If you lead collections, this post is for you. We wrote it for someone evaluating whether putting an AI agent on collections makes sense, what the business needs to get started, what mistakes to avoid and how to measure if it worked. It is not a sales pitch — it is a map.
What follows starts from a simple observation: collections, in most LATAM companies, is stuck in the same place. The first days after the due date are worked well. The +60 portfolio is half-worked. The +90 portfolio is written off or sent to lawyers. And nobody has time to check whether that distribution is optimal.
The problem before talking about the solution
Collections is not just about recovering money. It is about maintaining the customer relationship, reading each situation and applying judgment. When volume rises, judgment thins out: the team ends up sending the same template message to 300 different customers, losing precisely the human element that used to make the difference.
Three symptoms always appear when a collections operation overloads:
The team works by urgency, not by strategy. The newest cases are contacted fast because they're fresh. The old cases get contacted less because "the moment has passed". The result is that the +60 portfolio grows and becomes permanent.
Consistency breaks. Monday at 9, every customer gets a message. Thursday at 16, nobody does. When someone is on vacation, their cases go cold. When a new case comes in, whether it gets worked depends on the collector.
Traceability is weak. Conversations live across several phones, some in the CRM, others in spreadsheets, others only in the collector's memory. When you have to audit, reconstructing the case takes time.
If you recognize two of the three symptoms, automating has a return. If you recognize all three, automating is the most urgent thing on the quarter.
A simple framework to evaluate
Three questions, in this order, define whether your operation is a candidate.
1. How much overdue receivables do you have per month and how is it distributed by age?
If the total overdue is small and operations covers it without stress, automating is optional. If overdue receivables are high and the distribution shows more than half sits in the +30 days bracket, there's concrete potential.
2. What percentage of the team's work is repetitive?
We mean opening message, reminder, promise confirmation, promise follow-up. Tasks where the content barely changes and where human skill adds little. If more than half the team's time goes into this, automation frees up significant capacity.
3. What happens today with +60 days receivables? Are they worked, or written off?
This is the most important question. If the answer is "worked partially when there's time", there's invisible portfolio an agent can recover. If the answer is "sent to lawyers without going through management first", the cost of litigation is much higher than the cost of automated management.
When all three answers point in the same direction — high volume, high repetitiveness, poor +60 work — an AI agent pays for itself in weeks.
What operations looks like with a collections agent
The daily routine changes in three places.
At the start of the day, instead of the team building the work list manually, the agent reads the ERP overnight and prepares the list. Each case comes with age, last contact, registered customer intent and a recommended next action.
During the day, the team works the cases the agent escalates — those requiring judgment, conversation or decision. Repetitive cases run in parallel via the agent. The customer gets the greeting, replies "I'll pay Friday", the agent registers and schedules, without touching the team.
At the end of the day, the team doesn't close the list — the list never closes. The agent keeps working, registers promises that come in late afternoon, escalates whatever crosses the threshold, leaves everything ready for the next morning.
The most palpable difference teams report after running this way is that Monday morning is no longer the most stressful moment of the week. Operations level out the flow and remove the spikes.
Criteria to choose a solution
When evaluating tools to automate collections, it pays to review five criteria. Failing even one is enough reason to drop the option.
Connects to YOUR ERP, doesn't ask you to migrate. Any proposal that requires changing ERP is non-viable. Your ERP is the business truth; an agent that doesn't query it is useless. The three typical integration paths (database, REST API and MCP) are explained in how to connect your ERP with WhatsApp.
Respects country-specific hours and compliance. Without rules for hours, holidays, contact limits and opt-out, the risk of being blocked on WhatsApp Business and reported to the local data protection regulator is high. Country-by-country detail in WhatsApp Business compliance for LATAM companies.
Escalates to humans with context, not blind. If the escalation passes the case without a summary, the human spends time rebuilding it. The agent's value is lost right where it matters most — on complex cases.
Has clear metrics from day one. If the tool doesn't show what happened with each customer, how many replied, how many promised, how many paid, you're operating blind. Measurement should be part of the product, not an annex.
The vendor builds the connector with you, not via a form. Each ERP is different, each operation is different. If the vendor hands you a generic onboarding and leaves you to figure it out, you'll spend three months fighting the integration. You want a vendor that puts people on building your specific connector during the rollout.
Common mistakes when rolling out
Launching to the whole portfolio without segmenting. The first month, take a single aging bracket — say, 30-60 days — and run only that. Validate tone, metrics, escalation. Then expand. Launching to the full portfolio on day one is the fastest way to torch the launch.
Generic template messages. Already mentioned: "dear customer, this is a reminder…" doesn't convert. The agent must use customer name, invoice number, exact amount and, where applicable, acknowledge history.
No compliance rules. Messages at 21:30 are the number-one cause of complaints and spam reports. Every serious operation defines hours and respects them.
No baseline before measuring. If you didn't write down what your portfolio looked like before — average DSO, percentage in +60, reply rate to manual contact — every later number will look like improvement. Serious measurement requires comparison against a verifiable baseline.
Expecting results in the first week. The first two weeks are adjustment. Tone tweaks, time changes, escalation threshold calibration. The operation stabilizes around the fourth week. Asking for results on day seven is unfair to both the tool and the team operating it.
How to measure well
Serious measurement of a collections-with-AI project has three planes.
Operational (weekly). Coverage of the portfolio contacted within the first 7 days of overdue. Reply rate. Human escalation rate. Volume of registered promises. These numbers show whether the engine works.
Financial (monthly). Average days sales outstanding (DSO). Evolution of +30, +60, +90 receivables. Promise keep rate. Effective recovery as a percentage of the portfolio worked. These numbers show whether the engine converts.
Team (monthly). Time spent on repetitive management, before vs. after. Percentage of time spent on complex cases. Self-reported team satisfaction. These numbers show whether the engine is freeing capacity for what matters.
A well-instrumented operation reviews all three planes every month and adjusts. Automation is not "install and forget" — it is operating with one more team member that requires supervision, feedback and ongoing tuning.
When not to automate
There are cases where the right answer is not to automate yet.
When the portfolio is too small. If the recurring volume of overdue invoices is low, the gain from automating is marginal. Better to wait for growth or attack another process.
When ERP data isn't trustworthy. If the invoice base has wrong amounts, invalid dates, duplicate customers, the agent will send wrong messages. Before automating, clean the source. An agent on dirty data is worse than no agent at all.
When collections has no owner. If there's no one in the company accountable for collections, there's no one to review what the agent does. Automation requires human supervision — someone who approves escalations, adjusts thresholds and measures.
When the team doesn't want it. If the collections team sees the agent as a threat rather than a tool, the rollout gets sabotaged. The project has to be built with the team, not against it.
How it fits in Pacunex
The collections agent is one of the platform's specialized agents. It connects to the ERP, operates WhatsApp Business with approved templates, respects per-country hours and compliance, and escalates to humans with a summarized context. The specific integration with your ERP is built by the team during rollout. More detail in AI agents for collections and receivables recovery with AI.
Next steps
If you're evaluating rolling out a collections agent at your company, let's talk on WhatsApp with a founder. We'll walk through what your specific ERP requires, what you'd measure in the first month and how we evaluate whether your operation is a candidate, with no commitment.