrete terms, exactly how the loan repayments would fit into her monthly expenses.
The score said borderline. The person sitting across from me said low risk. These were two different reads of the same underlying question, and they pointed in different directions.
The scoring model had no way to capture forty years of financial discipline that happened outside the formal system. The credit officer did. Not through any magical ability to read minds, but through the accumulated pattern recognition that comes from sitting across from thousands of people and paying attention to what they say and how they say it.
Trading, Psychology, and the Six Inches Between Your Ears
The clearest view I ever got of the gap between analytical frameworks and human decision-making came not from lending but from trading.
For several years, while working at the bank, I traded markets in parallel. I had studied technical analysis seriously, understood chart patterns and indicators well enough to form views. I had a framework. I had risk management rules. On paper, I had everything you need to trade systematically.
In practice, I kept doing things my framework said not to do.
I would hold a losing position longer than my rules allowed because I was sure it would come back. I would exit a winning position early because I was afraid of giving back the gain. I would increase my size after a series of wins, when the rational response was to hold size constant, because winning felt like confirmation of skill rather than a temporary run of favorable variance. I would freeze on a setup I had analyzed correctly because the moment of actual execution felt different from the moment of analysis.
What I was experiencing was the same thing that operates in every financial decision: the gap between what rational analysis recommends and what a human being under uncertainty actually does. The gap that no scoring model, no risk framework, no technical system fully closes.
In trading, the consequences of that gap were immediate and measurable. A bad lending decision might not show up as a loss for months or years. A bad trading decision shows up on your screen within minutes. This feedback loop made trading an unusually good teacher of something that applies far beyond trading: the way that psychological state, not just analytical ability, determines the quality of decisions under uncertainty.
The traders and analysts I most respected, both inside the bank and in my own reading, were not the ones who had the most sophisticated analytical frameworks. They were the ones who had developed the clearest understanding of how their own psychology affected their decision-making, and who had built disciplines to manage that effect. They did not eliminate the human variable. They studied it in themselves and worked with it rather than pretending it was not there.
This distinction matters enormously for how we should think about artificial intelligence in financial decision-making. AI systems are very good at removing certain kinds of human error from decisions: the inconsistency that comes from fatigue, the anchoring to prior cases that distorts judgment, the availability bias that makes recent events feel more predictive than they statistically are. These are real improvements. They make certain categories of decision more reliable.
What AI does not do, and cannot yet do, is replicate the kind of judgment that comes from understanding a specific human situation in its full complexity. Not because AI lacks computing power, but because the information required to make that judgment is not in the documents. It is in the conversation, in the behavior, in the long chain of micro-signals that an experienced person reads without consciously naming them.
When Institutional Trust Is Tested
The discussion of trust in financial systems becomes most concrete when you look at what happens during a crisis.
Financial crises are almost always, at their core, trust failures. The 2008 financial crisis was not primarily caused by bad mortgages. Bad mortgages existed. But what turned a housing correction into a global financial catastrophe was the sudden, comprehensive collapse of trust between financial institutions. Banks stopped lending to each other overnight because no one could determine which institutions were holding how much exposure to which securities. The information problem was real, but its effect was amplified enormously by the trust problem sitting beneath it.
When you cannot trust your counterparty, you stop transacting with them. When enough institutions stop transacting with each other simultaneously, the system stops working.
In 2020, my country went through a period of intense political and economic stress. The institution where I worked faced circumstances that no risk model had been specifically designed for. The chairman, who had been the face and the direction of the bank for years, was suddenly absent. The bank’s future was genuinely uncertain. The people inside it, several thousand employees, did not know from week to week what the next development would be.
In that environment, clients came to us every day with transactions to process, loans to discuss, questions to ask. And what I observed was that the bank continued to function not because the technical systems continued to work, though they did, but because the people inside it continued to show up and make decisions.
Clients who came in during that period were not coming in because the bank’s risk-adjusted return on equity looked favorable. They were coming in because they had accounts there, because they had talked to specific people there, because they had a history there that meant something to them. They were drawing on years of accumulated trust that had been built one transaction at a time, one conversation at a time.
I watched clients who clearly knew that the institution was under pressure still choose to keep their accounts there. Not out of ignorance, but out of a judgment they were making that went beyond the balance sheet. They were betting on the people they knew, on the track record they had experienced, on something that did not show up in any financial statement.
That bet, in our case, was rewarded. The institution navigated the period. It continued to operate. The trust held.
But what I carried away from watching that process was not a lesson about institutional resilience in the abstract. It was something more specific: that trust, in financial systems, is not just a nice feature. It is load-bearing. It is the thing the structure actually rests on. And it accumulates slowly, through thousands of small interactions that feel routine in the moment, and dissipates fast, in ways that no technical system is equipped to reverse.
The Asymmetry That Changes Everything
Trust in financial systems has one property that makes it unlike almost any other asset a bank holds. It accumulates slowly and dissipates fast. The ratio between the two is not even close.
A bank builds its reputation through years, sometimes decades, of consistent behavior. Payments processed correctly. Commitments honored. Problems resolved fairly. Each positive experience contributes a small increment to the trust balance of everyone involved. No single interaction builds it dramatically. It compounds through repetition.
A trust crisis, by contrast, can develop in hours. News of a single significant failure travels through depositor networks at a speed that no risk model predicted until relatively recently, when social media made it possible for financial panic to organize itself in real time in ways that would have taken days or weeks in earlier eras. The depositor runs that destroyed institutions in previous centuries took time to build because information moved slowly. Modern depositor behavior can escalate in a single afternoon.
I became acutely aware of this asymmetry during the stress period of 2020. The dynamics I observed were not primarily digital. They were human, moving through conversations in queues, through phone calls between family members, through the particular kind of social information transmission that happens when people are frightened and uncertain. But the speed was instructive. Years of accumulated institutional behavior were being weighed against days of uncertainty. The weights were not equal.
What protected the institution during that period was not any single dramatic intervention. It was the depth of the trust reserve. Clients who had interacted with the bank over many years, who had processed significant life events through its services, who had relationships with specific people on the staff, were drawing on something real when they chose to remain. Their trust had been built through hundreds of individual experiences, none of which had felt significant at the time. In aggregate, they formed a buffer.
Institutions that had been present for a shorter time, that had attracted clients primarily through pricing rather than relationship, had a thinner buffer. The same level of uncertainty produced more client movement, more outflow, more fragility.
This is a lesson that gets rediscovered in every financial crisis and then forgotten again in the periods of stability between them. The lesson is that the trust buffer is built during calm periods and spent during stressed ones. Institutions that manage it only reactively, that think about trust only when it is already under threat, are perpetually underinvested in the asset that matters most.
The Geography of Trust
One thing that became clear to me over years of working in a specific regional context is that trust in financial systems is not uniform. It is shaped by history, and the history differs dramatically across different parts of the world.
In markets where institutions have been stable for generations, where property rights have been consistently enforced, where banks have not disappeared with depositors’ savings within living memory, the default posture of the population toward financial institutions is something close to ambient trust. People keep their savings in banks without thinking much about it. They take out mortgages that run for thirty years because they assume the counterparty will still exist and honor the terms three decades from now.
In markets where these conditions have not held, where institutions have failed or been seized, where currencies have collapsed and savings have evaporated, the relationship between the population and financial institutions looks quite different. Trust is more conditional. It is extended more guardedly, based more heavily on recent experience and personal relationships than on institutional reputation or regulatory assurance.
This geographic variation in trust has enormous practical consequences for how financial products work, what risk models are valid, and what kinds of fintech solutions actually address the underlying need versus which ones are transplants that fit a different context.
I watched this play out directly in my work. The credit products and decisioning approaches that worked in Western European contexts were not simply transferable. The assumptions embedded in them, about client information quality, about the reliability of credit history, about the stability of income patterns, about the social meaning of default, were not the same. Adapting them required not just technical modification but a genuine rethinking of what the product was actually solving for.
Fintech companies that entered post-Soviet markets with products built for higher-trust environments consistently underestimated this adjustment. The failure was not usually technical. The technology worked. What failed was the implicit model of human behavior embedded in the product design. Real people, operating in a context of legitimate historical skepticism about financial institutions, did not behave the way the model expected.
This does not mean innovation is impossible in lower-trust markets. It means the innovation has to be appropriate to the trust environment as it actually is, not as it is assumed to be. Building trust in these markets is a different project from maintaining trust in markets where it already exists, and conflating the two is one of the most common and costly errors in financial services expansion.
Why Fintech Keeps Running Into This Wall
For the better part of two decades, financial technology companies have been promising to transform banking by replacing the costly, inconsistent, relationship-dependent human layer with cleaner, faster, more scalable technology.
The promise is not dishonest. Technology has genuinely improved many aspects of financial services. Payments are faster. Account opening is easier. Credit decisions for standard consumer products are more consistent and less subject to certain biases. These are real improvements.
But the more fundamental promise, that technology could displace the relationship layer and produce better financial outcomes at scale, has proven considerably more difficult to fulfill than the early enthusiasm suggested.
The reason comes back to trust, and to where trust actually comes from.
Trust in financial relationships is not primarily about efficiency. It is about accountability. When a specific person at a bank makes a decision about your application, there is a human being who made that call and who can be held to it. When an algorithm makes the decision, the accountability structure is different in ways that most people experience as a loss, even if the outcome is identical.
This is not mere irrationality. It reflects something real about the nature of complex financial relationships. When circumstances change, when something unexpected happens to your business or your income, you need to renegotiate. You need to explain. You need to be heard by someone who can adapt the terms of your relationship to the new situation. Algorithms adapt slowly and through formal channels. Relationships adapt in conversations.
I saw this clearly when working on the early phases of automated credit decisioning. The cases where automated decisions worked well were cases where the situation was genuinely standard. Clear history, clear purpose, clear repayment capacity. The cases where automation struggled were the cases where reality was more complicated, where the right decision required understanding context that did not fit the input fields.
And reality is usually more complicated. The clean cases are the exception. The real distribution of financial situations that banks deal with every day is heavily populated by cases that require judgment, discretion, and the ability to hold multiple considerations in tension. This is not a temporary condition that better technology will eventually solve. It is a structural feature of the problem domain.
What This Means for the Future of Financial Intelligence
The genuine question for the next decade is not whether AI will replace human judgment in finance. It will replace it in the domains where judgment has always been replaceable: the mechanical application of rules to clean data under stable conditions. These domains are large, and their automation will produce real efficiency gains and create real pressure on the employment structures of financial institutions.
The more interesting question is what happens at the boundary. What happens to the human judgment layer that operates in the complex, ambiguous, relationship-dependent cases that automated systems handle poorly?
My observation, from inside the industry, is that this layer does not disappear as automation expands. It concentrates. As routine decisions are automated, the human experts who remain become responsible for a higher proportion of hard cases. Their judgment is more frequently exercised on situations where the stakes are high and the data is incomplete. The average difficulty of the decisions they face increases.
This has implications for how we should think about developing and retaining financial expertise. The skills that matter most are not the skills that make someone good at processing clean applications efficiently. They are the skills that make someone good at reading incomplete situations, managing relationship complexity, and making consequential calls under uncertainty. These skills are developed through experience, and they are not straightforwardly taught. They accumulate through years of exposure to the full variety of human financial situations, learning what the edge cases look like and developing the judgment to navigate them.
The institutions that will perform best in a world of expanded automation are the ones that most effectively combine automated efficiency on routine cases with genuine human expertise on complex ones. Not the ones that replace human judgment entirely, and not the ones that resist automation in domains where automation is clearly superior. The ones that understand where the boundary actually falls.
The Thing the Balance Sheet Does Not Show
Here is what I want you to carry out of this chapter.
Every financial institution you interact with is, on the surface, a collection of technical systems, regulatory structures, risk frameworks, and balance sheet positions. These things are real and they matter. But underneath them, holding them together, is something that the balance sheet does not show and that no risk model fully captures.
It is the accumulated belief of everyone who has ever chosen to deposit their savings there, take a loan there, send a transfer through there, that this institution will do what it says it will do. That the money will be there when they need it. That the loan terms will be honored. That the counterparty will not suddenly disappear.
This belief is not passive. It is constantly being renewed or eroded by every interaction the institution has with every person it serves. The loan officer who listens carefully and explains the decision clearly. The transfer that processes without errors. The complaint that gets resolved without requiring the client to fight for it. Each of these small events deposits something into an account that no financial statement measures.
And this account, the trust account, is the most important one the institution holds. Without it, the technical systems are just software. The balance sheet positions are just numbers. The correspondent banking relationships are just contracts waiting to be terminated.
I learned this not from a textbook but from watching what happened when that account was under stress. From watching clients make irrational-looking decisions that turned out to be deeply rational once you understood that they were not optimizing for yield. They were optimizing for something harder to quantify and more important: confidence that the institution they were dealing with was genuinely on their side.
Finance presents itself as a discipline about numbers. At its foundation, it is a discipline about belief.
And belief, unlike a balance sheet, cannot be audited. It can only be earned, slowly, through every interaction and every decision, by people who understand that their work is not just processing transactions but building and maintaining the invisible architecture of trust on which everything else rests.
That is a weight worth feeling.
Not as a burden, but as an orientation. The financial professionals who understood this, who moved through their work with some awareness that each interaction was adding to or subtracting from something larger than any individual transaction, consistently made better decisions than those who did not. Not because awareness made them more technically accurate. Because it made them more human, in a domain where being genuinely human turns out to be the most practically valuable thing you can be.
The woman who came to me about the transfer her son had sent. The contractor whose loan I approved on a borderline application. The clients who stayed during the worst weeks of 2020. They were all, in their different ways, extending something that had no formal name in any of the risk frameworks I worked with. They were extending the benefit of the doubt. They were acting on a belief that the institution and the person in front of them would reciprocate.
Most of the time, in functioning financial systems, that belief is rewarded. The transfer arrives. The loan is repaid. The institution stays open. Trust is vindicated, and the cycle continues.
But here is what I want to leave you with: the fact that trust is usually vindicated does not mean it was rational to extend it. It means the institution earned it. Every day, through thousands of small decisions made by thousands of people who understood, at some level, that they were not just processing paperwork. They were tending something fragile and extraordinarily valuable.
We built the most complex financial architecture in the history of civilization. And the thing that makes it work is not the architecture.
It is whether the people inside it are worth trusting.
That has always been the real question. And it always will be.
