Contact Center Architecture for Debt Collection

There is no per se safe mode. Predictive and power dialing generally fit consented, first-party, and landline-heavy paper; preview and manual approaches fit litigation-prone segments and stricter states; human-initiated architectures exist for cell phones but rest on contested case law, not settled law. We map modes to your paper mix, states, and risk tolerance with your counsel, then configure the platform to enforce the choice.

It creates a rebuttable presumption, not a safe harbor. More than seven call attempts about a particular debt in seven days, or a call within seven days after a conversation about that debt, is presumed harassment; staying under is presumed compliant, and both presumptions can be overcome by facts. The counting runs per debt and per person, and stricter rules layer on top, including a two-in-seven state rule in Massachusetts.

Federal debt collection rules require each recording to be kept for three years after the call, and records evidencing compliance until three years after the last collection activity on the debt. Roughly a dozen states require every party's consent to record, so multi-state floors disclose at call open everywhere. Retrieval matters as much as retention: exams and attorney demands are answered in days, so export rights, search, and speed are procurement criteria.

New York City's newest debt collection rules cap contact attempts at three per seven days across calls, texts, and email combined, effective September 1, 2026, with mail and consumer-initiated contact carved out. That is an architecture problem as much as a policy problem: dialer, SMS, and email platforms that each count separately cannot prove a combined cap. Unified attempt counting is exactly the kind of vendor-neutral integration we design.

Only inside consent. The FCC ruled in February 2024 that AI-generated voices are artificial voices under the TCPA, so outbound AI calls to cell phones need prior express consent, and collections layers its own frequency and disclosure rules on top. The defensible pattern today is AI on consented reminders, inbound self-service, summaries, and quality review, with humans on everything requiring judgment. Vendors pitching AI voice to raise attempt volume are selling risk.

Real-world negotiated mid-market UCaaS deals typically land at $15 to $27 per seat per month depending on volume, term, and licensing. CCaaS runs higher, typically $50 to $150 and up per agent per month depending on features and AI capabilities, and collections adds structural traps: per-minute pricing that punishes predictive dialing, per-seat pricing that punishes seasonal surges, and compliance modules priced as add-ons at renewal. Consolidating and renegotiating typically recovers 15 to 30 percent. We bid specialist and horizontal platforms across 250+ suppliers and right-size before we negotiate.