Most collections teams prioritise by age and amount because those are the figures already to hand. Prioritising by probability of recovery changes the order of the work, and order is the only thing that can be optimised when the team is finite.
What follows: why age orders badly, what can be modelled, what still requires judgement, and which obligations come with this data.
An overdue portfolio is not a uniform block. It contains cases that will resolve themselves with a reminder, cases that need a negotiation, and cases where effort will not change the outcome.
Treating them alike spreads the same effort across very different situations, and that is what makes the work feel intense and return little.
Why age orders badly
Age measures how long an obligation has been overdue, not the willingness or the capacity to pay of whoever holds it. They are different things and they do not move together.
A customer twenty days late with an impeccable history is usually an administrative oversight. Another at the same twenty days with three prior restructurings is a different situation, and age does not tell them apart.
Amount adds the same bias by another route: it concentrates effort on large obligations even when their probability of recovery is low, while leaving unworked a volume of small cases that a single contact would have resolved.
What can be modelled
Probability of recovery. Behavioural history — prior punctuality, number of defaults, response to earlier contacts — supports an estimate of how likely each case is to be recovered within a given window.
The channel with the best response. Customers do not respond alike to a message, a call or an email. Learning which channel works by segment avoids spending the most expensive interaction on someone who would have answered the cheapest.
The moment of contact. The hour and day a customer responded before is information that is usually recorded and almost never used to schedule the next attempt.
The early signal. The most valuable case is not the overdue one but the one about to fall due. Changes in payment pattern often anticipate default with enough margin to act before it happens.
What requires judgement
Deciding what to do with each segment remains the business’s call. A model can say a case has a low probability of recovery; deciding whether that means negotiate, provision or stop working it is a policy, not a prediction.
The limit on intensity also has to be set. Analytics may indicate that contacting more often raises recovery, and there is still a threshold above which the practice stops being acceptable, both by regulation and by the relationship with the customer.
And the model’s bias is worth reviewing. If it learns from historical work that concentrated effort on certain segments, it will tend to reproduce that concentration and to confirm that the rest do not respond — simply because they were never worked.
Which obligations come with this data
Portfolio information is personal data of a financial nature, and its processing is conditioned. In Colombia, Habeas Data (Ley 1581) governs processing and reporting to credit bureaus; in Mexico, the LFPDPPP imposes its own conditions of purpose and consent.
That has a design consequence: the purpose for which the data was collected limits what it may later be used for, and a model that widens that use without revisiting the legal basis builds a risk that surfaces later.
Being able to explain the prioritisation matters too. A criterion that orders the work and cannot be justified in response to a complaint is hard to sustain, even when it works.
How to sequence the project
Start by measuring current practice: how many contacts per case, with what result, through which channel and at what time. Without that baseline there is nothing to compare against and any improvement is an impression.
Then build the simplest segmentation that beats age. A sophisticated model is not required to improve on an order based on a single variable.
And compare properly: keep a group worked under the previous criterion for a period. It is the only way to know whether the improvement came from the model or from the team paying closer attention because a project was under way.
What has to exist first
Behavioural history with enough depth, a reliable record of the contacts made and their result, and a written policy for working each segment.
The second is usually the one missing. When the result of a contact is not recorded in a structured way there is nothing to learn from, and that is the project’s first real deliverable.
Is a complex model needed to start?
No. A segmentation based on prior behaviour usually beats ordering by age, and it builds the case for investing in something more elaborate.
How do you know the model improved recovery?
By keeping a group worked under the previous criterion for a period. Without a comparison, the improvement may simply be that the team was paying closer attention.
Which piece of data is most often missing?
The structured result of each contact. Without it there is nothing to learn from, and recording it properly is usually the project’s first deliverable.
What limits does regulation impose?
Processing financial data is conditioned by Habeas Data (Ley 1581) in Colombia and by the LFPDPPP in Mexico, including the purpose for which it was collected. That basis is worth reviewing before widening the use of the data.
