AI Collection Agents Meet Strict Liability: What Automating Dunning Actually Costs
Consumer collection law is one of the few regimes where a single sentence is a complete cause of action. It does not require proof of harm, and in large part it does not care whether anyone meant to mislead. Companies are now pointing generative agents at exactly that surface — writing payment messages, negotiating over chat, handling inbound calls — and reviewing them the way they would review a marketing campaign.
Why Collections Is the Worst Place to Improvise
Most consumer-protection regimes ask whether a practice was unfair in context. Collection law goes further in two ways that matter enormously for generative systems. It prohibits false or misleading representations about the debt with essentially no intent requirement, and it is commonly evaluated from the perspective of an unsophisticated or least-sophisticated consumer rather than a careful reader. A message that a compliance lawyer would parse as technically accurate can still fail if a struggling consumer would reasonably read it as a threat of imminent legal action.
Now consider what a language model does well: it adapts tone to the person it is talking to, it fills gaps confidently, and it produces urgency when the prompt asks for persuasion. Every one of those strengths is a liability generator in this context. The agent that improvises "this could affect your credit for years" or "our legal team will review your file this week" has produced a fresh violation, in writing, timestamped, in a system you control and will later be required to produce.
Where Automated Collections Breaks
- •Improvised statements about litigation, garnishment or credit effects
- •Amounts, interest and fees restated inaccurately from context
- •Urgency language that implies consequences the creditor cannot impose
- •Required disclosures dropped when the model rewrites for tone
- •Approving a sample output does not clear the next generation
- •Cease-communication requests recognized late or not at all
- •Disputes treated as objections to overcome rather than legal triggers
- •Requests for validation answered with another payment prompt
- •Attorney-representation statements missed in free text
- •Bankruptcy and deceased-consumer signals need hard stops, not soft prompts
- •Contact counters kept per campaign instead of per consumer and per debt
- •Multiple channels each running their own independent cadence
- •Retries after delivery failures counted as zero contacts
- •Time-of-day rules applied in the wrong time zone
- •Messaging consent and opt-out obligations tracked separately from calls
- •Debt details revealed to whoever answers the phone or reads the screen
- •Voicemail and SMS content visible on shared or family devices
- •Workplace contact rules ignored by an always-on dialer
- •Wrong-number continuation after the recipient says they are not the debtor
- •Rich message previews exposing balance information on a lock screen
The Architecture That Contains the Risk
The workable pattern is not "a better prompt." It is constraining what the model is permitted to emit. Use generation for comprehension — understanding what the consumer said, classifying intent, routing — and use deterministic templates for anything that asserts a fact about the debt or its consequences. The model selects among approved messages; it does not author them. That single boundary eliminates the largest category of exposure while keeping most of the operational benefit.
The second boundary is escalation bias. Every ambiguous signal — anything that might be a dispute, a cease request, a bankruptcy mention, a wrong number, a statement of hardship, an attorney reference — should stop automated contact and route to a human. Tuning that threshold toward more escalation costs conversation volume. Tuning it toward fewer costs statutory damages and, eventually, a pattern-and-practice allegation built from your own transcripts.
The Regimes That Arrive Together
Federal collection rules are the headline, but an automated program rarely triggers them alone. Telephone and text consent statutes govern how the outreach is delivered, with their own per-message exposure. State collection and unfair-practices laws extend to creditors the federal statute does not reach. Credit reporting rules apply the moment the model discusses furnishing or disputes. Recording-consent laws attach to voice agents. Comprehensive state privacy laws add rights over profiling that decides collection treatment. Bot-disclosure rules apply in several states. Six regimes, one deployment, usually one review.
Deployment Checklist
Before Go-Live
- ☐Restrict the model to selecting approved templates for any factual assertion about the debt
- ☐Build deterministic detection for dispute, cease, attorney, bankruptcy and wrong-number intents
- ☐Disclose the automated nature of the agent up front, in the first message
- ☐Centralize contact counting per consumer and per debt across every channel
- ☐Red-team the agent with hardship, hostility, confusion and third-party-answer scenarios
In Ongoing Operation
- ☐Sample transcripts continuously — generated text drifts even without a model change
- ☐Alert on any output containing legal-action, credit-impact or fee language outside a template
- ☐Verify escalated conversations actually stopped automated contact within the same cycle
- ☐Re-test the full scenario suite after every vendor or model version update
- ☐Retain transcripts under a defined schedule and know what production would show
Frequently Asked Questions
We are a creditor collecting our own accounts. Does any of this apply?
More than most creditors expect. The federal statute focuses on third-party collectors, but several states extend comparable duties to first-party creditors, unfair-practices law reaches abusive conduct by anyone, and calling and texting consent rules apply universally. Running two message standards is usually more expensive than building to the stricter one.
Can we let the model write the messages if we review a sample?
Sampling proves what the model did, not what it will do. Because liability attaches per message and the standard is strict, the practical control is architectural: let the model classify and route, and let approved templates carry every statement about the debt, the consequences or the amount.
What happens if the consumer disputes the debt mid-conversation?
Obligations start at that moment. If the agent keeps negotiating a payment plan afterward, the transcript establishes both the trigger and the failure to honor it. Treat dispute and cease language as hard stops detected deterministically, not as sentiment the model weighs against continuing.
Do we need to say the agent is not a person?
Several states require disclosure when an automated system converses with a consumer commercially, and collection rules separately forbid misleading impressions about who is making contact. Disclose in the first message. A friendly human-sounding name with no clarification is an avoidable fact for a complaint.
How do frequency limits break in automated systems?
Almost always through counter fragmentation. Calls, SMS, email and chat each maintain their own cadence, the same consumer appears under multiple account IDs, and delivery retries do not increment anything. Count per consumer and per debt in one place, across all channels, and include retries.
Our platform vendor advertises FDCPA compliance. Is that enough?
It describes what the product can be configured to do. The obligations sit with the entity collecting the debt and the entity that chose the cadence, disclosures and escalation thresholds. Negotiate documentation, change notification and audit rights, and re-run your scenario tests after each update — updates are when reviewed behavior silently changes.
Your Payment Pages Are Part of the Conversation
Every automated message points somewhere — a payment portal, a dispute form, a hardship page. If a consumer cannot use those pages, or if what they say differs from what the agent said, that gap becomes part of the complaint rather than a UX issue.
See what your site currently says. Run a free scan and review every page your collection messaging sends people to.