Automating WhatsApp collections isn't about blasting templates. It's about building an AI agent that talks with each customer, respects the rules and escalates to a human only when needed. This guide walks through the flow step by step for a LATAM company that already handles WhatsApp collections manually and wants to take the next step.
The difference between manual and AI-driven WhatsApp collections isn't speed. It's consistency. A human team contacts the first 50 customers of the day well and the last 50 worse. An agent works the same way with customer number 1 as with case 800. And collections, above anything else, is a consistency game.
The scenario
To set the stage: we're talking about a company with 800 to 1,500 overdue invoices on a recurring monthly basis, a standard ERP (Microsip, Defontana, Aspel, Bind, Tango, Alegra or something custom), two people dedicated to collections, and WhatsApp already in use but operated manually. They likely share a phone or each collector works from their own mobile. Templates are pasted from a spreadsheet.
That setup works up to a point. When volume rises, consistency drops. When a collector goes on vacation, their cases go cold. When an invoice hits +60 days, no one works it anymore. And auditing is painful because conversations are spread across six phones.
Automating changes all three things: consistency is guaranteed, cases don't depend on individuals, and everything ends up in a single auditable log.
What you'll need
Before building anything, check that the four basic inputs are in place.
- Read access to the ERP database or its API. You need to extract pending invoices, customer data, recorded payments and assigned salesperson. Read-only permissions are enough — the agent doesn't modify the ERP, it only queries it.
- WhatsApp Business API account through a BSP (Business Solution Provider) authorized by Meta. The regular WhatsApp Business app doesn't support professional volume or approved HSM templates. The difference between both options is documented in the official WhatsApp Business Platform guide.
- Approved HSM templates for messages outside the 24-hour window since the customer's last interaction. Without approved templates you cannot start a conversation. There's a dedicated post on what they are and how to write them in official WhatsApp Business templates.
- Basic written rules: business hours per country, local holidays, minimum amount to manage, maximum contacts per customer per month. Without rules, even the best-configured agent ends up annoying customers.
With those four inputs in place, you can start.
Step 1 — Map the receivables portfolio by age and priority
The agent reads the full receivables portfolio daily and sorts it by aging brackets. The classic segmentation works well: 1-7 days overdue, 8-30, 31-60, 61-90, +90. Each bracket gets a different message pattern, frequency and escalation level.
The reason to segment is operational. A customer 5 days overdue probably only needs a friendly reminder; one at 75 days requires a more serious conversation and possibly the team's involvement. Treating them the same wastes the opportunity in the first case and underestimates the problem in the second.
It also helps to segment by amount and by relationship. A long-standing customer with a small delay shouldn't get the same message as a new customer who has hit collections three times already. The agent can read the full history and adjust tone — but only if the information lives in the ERP.
The output of this step is a daily work list. Each customer with their invoice, amount, age, assigned salesperson, last contact. That list feeds the next steps.
Step 2 — Write messages that sound human
The most common mistake in automated collections is using a generic template. "Dear customer, this is a reminder that you have a pending invoice." That message doesn't work because the customer knows it's robotic and ignores it.
Better: the agent builds each message with the customer's name (not "dear customer"), the specific invoice number, the exact amount and the due date. If the customer has a history of paying on time, that gets acknowledged ("We know this is unusual for you"). If there's a preferred payment channel, it gets offered.
The Meta-approved template is only the shell. The real writing happens inside it: the agent fills variables with real data and, when relevant, adjusts tone. A well-designed HSM template leaves room for personalization without losing its approval.
Three concrete rules:
- Customer name first, not "dear customer".
- Specific data in the middle — invoice number, amount, date. This turns the template into a personal message.
- A concrete question at the end — "Should we coordinate the payment today, or would you prefer to schedule it for this week?" — to invite a reply.
Step 3 — Set up basic compliance
Before sending a single message, the compliance rules need to be in place. These are the five that any serious operation in LATAM applies:
- Business hours by country. In most jurisdictions, collections happen between 8:00 and 20:00 local time. Some laws are stricter. The agent respects the customer's local time, not the company's.
- Local holidays. No messages on national holidays. The calendar is loaded per country and refreshed once a year.
- No messages on Sundays, unless the customer replied. It is the industry standard.
- Maximum 3 contacts per customer per month, unless the customer replies and the conversation stays open. This prevents the feeling of harassment.
- Automatic opt-out. If the customer writes "stop contacting me" or anything equivalent, the agent registers it, stops contacting and notifies the team.
The country-specific legal frameworks (Ley 1581 in Colombia, Ley 19.628 in Chile, LOPDP in Ecuador, LFPDPPP in Mexico, Ley 25.326 in Argentina, Ley 29733 in Peru) are covered in more detail in WhatsApp Business compliance for LATAM companies.
Step 4 — Process replies with context
This is where the AI agent diverges from a chatbot. Customer replies don't fall into five clean categories — they're free text, in colloquial language, sometimes audio, sometimes with misunderstandings.
The agent should recognize at least seven typical intents:
- Confirmation of a payment already made. The customer says they already paid. The agent verifies against the ERP on the next sync. If found, it acknowledges and closes. If not found, it asks for the transfer reference or receipt.
- Payment promise with a date. "I'll pay on Friday." The agent logs the promise, schedules a reminder for Friday afternoon, and doesn't follow up until then.
- Request for an extension or refinancing. The agent recognizes the case, offers predefined options if policy allows, or escalates to collections for complex cases.
- Debt dispute. The agent never argues. It registers the dispute and escalates to the team with the history.
- Question about invoice details. The agent queries the ERP and replies with line items, amount, issue date.
- Request to speak to a human. The agent hands off the case and doesn't push.
- Total silence. The agent waits the configured period and applies the monthly contact cap.
Every reply is logged in the CRM with timestamp, content, detected intent and action taken. That is what later supports auditing and improvement.
Step 5 — Escalate to humans when it matters
There are three case families that always escalate, no debate.
Disputes. Any case where the customer denies the debt, questions the amount or claims an undelivered product. The agent registers it, thanks them for the information and passes it to the team. The company decides how to respond.
Amount above threshold. If the invoice exceeds a company-defined threshold, the agent doesn't negotiate. Large cases require human judgment.
Customer asks for a human. If the customer explicitly asks to speak with someone on the team, the agent hands the case off immediately.
The quality of the escalation depends on the summary. The human receives: customer name, specific invoice, age, recent contact history, intent detected in the last message, agent recommendation on how to proceed. Four lines. The human enters the case with context in thirty seconds instead of five minutes.
What to do if something goes wrong
Customer doesn't respond after 3 contacts. Escalate to the team or mark as stalled, per policy. No automated nagging beyond that point.
Template rejected by Meta. Have at least two pre-approved variants per message type. If one fails mid-campaign, the agent pivots without halting the operation. How to avoid this scenario lives in how to avoid blocks in WhatsApp campaigns.
ERP returns wrong data. The agent never denies a debt to the customer or makes up data. If it detects inconsistency (invoice with no amount, customer with no name, invalid date), it pauses the case and notifies the team.
Customer sends audio. Three paths: transcribe and process as text, always escalate audios to a human, or use a standard reply asking for text. The best option depends on the customer profile.
Suspected wrong customer. If the customer says "I think you have the wrong number", the agent pauses and verifies. It doesn't keep pushing against a number that may belong to someone else.
How to measure if it's working
Automated collections KPIs split into two groups.
Qualitative (what you feel):
- Portfolio coverage — what percentage of the total was contacted in the first 7 days.
- Reply rate — what percentage of contacts generates a customer reply.
- Team time spent on repetitive work, before vs. after.
Quantitative (what you measure in money):
- Promises registered vs. promises kept.
- Average days sales outstanding (DSO), before vs. after rollout.
- +60 days portfolio, month-over-month evolution.
- Human escalation rate (how many cases end up with the team). A well-tuned operation escalates 15-20% of cases; the agent resolves the rest.
It pays to set a baseline before rolling out. If you don't measure the starting point, any later number looks like improvement. Serious measurement requires at least four weeks of operation before comparing.
The next level
Once WhatsApp collections is automated and stable, the natural move is to add adjacent capabilities. Automatic reconciliation against payments flowing into the ERP. Aging reports in natural language. Sales briefings for the commercial team on customers with open receivables. Each new layer reuses the data and the connector you've already built.
The full map of what can be automated in a LATAM enterprise lives in what AI agents your LATAM company should have, and the broader view of recovering receivables with AI in recovering overdue receivables with AI.
Next steps
If you want to see how a collections agent works at a company similar to yours, we'll show you on WhatsApp. We'll also walk through what your specific ERP requires and what you'd measure in the first month.