Superintelligent Agent for Collections

A superintelligent agent for collections is, in practice, an AI agent that keeps getting better at negotiating debts. That capability only helps when it comes with written limits: a discount range, rules of tone and clear triggers for handing the conversation to a person. What protects the debtor and the company is not how smart the model is. It is what the model is allowed to do, within what Article 42 of Brazil's Consumer Defense Code (CDC) permits.

Key takeaways

• The more capable the agent, the more its limits should move out of the prompt text and into the system, where the model cannot work around them.
• The discount policy should be a table queried by a tool that rejects values outside the range, not an instruction the agent interprets.
• Collection tone is a code of conduct: identify yourself, never expose the debt to third parties, never threaten, and accept a 'not now'.
• Escalation to a human needs objective triggers, such as a disputed debt, vulnerability, a request to stop contact and a discount outside the table.
• A pilot succeeds when agreements are honored and complaints are avoided, not when agreements are merely closed.

What does a superintelligent agent for collections actually mean?

The term shows up in searches, but it helps to be precise. Superintelligence, in the strict sense, is a hypothesis: a system that outperforms humans at practically every cognitive task. What a lender, a school, an internet provider or a clinic operates today is an AI agent made of a language model, tools and rules. It improves every time the model is swapped for a better one, which is why the topic calls for more care, not less. To separate hypothesis from operation, read the difference between AGI, ASI and AI agents.

Collections is a good test because it mixes money, emotion and law. A weak agent fails through lack of understanding. A strong agent fails through conviction: it negotiates well, sounds persuasive and may push a deal the customer cannot afford, or promise a discount nobody authorized. If the only instruction is to recover the amount owed, a competent model will find efficient ways to do it, and some of those ways are exactly the ones the law forbids.

What do consumer law and data protection law require of automated collection?

The starting point is Article 42 of the CDC: when collecting debts, a customer in default must not be exposed to ridicule or subjected to any kind of embarrassment or threat. The CDC also treats collection by threat, coercion or embarrassment as an infraction. The agent has no bad intent, but it follows the goal it is given. That is why the legal rule has to become an operating rule.

One practical translation for the agent's daily work:

• Never mention the debt to third parties: relatives, colleagues, group chats or the employer.
• Never threaten consequences the company cannot or does not intend to carry out, such as protest, credit bureau listing or a lawsuit.
• Never use false urgency, such as a deadline that does not exist or a discount that 'ends today' when it does not.
• Identify who is collecting, the creditor, the origin of the debt and the amount, always from the source system.
• Respect a request to stop contact or to switch channels.
• Never charge an amount the source system does not confirm.

On the data protection side (LGPD, Brazil's general data protection law), purpose and necessity apply: the agent only needs what is useful for negotiating that specific debt. The debtor's health, family or employment data is rarely necessary and should not enter the agent's context. If the agent influences decisions about credit or payment terms, the right to review under Article 20 becomes relevant. We wrote about it in LGPD automated decision-making. The national data protection authority (ANPD) is the official reference for guidance on the subject. None of this replaces legal advice: your legal team should review the final policy.

How do you set discount limits the agent cannot exceed?

The central rule is that the limit lives in the system. If the discount ceiling exists only in the prompt, it depends on the model obeying in every conversation, including those where the customer insists, argues or claims that 'the manager already approved it'. If the ceiling lives in a tool that validates the proposal, insistence changes nothing.

A decision guide for building the policy:

• Store the policy in a table with ranges by debt type, days overdue and amount, with an owner and a review date.
• Make the proposal tool query that table and reject any condition outside it, with a clear error message back to the agent.
• Define the order of offers: first a new due date, then installments without discount, then a discount up to the ceiling of the first range. Later ranges only with a recorded reason.
• Ask about ability to pay in a simple way, without requesting sensitive data, and never offer an installment the customer said they cannot afford.
• Above the ceiling, the agent does not negotiate: it escalates to a person with authority.
• Record in every agreement the original amount, the offer, the acceptance, the date and the policy version applied.

That last item matters for audits and for the right to review. If someone disputes an agreement, the company has to show what was offered, why, and under which rule.

What tone should you use for collections on WhatsApp?

WhatsApp is a personal channel, and the customer's phone may be visible to other people. The first message should not detail the debt before confirming that you are speaking with the account holder. For outbound contact, the channel requires an approved template and prior permission. The subject is covered in WhatsApp outbound messages, templates and opt-in. Collection without a clear basis for contact is a legal risk and also a risk to the reputation of your number.

Tone rules that fit into instructions and evaluations:

• Open with identification and purpose, in one short sentence.
• Use neutral, non-judgmental language, for example: 'We found an outstanding balance and want to find a way that works for you'.
• Offer choices, not ultimatums.
• Acknowledge difficulty without faking intimacy or dramatizing.
• Accept 'not now' and record the best time for another contact, within a time window defined with your legal team.
• Never compare the customer with other debtors or suggest bad faith.

A good test: could a supervisor read the message aloud next to the customer without causing embarrassment? If not, it is off tone.

When should the agent hand the conversation to a person?

Escalation is not a failure of the agent. It is part of the design. The more capable the model, the more tempting it is to let it handle everything, and the more important it is to have objective triggers that do not depend on its judgment:

• The customer does not recognize the debt, or claims fraud, an incorrect charge or a payment already made.
• The conversation mentions illness, bereavement, unemployment, violence or another vulnerable situation.
• The customer asks to speak to a person, asks to stop contact, or mentions Procon (the consumer protection agency), Reclame Aqui (a public complaint site) or a lawsuit.
• The request for a discount or deadline falls outside the table.
• The person replying appears to be a third party, or the data differs between systems.
• There are repeated refusals combined with growing irritation.

At handoff, the agent delivers a summary: who the customer is, what was said, what was offered and why it escalated. The customer should not have to repeat the story. When the issue is a public complaint, the guide on handling consumer complaint sites with AI shows how to keep a record and human review in the loop.

Illustrative scenario: a negotiation that ends with a human

Illustrative scenario: a recurring-services company sends a message, using an approved template, to a customer with overdue monthly fees. The agent confirms it is speaking with the account holder, names the creditor and the amount as shown in the system, and offers a new due date. The customer says they lost their job and cannot pay the full amount. The agent proposes installments without discount, and the customer asks for a bigger discount, claiming a previous agent had promised one.

The proposal tool rejects the value outside the table. The agent does not argue or invent an exception. It explains that this request needs review, records the mention of unemployment as a vulnerability situation and transfers the conversation to a person with the full summary. That person checks the history, confirms whether an earlier promise was made and decides within their own authority. No threat was made, no third party was involved and the log shows every step.

How do you pilot the agent and show it is safe?

Our way of working follows four verbs: map, limit, pilot and hand over. In collections, that translates as follows:

• Map: document the current flow, who negotiates today, the discount ranges in use, the most common reasons for delay, and the starting point for complaints and honored agreements.
• Limit: define which data the agent sees, which tools it can call, which hours and channels it operates in, and which actions require human approval.
• Test before talking to customers: build cases in a test environment, using mock tools and a sandbox for AI agents. Include the aggressive debtor, the one who claims fraud, the one who tries to induce a discount by citing an authorization that does not exist, and the one who asks to stop contact.
• Pilot a real flow: choose an early-stage overdue portfolio, with human review of a sample of conversations and logs of every decision and every tool call.
• Agree on the metric: honored agreements, complaints and correct escalations weigh more than agreements merely closed, because an unpayable agreement becomes new default.
• Hand over and decide: at the end of the pilot, your team takes over the operation and decides whether to expand, adjust or stop.

For structuring risk management, the NIST AI Risk Management Framework is a useful reference, and Anthropic's guide to building agents reinforces a preference for simple, predictable flows where errors are costly. The same logic applies to evaluations: agent evals help turn tone and limit rules into repeatable tests that run every time the model or the policy changes.

Frequently asked questions

Can an AI agent collect debts on its own? It can lead the conversation within defined limits, but it should not be the last instance. Disputed debts, vulnerability and discounts outside the table need a person. Autonomy should grow only as logs and evaluations show the limits hold.

Can I let the agent decide the discount? That is not recommended. The agent can choose among offers that are already authorized, but the ceiling and the ranges should sit in a table validated by a tool. That way the discount decision stays with the company, and does not depend on how a model performs in one specific conversation.

Does the customer need to know they are talking to AI? LGPD requires transparency about data processing, and stating that the service is automated is the safest practice and the most consistent with it. Always offer a path to a person. Your legal team should validate the exact wording of the notice.

Does a more capable model justify more autonomy? Capability and autonomy are separate decisions. A better model tends to follow rules better, but it is also more persuasive when it gets something wrong. Increase autonomy only after testing the new model against the same hard cases and reviewing the pilot logs.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents