Volume Arrives Faster Than Capacity
Brazil closed July with close to 84 million delinquent consumers, the highest figure ever recorded by Serasa Experian, in a month when household indebtedness, corporate delinquency and judicial recovery filings all hit records or their worst readings on their respective series, with the Selic at 14% a year. The share of households carrying some form of debt reached 81.6% in May, the highest in the CNC survey history.
Two details inside those numbers matter more than the headline. The first is recurrence: 42% of Brazilians delinquent in 2026 were already negativados a decade ago, roughly 34 million people, and the average debt per consumer rose 12.2% in inflation-adjusted terms over that period, from roughly US$1,150 to roughly US$1,300. The second is the corporate side, where more than 9.1 million CNPJs carried about US$46 billion in accumulated debt in June, 16.67% above June 2025 and a record since the series began in 2016.
Mexico is moving in the same direction with different mechanics. Non-performing loans across the banking system reached about US$11.4 billion in June, a real increase of 14.16% year over year and the highest level in at least 26 years according to CNBV data. Consumer credit delinquency hit 3.35% in May, its highest since 2021, per Banco de México. The adjusted delinquency index for credit cards, which includes loans written off over the previous twelve months, reached 13.7% in June, the highest reading since December 2021. Among Sofipos the delinquency index climbed to 9.41%, and Moody’s has flagged that the consumer write-off rate rose almost 100 basis points in three years to around 8% in May.
For banks, retailers, Sofoms and the BPOs that run their recovery operations, the operational consequence is specific. Portfolios are growing in number of accounts, not only in outstanding value, and the cost of a human contact attempt stays roughly flat regardless of the balance behind it. Small and medium balances are where the arithmetic breaks first, and they are also where recurrence concentrates. Local-currency figures in this edition are converted at September 2026 rates, close to R$5.09 and MX$16.92 to the dollar.
References: Terra Brasil Notícias, September 2026 / Serasa Experian, March 2026 / CNN Brasil, August 2026 / El CEO, August 2026 / Mundo Ejecutivo, August 2026 / El Universal, August 2026
AI in Finance: Recovery Operations Under Two Kinds of Pressure
The industry-wide picture explains why technology budgets in collections are moving now. TransUnion’s 2026 debt collection industry report, based on a survey of more than 200 collection professionals, found that 64% of collection companies saw account volume rise over the previous twelve months while only 39% reported improved liquidity. Nearly half named agent productivity and margin improvement as the main motivation for technology spending.
The performance figures circulating come mostly from global deployments and should be read as a ceiling rather than a regional benchmark. McKinsey’s research on generative AI in credit customer assistance and collections, widely cited across the industry, points to operating expense reductions of up to 40% and recovery improvements of around 10%, with digital-first collections strategies associated with non-performing loan reductions in the 20% to 25% range at leading institutions. A randomized field experiment on algorithmic calling decisions reported repayment rates roughly 23% higher than when human agents made the same decisions, which is a narrower claim than it first appears: what improved was the choice of whom to call and when, not the persuasion inside the call.
Latin America has a channel advantage here that does not require customer education. WhatsApp reaches roughly 80% of consumers in the region and 77% already use AI tools regularly, so a payment negotiation over messaging is an ordinary interaction rather than a new behavior. Regional supply is forming around that reality. Belvo launched an AI Collections product this year combining voice and WhatsApp agents, integrating through API or file upload, and positioned explicitly for high volumes of delinquent accounts, for small and medium balances where the cost of human servicing makes individual recovery unviable, and for operations currently running on call centers, BPOs or manual WhatsApp. Its outreach logic includes timing proposals to when income actually arrives and avoiding weeks loaded with fixed expenses, which is a reasonable proxy for what conversation design in this category is starting to look like.
The second kind of pressure is conduct. In January 2026 alone, Condusef registered 2,638 complaints against collection agencies in Mexico over how they conduct their work, 19% above the same month of 2025. The second most frequent cause was collection contact, negotiation or restructuring carried out with someone who is not the debtor, with 876 cases, up 23% year over year. Collection complaints accounted for 17% of the 15,292 complaints filed against banks that month, close to two in every ten. Under the REDECO framework, Condusef sanctions the financial institution that contracted the agency, not only the agency.
Brazil applies pressure from a different direction. The Consumer Defense Code prohibits exposing a debtor to ridicule or subjecting them to undue pressure during collection, and Article 20 of the LGPD gives data subjects the right to request review of decisions taken solely by automated processing. The ANPD included artificial intelligence and emerging technologies in its priority agenda for the 2026-2027 biennium, with regulation of Article 20 on the list and a regulatory sandbox in its testing phase during the year. The general AI framework, PL 2338/2023, remains in the Câmara after being pushed past its original 2025 timeline, and any final text returns to the Senate if substantially modified.
Put those two pressures together and the requirement for a recovery operation becomes concrete. Contact has to scale without headcount scaling with it, and every contact has to be reconstructable: who was called, whether identity was verified, what was offered, what the debtor agreed to, and what the system was instructed to never say.
References: Kaplan Group, 2026 / McKinsey / Galileo / Belvo, 2026 / La Jornada, March 2026 / Confidata
Coru Product: ADI Platform as the Operating Layer for Recovery
ADI Platform is an enterprise platform for running AI conversational agents across voice (PSTN inbound and outbound), WhatsApp Business and an embedded web widget, with the agent logic held in one place and the channel abstracted away from it. In collections that consolidation carries weight, because a portfolio is worked through outbound calls, answered on inbound returns, and continued over messaging, usually across several days.
Agents that are configured, versioned and constrained
An agent is a defined object: system prompt, LLM model, voice configuration, HTTP tools, dynamic variables, guardrails and webhook target. Every element is versioned. An operation can change how an agent handles a hardship claim or a discount request, keep the prior version, and know which configuration was live on a given date. Guardrails are enforced by the platform rather than requested in the prompt, which is the difference between a policy and a suggestion when the instruction is to never contact anyone other than the account holder.
Integration without touching the core
HTTP tools let an agent call existing systems inside the conversation: balance lookup, eligibility for a payment plan, generation of a payment instrument, registration of a promise to pay, ticket creation. An MCP gateway extends the same pattern to external context servers. The collections engine, the core banking system and the CRM stay where they are.
Persistent context across sessions and channels
Conversations retain full context. Prior interactions, commitments made, sentiment signals and pending next steps persist. A debtor who negotiated an installment on a Tuesday call and writes on WhatsApp on Friday is continuing one negotiation, not opening a second one with an agent that has no memory of the first.
Human authority where money is at stake
Which systems an agent can reach, which data it can extract and store, and which actions require human sign-off are explicit configuration. An agent can verify identity, present eligible options and capture a commitment, while authorization of a discount beyond policy routes to a human supervisor or to a decision engine holding that mandate.
Evidence produced by running, not by exporting
Each conversation yields a full transcript, an AI-generated summary, defined extraction fields and HMAC-signed webhook events delivered to the client endpoint. When a debtor files a complaint about how they were contacted, the record already exists in the shape a regulator or an internal audit needs, per account and per attempt.
Governance and multi-portfolio structure
Granular role-based access control, audit logs on every configuration change and server-side tenant isolation support both multi-country operations and BPOs running several portfolios from a single dashboard, with an Organization holding many Clients, each with its own agents, plan and consumption. Operational dashboards cover AgentOps, FinOps and SecOps, with metrics per client, per agent and per organization, including consumption against cycle limits, which is what makes cost per recovered account measurable rather than estimated.
Deployment that fits enterprise reality
Shared multi-tenant SaaS, dedicated cloud per client, partial on-premise and hybrid voice are all supported, with separate dev, staging, sandbox and production environments. Multicloud by design, without dependency on a single provider. Onboarding for a new client runs one to three weeks, and voice can be live in two to five days when the client already operates a SIP trunk. Voice latency targets under 800 ms to first byte, close enough to human cadence that a negotiation holds together.
Coru Weekly Picks
Film Recommendation: Sorry to Bother You (2018)
Boots Riley’s satire opens on a telemarketing floor where a young operator discovers that changing how he sounds changes what he can close. The film has more on its mind than call centers, but its first act is an unusually sharp portrait of what an outbound operation does to the people staffing it and to the people receiving the calls: scripts, quotas, the pressure to keep someone on the line. Worth two hours for anyone designing conversational systems whose job is to ask people for money.
Book Recommendation: Scarcity: Why Having Too Little Means So Much, Sendhil Mullainathan and Eldar Shafir
Mullainathan and Shafir argue that scarcity itself captures attention and degrades decision-making, which is why people in financial distress often behave in ways that look irrational from the creditor side of the table. The relevance to recovery design is direct. Timing, channel, wording and the number of options presented all change outcomes independently of ability to pay, and the book explains the mechanism behind effects that collections teams usually discover empirically.
Looking Ahead
Collections is the highest-volume regulated conversation in Latin American financial services, and the volume is still climbing in both of the region’s largest markets while the capacity to work it manually is not. That gap is where the next two years of operational investment will land.
The institutions that convert it into recovered value will be the ones treating the platform layer as the asset. Persistent context so a negotiation survives a channel change. Guardrails the operation cannot accidentally disable during a month-end push. Approval gates on anything carrying financial authority. Evidence accumulating as a byproduct of execution rather than as a reporting project. In a category where the regulator sanctions the institution for the conduct of its contractor, the record is part of the product.
By the time supervisory attention in the region turns from how agencies behave to how automated agents behave, the operations with two years of documented, auditable contact history will be answering questions rather than assembling answers.
At Coru, we build the operational intelligence that sits between an AI model and a regulated process. In Latin America, that means conversational systems fluent in local channels and local rules, with compliance-first architecture underneath. Every interaction becomes a resolved case with a record behind it.