Superintelligence for Debt Collection Firms: Settlements and Audit Logs

Superintelligence for Debt Collection Firms: Settlements and Audit Logs

See which settlement offers an AI agent can build in debt collection, which amounts need human approval, and how to log every contact within legal limits.

•

Want to apply this in your company?

Book a call with Juan →

In a debt collection firm, the AI should build settlement offers only within ranges the company has approved in advance, with fixed rules for discount, term and payment method. Anything outside that range, including exceptions, disputes and requests for different terms, passes to a person with the authority to decide. Every contact must be recorded with the exact message, the rule used, the offer generated and who approved it.

Key takeaways

  • The AI assembles settlement options from a rule table approved by management, not from the model's own judgment.
  • Amounts, discounts and terms outside the approved range, along with any dispute, require human approval within a defined authority limit.
  • Data used in negotiations must follow the purpose and necessity principles of the LGPD, with restricted access.
  • Messages must respect what the Consumer Protection Code prohibits in collection and the rules of each channel, such as WhatsApp.
  • Every contact creates an auditable record: channel, text, rule, offer, decision and the reason for any handoff.

Superintelligence is a label, the real work is an agent with rules

Many people arrive at this topic through the word superintelligence, but for a collection firm that word does not describe any product you can buy and configure. What exists today is a language model inside an agent, with instructions, access to a few tools and explicit limits. If you want to understand the gap between what is sold under that name and what actually operates, read ChatGPT is superintelligence? What it is and what it is not before any conversation with a vendor.

In practice, the useful question is not how intelligent the AI is. It is what the AI can do, with which data, and who answers for the result. This way of thinking supports a safe collection operation because it shifts the focus from the promise to the design of the process.

What settlement offer the AI can build

The AI can do three things safely: read the balance and the case history, present the options the company has already approved, and collect the debtor's response. For example, it can offer a lump-sum payment with a discount up to the ceiling the company has set, or an installment plan with the number of payments listed in the table, with a due date chosen by the customer within an allowed window.

It should not calculate new conditions on its own, promise terms that are not in the table, or suggest that an offer is a final chance or valid only today unless that is a real rule. When the model invents phrases like these, it creates legal and reputational problems.

A practical way to structure this is to separate the rule table from the message text. The table states what is permitted. The text only communicates what the table says. If the conversation requires information that is not in the table, the agent hands off instead of filling the gap with assumptions.

Which amounts require human approval

The most important rule is simple: the further an offer moves from the approved standard, the closer it must be to a person. A decision guide can look like this:

  • An offer within the approved range with no dispute: the agent can present the condition and record the acceptance.
  • A discount request above the ceiling or a term beyond the limit: the agent routes it to an operator with authority, with a short summary of the history.
  • The debtor disputes the amount, the origin of the debt or whether the debt exists: the agent pauses the negotiation and hands off without arguing.
  • Mention of a lawsuit, a lawyer, a protest or a credit listing already in progress: immediate transfer to legal or to the responsible person.
  • Signs of vulnerability, a declared financial hardship or a request not to be contacted: treat as a special case and escalate without pushing an offer.
  • A request to speak with a person: transfer without trying to persuade the customer another way.

The authority limits should be defined by the company's finance and legal teams, not by the technology team. Implementers translate that decision into a rule, a test and a log. You also need to decide what happens when the operator does not respond within the agreed time. Usually, the correct response is to tell the debtor the actual next step, without promising a condition that has not been approved.

Legal and contact limits the agent must respect

In Brazil, debt collection has two areas the agent must respect from the first message. The first is data protection. The LGPD requires that processing have a clear purpose, that only what is necessary be used, and that access be restricted to those who need it. In practice, the agent should not receive the entire customer base, only the fields for the case in progress. If the company uses a model provider that processes data outside the country, the international data transfer also needs review with legal counsel.

The second area is collection practices. The Consumer Protection Code addresses abuses such as threats, exposure of the debtor and inappropriate contact. Which hours, frequency and language are acceptable should be defined by the company's legal team, based on its portfolio and contracts. The agent should treat these rules as technical constraints: if the hour is not allowed, the message simply does not go out.

WhatsApp adds another layer. Messages initiated by the company depend on the customer's permission and on approved templates. The customer service window also limits what can be sent freely after a customer reply. These topics are covered in WhatsApp outbound messages: templates, opt-in and best practices and in the WhatsApp customer service window: how an AI agent should act.

When the agent makes or suggests decisions that affect the debtor, the question of human review also comes up. It is worth understanding what the LGPD says about automated decisions before designing the flow, especially when the offer depends on data about the customer's profile.

How to document every contact

Documentation does not mean storing the whole conversation. It means recording enough to reconstruct what happened, why it happened and who decided. A minimum record for each contact should include:

  • The case identifier and the channel used.
  • The exact text sent to the debtor, with date and time.
  • The categories of data consulted, without copying the entire base into the record.
  • The version of the rule table applied at that moment.
  • The offer generated, with amount, discount and terms.
  • The automated decision and, when applicable, who approved it and when.
  • The reason for any transfer to a human.
  • The debtor's response and the result recorded by the operator.

Records need access controls and a retention period set by the company. It is also wise to keep sensitive data out of free text, such as health information the debtor mentioned, and to treat those fragments as restricted fields. Records also help you review the operation itself. If many offers are transferred for the same reason, the rule table probably needs adjustment, not more instructions for the agent.

When thinking about logs, traces and metrics for an agent in production, the observability concepts from OpenTelemetry help organize the idea, even if your company does not use that tool. For governance, the NIST AI risk management framework organizes mapping, measurement and monitoring. In Brazil, the ANPD is the reference point for questions about data protection.

Illustrative scenario: a settlement request beyond the table

Illustrative scenario: a debtor replies on WhatsApp asking to split the debt into more installments than the approved table allows, and asks whether a bigger discount is possible because they paid other invoices on time. The agent recalculates nothing. It tells the customer it will check with the team, records the request with both conditions the customer asked for, transfers it to an operator with authority, and states a return time the team can actually meet. The operator approves a special condition, and the record shows who approved it and under which rule. If the company has no authority for this case, the operator replies with the standard condition and explains the next step.

None of this requires the AI to be brilliant. It requires limits, records and a person with the authority to decide.

How to apply the method in a pilot

The safest path starts by mapping the current process: who negotiates, with which tables, through which channels and with which exceptions. Next, limit what the AI can see and do. Only then should you pilot a real flow, with a defined period and volume, logs, and a success metric agreed before launch, such as the share of appropriate handoffs or the time to the first human response. At the end, the company decides whether to expand, adjust or stop.

This is the method we follow in agent projects that include limits and human handoff. No result is promised. What gets defined at the start is what will be observed and who decides afterward.

Frequently asked questions

Can the AI approve a discount on its own? No. It can present conditions the company has already approved and record the debtor's acceptance. Any condition outside the table needs human approval, recorded with the approver's profile, the date and the rule applied.

Which data should the agent access? Only what is needed to negotiate that specific case, such as minimal identification, updated balance, due date, settlement history and the authorized channel. Data that does not change the offer should stay out of the agent's context. This minimization principle comes from the LGPD and should be reviewed by the company's legal team.

Do I need permission to collect through WhatsApp? The answer depends on who starts the conversation and the type of message. Outbound messages from the company follow their own rules for permission and approved templates, while replies within the customer service window follow different logic. Check with legal counsel before designing the flow.

Is this the same as superintelligence? No. What operates in collection is a restricted AI agent, with rules, transfers and records. The term superintelligence describes a theoretical category and market rhetoric, not what your company will contract for or configure.

Sources

debt collectionAI agentsLGPD complianceaudit logshuman handoff