For years, automating collections meant automating email. Templates, reminders and sequences that fire 5, 15 and 30 days after the due date. It works well until the debtor stops opening the email. And then someone on your team has to pick up the phone.
In our most recent webinar, we showed, live and without a safety net, what happens when it is an AI voice agent who picks up. Two calls to the same debtor, two agents with opposite tones and two different outcomes. And afterwards, what almost no one shows: the configuration screen behind it.
The question is who decides who gets called, when, through which channel, in what tone, and what happens to what the debtor says in response.
Here is what was covered in the session, with the full video so you can listen to it yourself.
From spreadsheet to orchestrator: the three models that coexist today
In the meetings we hold every week with credit and collections teams in Chile, Peru, Colombia, Mexico and Spain, we almost always come across one of these three realities.
1. Manual collections
The manager sends emails one by one or calls directly and logs payment commitments in a shared spreadsheet. Fewer and fewer operations work this way, for a simple reason: traceability depends on someone remembering to write it down.
2. Static automated flow
Actions are triggered according to the stage each invoice is in (pre-due, due, post-due) and according to fixed rules defined by segment. This is a huge step up from the previous model, but it remains a rigid decision tree: always the same channel, always the same message, always the same day. And if the client replies to the email, the reply dies in one manager’s inbox.
3. AI-orchestrated collections
An orchestrator is layered onto that flow to decide when, how and what to claim, invoice by invoice. It assigns tasks to each manager’s calendar, chooses the most effective channel for each client type (email, WhatsApp, SMS or a voice call) adjusts the tone and, above all, reads and interprets the debtor’s response.
This is the difference between automating tasks and delegating operational judgement. The orchestrator handles the volume; the manager handles the decisions that require human judgement.
Listen to the calls: the same debtor, two different agents
In the demo, Macarena Donoso, who is responsible for Chile, Peru, Colombia and Mexico at Atlax 360, played both the collections manager and the debtor: she configured the agents and then answered the phone herself.
The first call was deliberately set up to go wrong. The task concerned an invoice more than 90 days overdue, close to being declared a bad-debt claim, and it should have been handled by the escalation agent. It was mistakenly assigned to Marcela Montes, the initial-claim agent. The result was a flawless, friendly call and entirely wrong for that stage of the debt. That live mistake explains, better than any slide could, why tone is not a cosmetic detail: it is part of the collections policy.
Next came the second call, from Ramón Oviedo, the legal-escalation agent. You can listen to it here: video.
| Agent | Assigned role | Tone | Point in the cycle | How the call ended |
| Marcela Montes | Initial invoice claim | Warm and collaborative | Early days after the due date | The debtor commits to checking the payment with Finance and agrees to a follow-up call |
| Ramón Oviedo | Escalation prior to legal action | Firm and direct | More than 90 days overdue | Rejects instalment terms, sets a final deadline and closes with a specific day and time for follow-up |
| Marcela Montes | Incident verification | Warm and collaborative | Overdue invoice with payment blocked | Identifies a commercial dispute and refers it to the relevant team |
The third call is the most interesting one for any collections manager: the debtor had not paid because she had received eight boxes instead of ten, and damaged ones at that. The agent did not press for payment. It acknowledged the problem, logged the case and escalated it. When the orchestrator detects a commercial dispute, it automatically notifies the salesperson assigned to that account, because the issue is no longer one of collections.
The 7 parameters that define a collections voice agent
A voice agent is not “hired“, it is configured. These are the seven parameters that determine its behaviour within Yndika’s collection management module.
- Role and purpose of the call. Initial claim, legal escalation, dispute verification, or an administrative task, such as requesting the purchase order without which no invoice can be issued.
- Tone and personality. From warm to firm. Defined in writing and tested before it is switched on.
- Language, voice and accent. Configurable by segment. This has a real impact: in some markets we have found that a particular accent gets far more responses than the local one.
- Agent disclosure. Whether the agent must identify itself as a virtual agent at the start of the call. In countries such as Colombia, regulations require the call to be identified as debt-collection management, and this rule is built into the flow.
- Negotiation scope. The maximum acceptable term, whether instalments are permitted, and what interest or collection costs apply under your arrears policy.
- Originating number. Calls are placed from a number belonging to your own company, not from a call-centre prefix that no one picks up any more.
- Escalation and referral. Which channel is triggered if there is no answer (WhatsApp, SMS or email) and which manager or salesperson the case is referred to.
With these seven fields, a team of three collectors can operate with five or ten agents assigned to their department, each playing a different role. Capacity grows without growing headcount.
Behind the tone lies the data: how the orchestrator decides
A voice agent with no data is simply a robot dialling in chronological order. What turns these calls into a recovery strategy are the risk indicators that Atlax 360 models for each client:
- Maximum recommended exposure per debtor.
- Probability of default within the next 90 days.
- Financial health, with daily alerts on any change.
- Payment behaviour, ranging from good to defaulted, based on days overdue.
- Collectability indicator per invoice, not just per debtor.
- Collection forecast and expected loss.
These indicators draw on three sources: the client’s own ERP history, between 12 months and 3 years of positive, negative and late-payment data; our own proprietary market database, built over a decade, covering more than 800,000 unique tax IDs (RUT) in Latin America and more than two million debtors in Europe; and the public or central risk-bureau sources available in each country. Machine learning recalculates them daily, so credit assessment stops being a snapshot and becomes a living picture.
On this basis, risk monitoring is defined and the portfolio is segmented. In the demo, four segments were used: large companies, repeat defaulters, the construction sector and debt overdue by more than 90 days, and for each one a decision is made on what percentage of the process is automated. Repeat defaulters with small invoices: 100% automated. Other segments receive only internal recommendations, with no agent intervention. You decide where AI comes in and where it does not.
What happens after the call ends
This is where collection management with agents sets itself apart from simple automated dialling. Every call leaves behind:
- The recording and the full transcript of the conversation.
- A conclusion on the outcome of the contact and the intent detected: payment commitment, commercial dispute, liquidity problem or administrative request.
- A new task in the relevant manager’s calendar, with a date and time, if the debtor has asked to be called back.
- A recommendation on the next action to take.
This repository is not simply an operational convenience. It is what you will be asked for when you need to write off a bad debt, when you notify a claim to your credit insurer, or when an audit asks how you assessed the risk and managed the collection. If you extend credit to your clients without being a bank or a lender, you need to be able to demonstrate the traceability of those decisions.
And it has a side effect that is rarely mentioned: when a manager goes on holiday and their portfolio passes to someone else, the person taking it over can read and listen to everything that was done before. The memory of the collections process no longer lives solely in one person’s head.
Questions from the audience
Can the accent and language of the calls be chosen?
Yes. Language, voice and accent are configurable by agent and by portfolio segment.
What happens if the debtor does not answer the call?
An automatic escalation to another channel is triggered, usually WhatsApp or email, depending on what proves most effective for that type of client. The sequence continues to escalate depending on the response received.
Are the tasks predetermined, or can we create our own?
There are three possible sources: those created by the orchestrator, those arising from a system recommendation, and those created manually by the manager. You can create and automate tasks by channel and by source.
Is the configuration set by the client or by Atlax 360?
The initial setup (business rules, payment and arrears policy, segments, users and permissions) is defined together with you during implementation. From that point, your team can create segments, assign tasks and amend the configuration, with our support included for any sensitive changes.
Does it connect with SAP or other ERP systems?
Yes. We have connectors and extractors already built that work on standard tables in read-only mode, with a range of connection methods, so that it does not become a development project for your IT team. We have delivered more than 2,000 implemented projects.
How long does implementation take?
The standard timeline, from the ERP connection through to releasing the modules, is around 8 weeks. For projects with technical particularities, we review potential bottlenecks in advance.
Do existing clients need to change platform?
No. The voice agents are integrated as an additional layer on top of the existing collection management module.
The conclusion from the webinar
What impressed attendees most was not that the AI spoke well. It was seeing that the same debtor, with the same invoice, receives two entirely different conversations depending on the risk of the transaction and the stage the debt is at. That is something an email flow simply cannot do.
Collections are not automated to save on calls. They are automated so that no small invoice goes unmanaged, so that no client response takes two weeks to be answered, and so that your team can spend its day on the decisions that genuinely move cash flow.
Everything else can now make the call by itself.
