By Heorhi Tratsiak
The algorithm reviewing your next loan application has processed millions of files and has a margin of error lower than any human. It has also never had a bad month, buried a parent, or rebuilt from zero at 47. That is not a minor detail. That is the entire problem.
I approved and rejected hundreds of loan applications over seven years inside a major financial institution. Some of the best decisions I ever made were technically wrong by every metric the model used. And some of the cleanest approvals I ever processed defaulted within eighteen months.
What I know, from the inside of that work, is this: AI is coming for credit decisions faster than the industry is admitting, the people most affected will not be warned in advance, and the consequences will be distributed in ways that are deeply unfair and almost impossible to appeal.
The Confession Nobody in Banking Will Make Publicly
Let me tell you about a woman I will call Maria.
She was fifty-three. She had worked the same job for twenty-two years. She had one credit card, used it occasionally, paid it off consistently. She had almost no credit history in the traditional sense because she had lived most of her adult life without needing to borrow.
Her credit score was mediocre. Not bad. Not good. The kind of number that makes a model nervous.
She wanted a small business loan to open a tailoring shop. The documentation was modest. The revenue projections were conservative. The business plan was handwritten in places.
By every quantitative measure I had, this was a borderline case leaning toward no.
I approved it. Because in twenty minutes of conversation, I understood something the score did not contain: this woman had been managing household finances for thirty years with the discipline of a CFO and the accountability of someone who had no safety net. Her lack of credit history was not evidence of risk. It was evidence that she had never needed credit before, which is the opposite of a red flag if you know what you are looking at.
She repaid the loan. She still runs the shop.
An AI model trained on historical data would have declined her. Cleanly. Consistently. Without hesitation. And it would have been statistically defensible, which is the part that should keep you up at night.
What AI Credit Scoring Actually Does Well
I want to be honest here, because the argument is not that AI is wrong about everything. It is not.
AI-based credit decisioning has real advantages, and ignoring them makes the rest of the argument weaker.
It is consistent. A human loan officer on a Tuesday after a bad weekend reviews applications differently than on a Thursday morning. The research on this is extensive and damning. Judges give harsher sentences before lunch. Loan officers approve fewer applications late in the day. AI does not have these fluctuations.
It is faster. Decisions that used to take days now take minutes. For straightforward applications, this is genuinely better for the applicant.
It removes certain biases. There is significant documented evidence that human loan officers discriminate, often unconsciously, based on race, gender, accent, and appearance. AI trained on clean data does not care what you look like. This matters enormously for communities that have historically been on the wrong side of human judgment.
And it scales. One AI system can process ten thousand applications simultaneously with the same quality of review. No human institution can match this, which is why every major lender is moving in this direction regardless of what their press releases say about the value of human relationships.
These are not small things. They are real improvements in a process that was badly flawed in its human form.
But here is what the industry is not telling you about what gets lost in that transition.
What 10 Million Data Points Still Cannot See
Credit models are built on history. They learn from what happened to people who borrowed before you, under conditions that existed before now, in an economy that no longer quite exists.
This is fine when your situation looks like the situations in the training data. The model has seen a million people who look like you on paper, and it knows what happened to them, and it gives you a score accordingly.
The model breaks down when your situation is genuinely new, genuinely unusual, or genuinely human in ways that data does not capture.
A thirty-eight-year-old who left a corporate career to care for a sick parent for three years has a gap in their employment history. The model sees the gap. The model does not see the reason. The model does not know that this person managed a household through a medical crisis, negotiated with insurance companies, handled estate paperwork, and emerged on the other side with a set of real-world competencies that make them more reliable, not less.
A recent immigrant with strong earning history in another country has thin credit history in the United States. The model sees the thin file. The model does not see the earnings trajectory, the professional reputation, the cultural context in which debt avoidance is a point of pride rather than a sign of financial weakness.
A small business owner in the middle of a difficult year who has consistently met obligations for a decade gets a snapshot of a bad quarter. The model sees the snapshot. The model does not see the decade.
These are not edge cases. These are tens of millions of real people who interact with the credit system every year and whose situations require exactly the kind of contextual judgment that AI credit scoring is structurally unable to provide.
The 50 Million People Who Will Fall Through the Gap
The lending industry processes roughly 150 million credit applications per year in the United States alone. Somewhere between twenty and fifty million of those involve applicants whose situations have genuine complexity that falls outside clean scoring categories.
These are disproportionately the people who can least afford to be wrong.
First-generation Americans building credit for the first time. Single parents who took a financial hit during a custody battle. Entrepreneurs whose businesses had a hard year in an economy that was objectively hard for everyone. Gig workers with income patterns that look volatile to an algorithm but are actually more stable than they appear when you understand the work.
When a human loan officer declined these people, they could ask why. They could provide context. They could speak to a manager. They could come back next month with additional documentation and try again. The process was inefficient and sometimes biased, but it was navigable. There was a person on the other side.
When an algorithm declines them, the interaction looks different. You get a letter. Or an automated email. You are told you did not meet the criteria. You are given a reason code from a predetermined list that may or may not accurately describe your situation. You are told you can reapply in thirty days.
There is no one to call. There is no conversation to have. There is no opportunity to explain the three years you spent caring for your mother or the quarter your business had after a major client went bankrupt.
You are, in the eyes of the system, a set of inputs that produced a declining output. Nothing more.
When the Algorithm Denies You, Who Do You Call?
This is the question that the lending industry has not answered and is not eager to answer, because the answer is uncomfortable.
When I made a wrong call as a loan officer, I was accountable. My manager reviewed my decisions. My approval rate and default rate were tracked. If I was consistently wrong in ways that cost the bank money or in ways that discriminated unfairly, there were mechanisms to identify and correct that.
Who is accountable when the model is wrong?
The bank will tell you the model is fair because it was trained on unbiased data. Researchers who study these models will tell you that the data itself encodes decades of biased lending decisions, which means the model learning from that data is, at minimum, learning patterns that reflect historical discrimination even when no one intended to build discrimination into the current system.
The company that built the model will tell you their validation process showed strong performance metrics. Those metrics measure how accurately the model predicts default among people similar to those in the training data. They do not measure what the model does to people who are not well-represented in that data.
The regulator will tell you they are monitoring for fair lending compliance. They are. And they catch some of the most egregious cases. They do not catch the structural problem, which is not that the model is maliciously biased but that the model is incomplete, and incomplete in ways that disadvantage specific populations consistently and invisibly.
The person holding the denial letter has no one to call.
What This Means for You Right Now
If you are about to apply for a mortgage, a business loan, or any significant credit product, here is what the industry will not tell you in the application materials.
Your application will almost certainly be scored algorithmically before a human ever looks at it, if a human looks at it at all. The factors that matter most to the model are your payment history, your utilization rate, the age of your accounts, and your recent inquiry history. Context does not matter to the model. Explanation does not matter to the model. Your story does not matter to the model.
If your file is clean, the system will serve you well. It will be faster and more consistent than the human process it replaced.
If your file has complexity, a gap, an unusual income pattern, a thin history, a hard year, the system will struggle with you. And you will have limited ability to get a fair hearing.
The practical implication is that you need to manage the inputs the model actually uses. Pay everything on time, every time, even small amounts, even when you have the cash to pay it off in full. Keep your utilization low even when you can afford higher balances. Do not close old accounts. Minimize new credit applications in the six months before you need a significant loan. These are the levers you actually control.
What you cannot control is whether the model is asking the right questions about your specific situation. That is the part that nobody in the industry is adequately addressing.
The Deeper Question Everyone Is Avoiding
The replacement of loan officers by AI is a productivity story and an efficiency story and in some ways a fairness story, and all of those framings are real.
But underneath all of them is a question that we are not asking clearly enough.
When we give an algorithm the authority to determine who gets access to capital, which in the United States is one of the primary mechanisms by which people build wealth and stability, we are making a decision about what kind of society we want to build. We are deciding that consistency and efficiency and scalability are more important than the kind of judgment that can hold complexity, context, and individual humanity in the same decision.
That may be the right decision. There are genuine arguments for it, and the human alternative was not clean or fair either.
But we should make that decision explicitly, with full understanding of what we are trading. Not by default, because the technology became available and the economics pointed in one direction and the banks responded to their incentive structures, which is what is actually happening.
The woman with the tailoring shop would have been declined by the model. She would have gone without capital and probably not tried again. The shop would not exist. That is a real loss. Not a tragedy, not a crisis, just a quiet disappearance of something that should have existed but did not because the tool reviewing her application could not see what I could see.
Multiply that by a million. By ten million. By the full scale at which these decisions are now being made.
That is what we are trading for consistency and speed.
I am not sure we have thought clearly enough about whether it is worth it.