How a Colombian Credit Union Cut Delinquency by 38% with Predictive AI

Pulse AI Editorial · ITSense — Edition #03 · The Case of the Week · April 29, 2026

A problem that scales badly

Every Tuesday morning at 9 AM, the risk officer walks in with the same folder. Color-coded printouts of the active loan portfolio. Green: current. Yellow: early delinquency. Red: accounts more than 90 days past due — the ones that, by experience, are already a loss.

The meeting takes two to three hours. The credit committee reviews yellow-flagged accounts one by one. An analyst explains the member's history. The manager asks about savings balances. The risk officer estimates exposure from instinct. A decision is made: continue monitoring, call the member, or initiate collections.

This scene plays out across hundreds of credit unions in Colombia every week. The process works. But it does not scale.

As portfolios grow — more members, higher loan volumes, broader geographic reach — the Tuesday committee becomes a bottleneck. Files pile up. Decisions take longer. And when early delinquency is not caught quickly, it becomes permanent default.

This case documents how a mid-sized solidarity-sector credit union in Colombia broke that cycle. Not with a sophisticated AI platform. With clean historical data and a predictive model that the risk department chose to own.

The sector most AI vendors overlook

Colombia's solidarity sector is larger than most outsiders expect. There are more than 4,000 active cooperatives in the country, according to Confecoop (Colombia's national cooperative federation). Of those, roughly 200 are savings and credit cooperatives with significant lending activity, serving more than 7 million members nationwide.

For a large portion of those members — independent workers, smallholder farmers, micro-entrepreneurs, teachers, retirees — the credit union is their primary source of formal credit.

The challenge: traditional credit scoring was not built for this population.

Mainstream credit risk models rely heavily on data from credit bureaus: bank loan history, credit card behavior, installment payment records. But the typical credit union member has a different profile — variable income, fragmented formal financial history, and a financial life that does not always leave a trace in the conventional credit system.

What they do have is a history inside the credit union itself. How many times they have borrowed. How they have repaid. How much they save. How long they have been members. What economic activity they are engaged in. That data exists. Until recently, it lived scattered across spreadsheets, legacy systems, and incomplete fields in the core banking platform.

That was the starting point of the intervention documented here.

The intervention: three steps, ninety days

The credit union in this case is a composite profile. Its characteristics reflect documented patterns observed across the Colombian solidarity sector over the past three years: a second-tier cooperative with active portfolios across multiple municipalities, a membership base between 15,000 and 40,000 individuals, and a core banking system between 5 and 12 years old.

The project was triggered by a board decision: reduce delinquency or accept a capital adjustment in the second half of the year. The pressure was financial and immediate.

The three-step intervention

Step 1: The data was there. It was just dirty.

The first month of the project had nothing to do with models or algorithms. It was a data audit.

The data team reviewed four years of portfolio history. The finding was blunt: key variables for risk analysis — member economic activity, historical payment frequency, savings-to-debt ratio — were missing in 40% of records. Some due to input errors. Others because the legacy system did not require those fields at origination. Others because the data was never properly migrated when the platform changed in 2019.

Without cleaning this, any model would simply be a mirror of that disorder.

Three weeks were spent normalizing variables, imputing missing data with business rules validated by the risk department, and building a master feature table per member.

Step 2: This was not generative AI. It was classical scoring with proprietary data.

This is worth stating clearly because the market conflates these categories constantly.

This project did not use large language models. It was not a chatbot. It was not generative automation. It was supervised machine learning credit scoring — RandomForest and XGBoost — trained on the credit union's own cleaned portfolio history.

The main predictive variables: member tenure, historical average loan amounts, on-time payment frequency in the previous 24 months, savings-to-active-debt ratio, and declared economic sector.

The model was trained on 70% of the historical data and validated on the remaining 30%. Accuracy in predicting 90-day delinquency was 78%. Not perfect. But significantly better than intuitive judgment applied to 150 files in two hours.

Step 3: The score is an input, not a verdict.

This design choice is the most important factor in understanding why it worked.

The predictive score does not replace the credit committee. It informs it. Every loan application now arrives with a number between 0 and 1 — the estimated probability that the loan will enter delinquency within 90 days.

• Score above 0.7: the committee approves with conditions or declines directly.
• Score below 0.4: expedited approval with minimal review.
• Score between 0.4 and 0.6 (the gray zone): full committee review, the same as before.

The practical result: 65% of loan applications no longer require full committee review. Only gray-zone cases — where the model has lower certainty — receive the same level of attention that previously every application received.

The pull quote that frames it

The result: 38% less delinquency. And something no one anticipated.

Six months after deployment, the results were clear:

• Non-performing loans: 38% reduction compared to the prior semester.
• Loan evaluation time: from 5 days to under 1 day on average.
• Credit committee workload: 65% reduction in full-review cases.
• Loan approvals: volume did not decline. The score was not used to restrict credit — it was used to concentrate attention on the cases that genuinely needed it.

But the most important outcome was not in the portfolio metrics.

The risk department became the owner of the model.

In most AI implementations that fail in the financial sector, the pattern is consistent: the technology team builds the tool, the business team receives it without understanding it, adoption falls to 20% within six months, and the project dies from lack of use.

Here, the opposite happened. From day one, the risk department was part of the project team. The risk officer participated in defining the predictive variables. He challenged the ones that did not seem intuitive from a business perspective. He proposed including a variable for "frequency of branch visits" that the data team had initially discarded as noise. When it was included with adjustments, it improved model accuracy by 3 percentage points.

When the system went live, the risk department did not receive a new tool. It received a formalization of its own institutional judgment — encoded in an algorithm it had helped design.

That is what explains the 38%. Not the algorithm. The ownership.

The lesson that connects to this publication's core thesis

In Pulse AI Edition #01, we reported the statistic that has become the anchor of this publication: 88% of organizations are now using AI in at least one business function. But only 6% capture real, sustained value. The gap is not about access to technology. It is about readiness — in data, governance, and organizational design.

In Edition #02, we explored how that gap widens in the agentic era: more investment, more agents deployed, less trust in outputs. The pattern is the same regardless of the AI paradigm.

This credit union is in the 6%. Not because it had better technology than the other 3,994 cooperatives in Colombia. Not because it hired a ten-person data science team. Not because it had an extraordinary implementation budget.

It is in the 6% because it started with the data. Because it put the business in front — literally: the risk department leading the model, not receiving it. And because it did not expect AI to solve a problem that its own internal process had created.

How to replicate this pattern at your institution

If you manage a loan portfolio at a credit union or community lending institution and this case resonates, here are four concrete moves before engaging any technology vendor:

Step 1 — Audit your historical loan data. Four years of credit, payment, and member history. Measure what percentage of key risk fields are empty or inconsistent. If that number exceeds 20%, the problem is not the model — it is the data. Start there.

Step 2 — Identify the variables you already have. You do not need external data sources. Member tenure, historical loan amounts and frequency, savings behavior, declared economic sector, loan origination channel. That data already exists inside your core banking system. The work is structuring it.

Step 3 — Run a pilot on your 10% highest-risk portfolio. Do not start with the full portfolio. Take early-delinquency loans from the past semester, build the model on that subset, and validate accuracy before scaling. This reduces project risk and generates internal evidence quickly.

Step 4 — Put the risk department in front, not IT. This is the step most frequently skipped and the one that matters most. The risk department must co-own the model from variable definition through score interpretation. If IT leads and risk only "receives," the model will die from disuse within a year.

88% have AI. 6% have results.

This credit union did nothing technologically extraordinary. It used open-source tools available to any data team. It took ninety days to implement and six months to measure meaningful outcomes.

What was extraordinary was the sequence: data first, then model, then integration with the existing process. And at each step, the business in front.

88% of organizations have access to the same technology. The 6% that captures value understands that AI is not a product you install. It is a process you build with the business inside it.

This credit union understood that. The question for yours is when you will start.

Pulse AI by ITSense covers applied AI cases across the industries that drive Latin America. This is a composite case built from documented patterns in Colombia's solidarity-sector cooperative landscape. No data is attributed to a specific ITSense client.

Is your credit union evaluating predictive AI for portfolio management? Let's talk.

Previous editions: - Ed. #01 — The Real State of Enterprise AI in 2026: 88% Adopt, Only 6% Capture Value - Ed. #02 — The Agentic Era Arrived. Enterprises Didn't: What Harvard and Stanford HAI Are Telling Us