The moment every credit team dreads
It arrives without warning. A letter from the CMF, the SBS, the BCRA, the BCB, the CNBV, or the Superintendencia Financiera. A request for documentation on a specific model. An explanation of why a particular segment is being approved at a lower rate than others. A justification for the cutoff applied to a specific product.
The risk team opens the request and begins the process of producing a response. They pull the model documentation — if it exists in complete form, which it often doesn't. They reconstruct the reasoning behind feature selection decisions made six months ago by a team member who has since left. They produce a narrative explanation of model behavior that they hope will satisfy the regulator's questions without raising new ones. The process takes weeks. And at the end of it, the institution submits documentation that is accurate in the narrow sense — it describes what the model does — but that fails to answer the deeper question the regulator is actually asking: does this institution understand its own model well enough to be trusted with the decisions it makes?
What regulators are actually asking
The regulatory demand for explainability is often framed as a documentation requirement. But behind the formal requirements is a substantive concern that the documentation is meant to address. Regulators across LATAM are asking a question that has three parts.
First: Does this model make decisions for the right reasons? A model that produces accurate predictions but does so by relying on proxies for protected characteristics is not acceptable regardless of its predictive accuracy. The regulator needs to know that the model's decision logic is legitimate, not just that its outputs are accurate.
Second: Can this institution control its model? A model that cannot be explained is a model that cannot be reliably controlled. If the institution cannot articulate why the model makes specific decisions, it cannot predict how the model will behave under conditions it hasn't encountered before. The regulator needs confidence that the institution is genuinely in control of its automated decision-making.
Third: Can affected customers be treated fairly? Across Chile, Peru, Argentina, Brazil, Mexico, and Colombia, regulations increasingly require that customers who are declined credit receive an explanation of the reasons for that decision — in plain language, not statistical notation. A model that cannot be explained in plain language cannot satisfy this requirement.
Why most explainability efforts fall short
The standard approach to model explainability follows a predictable pattern. The model is built. Validation is completed. Then, in the final stage before deployment, someone produces the explainability documentation — feature importance rankings, SHAP values, partial dependence plots.
This approach has three structural limitations. It is retrospective — it captures what the model does, not why it was designed to do it. It is static — it describes behavior on training data, not on the live population. And it is technical — it produces outputs only a data scientist can interpret, satisfying the formal requirement but failing the practical one.
What genuine explainability requires
Building AI that can truly be explained to a regulator requires integrating explainability into the modeling process from the beginning, not appending it at the end.
Explainability by design. Every modeling decision — feature selection, algorithm choice, hyperparameter configuration, validation approach — should be documented as it is made, with the reasoning captured in language that a non-technical stakeholder can understand.
Plain language explanations at the decision level. For every credit decision the model makes, the system should be able to generate an explanation in plain language that identifies the primary factors driving the decision. Not "feature X has a SHAP value of 0.23" but "this application was declined primarily because the applicant's digital payment activity over the past 90 days shows a pattern inconsistent with the income level reported."
Continuous monitoring with explainability. Model behavior changes over time. In Argentina, where macroeconomic volatility can shift the distribution of key features rapidly, continuous explainability monitoring is the mechanism by which an institution detects that its model is beginning to make decisions for different reasons than it was designed to — before the regulator notices it in the portfolio data.
Regulator-ready reporting as a default output. The documentation regulators request should be generated automatically as a byproduct of the normal modeling workflow, not produced on demand through a manual reconstruction process.
The conversation that changes everything
There is a specific kind of regulatory interaction that separates institutions with genuine explainability capability from those without it. The regulator asks a question that wasn't anticipated: "Your approval rate for self-employed applicants in this region has declined 15% over the past two quarters. What is driving that change?"
An institution without genuine explainability capability begins a manual investigation. Three weeks later, a response is submitted that answers the question — partially, with qualifications about data limitations.
An institution with genuine explainability capability answers the question in the meeting. The model's decision logic is live and queryable. The factors driving the change are visible in real time — a shift in the distribution of income volatility signals, driven by a change in how a specific data source reports self-employment income, which has been flagged by the monitoring system and is already under review. That demonstration is worth more in regulatory relationship terms than any amount of documentation produced after the fact.
Building for the regulatory environment that is coming
The regulatory environment for AI in financial services across Chile, Peru, Argentina, Brazil, Mexico, and Colombia is evolving rapidly — and consistently in one direction. The bar for explainability is rising, and it will continue to rise. Institutions that are building genuine explainability capability now are building for the regulatory environment that will exist in three to five years — not just the one that exists today.
Building AI you can explain to a regulator is not a documentation exercise. It is a discipline — one that, practiced consistently, produces better models, faster deployment, stronger regulatory relationships, and more defensible credit operations. It is also, increasingly, the price of admission to the markets where the greatest growth opportunities in Latin American financial services exist.