AI in debt collection is software that decides which accounts to work, when and through which channel, and checks the work for compliance. The proven uses are propensity-to-pay scoring, contact timing, channel choice, message testing and call monitoring. Every one of them still runs inside Regulation F, the TCPA and federal unfairness rules, and a creditor answers for what its agency’s AI does.
AI changes the order of the work more than the work itself. A program that used to call every account in a balance band in the same sequence can now rank accounts by likelihood to resolve, send a text where a call would go unanswered, and save the phone for the conversations that need a person.
Where AI helps in collections
| Use | What it does | What it needs to be safe |
|---|---|---|
| Propensity-to-pay scoring | Ranks accounts by likelihood to resolve, so effort goes where it pays | Fresh data; scores that update as the account changes |
| Contact timing | Picks the hours a consumer is most likely to answer | Hard limits on contact hours and Regulation F’s call presumption |
| Channel choice | Routes accounts to email, text, portal or phone | Consent records and an opt-out in every electronic message |
| Message testing | Tests wording and offers in digital messages | Pre-approved templates; every variant reviewed for accuracy |
| Self-service and virtual negotiation | Lets consumers set up payments or plans on their own | Clear disclosures; an easy path to a person |
| Call monitoring | Scores calls for compliance and quality, and flags risk | People review what it flags; it doesn’t decide outcomes on its own |
| Hardship and vulnerability flags | Finds signs of hardship, disputes or bankruptcy in notes and calls | A routed path to trained staff and a pause on collection |
What compliance requires of AI
AI doesn’t get its own rulebook in collections. It inherits every rule a human collector follows, and a few questions regulators and plaintiffs’ lawyers now ask about models in particular.
Contact limits count across channels
Regulation F presumes a violation when a collector calls more than 7 times in 7 days about a debt (12 CFR 1006.14(b)). An AI that picks the “best time” for each channel separately can blow through that limit without anyone noticing. The count has to run per debt, across every dialer and channel. Here’s what counts toward 7-in-7.
Electronic messages need consent and an exit
Every email and text needs a clear way to opt out (12 CFR 1006.6(e)), and texts to cell phones need the consent the TCPA requires. The FCC’s consent revocation order changes how fast opt-outs must be honored; the TCPA page covers the dates.
AI voices are “artificial” voices
In a February 2024 declaratory ruling (FCC 24-17), the FCC held that calls using AI-generated voices are calls with an “artificial or prerecorded voice” under the TCPA. Outbound AI voice calls to cell phones need the same consent as any prerecorded call.
Decisions have to be explainable and logged
Under the federal ban on unfair, deceptive or abusive practices (UDAAP), a creditor answers for what its service providers do. If an AI decided to contact a consumer at a certain time with a certain message, someone has to be able to say why, on a specific account and date, long after the fact.
Models should be tested for unfair outcomes
A model trained on past payment behavior can treat groups of consumers differently in ways nobody intended. Ask for evidence the vendor tested for that, and what changed when it found something.
The full map of which collection laws apply to creditors and which to agencies is on the debt collection laws page.
8 questions to ask any AI collections vendor
- How are contact attempts counted per debt across calls, texts and emails?
- Is every message checked against federal and state rules before it goes out, or after?
- How often are account scores refreshed after placement?
- Can you produce the decision trail for one account on one date?
- How was the model tested for unfair outcomes, and when was it last tested?
- Which conversations always go to a person: disputes, hardship, bankruptcy, attorney representation?
- Who reviews what the call monitoring flags, and who do they report to?
- What result did AI change, against what baseline, over what period?
The last question matters most. Results vary widely by portfolio, and a lift measured on one creditor’s accounts may not hold on yours.
How TSI uses AI in collections
CollectX, TSI’s analytics platform, scores each account on arrival, so collectors and digital outreach start with the accounts most likely to resolve. Channel and timing follow the contact strategy agreed with each client, inside the rules above. Ripple, TSI’s speech analytics tool, scores every call each day for compliance and quality and flags calls with risk markers, such as a dispute or an attorney mention. A quality coordinator who reports to the Chief Compliance Officer listens to and scores each flagged call; Ripple doesn’t take account actions or decide outcomes. How calls are recorded, scored and root-caused is on the call monitoring page.
Disputes, hardship and anything a model can’t explain go to trained collectors. That’s where the judgment is, and where most complaints start.
On the healthcare side, PULSE, TSI’s denial and appeals platform, has measured an 86% denial overturn rate, 43% higher payment per appeal and 50% less time to appeal on TSI’s own denial work.
Results vary by portfolio, payer mix, placement stage and client program.
Get a Recovery Performance Diagnostic to see your placements, liquidation and complaints side by side with what your agencies report.
Frequently asked questions
How is AI used in debt collection?
Mostly to decide which accounts to work, when and through which channel: propensity-to-pay scoring, contact timing, channel choice, message testing, and monitoring calls for compliance and quality. People still handle disputes, hardship and anything a model flags.
Are AI debt collection bots legal?
They can be, inside the same rules as any collector. The FCC ruled in 2024 that AI-generated voices count as artificial voices under the TCPA, so AI voice calls to cell phones need the consent the TCPA requires. Bots must also follow Regulation F’s contact limits, disclosures and opt-out rules.
Can AI replace debt collectors?
It replaces some tasks, not the accountability. AI handles scoring, routine digital messages and self-service payments well. Trained collectors still handle disputes, hardship, payment negotiations that need judgment, and every case a model can’t explain.
What should creditors ask an AI collections vendor?
How contact attempts are counted per debt across every channel, how each message is checked against federal and state rules before it goes out, how decisions are logged and retrieved, how the model was tested for unfair outcomes, and who reviews the exceptions.
What is digital debt collection?
Collecting through email, text, web portals and self-service payment pages instead of only letters and calls. Regulation F allows it with conditions, including a clear opt-out in every electronic message.