How AI Predicts Which Debtors Will Pay
AI captures intent, capacity, commitment quality, and engagement signals on every call. Resources route to the 80/15/5 bands that match reality.
TL;DR
In a typical collections portfolio, 80% of recoveries come from 20% of cases. Identifying which 20% matters more than working harder across the whole book. AI voice agents capture structured signals on every call - intent language, payment capacity markers, emotional engagement, follow-up commitment quality - and feed them into a propensity-to-pay model that updates after every interaction. Resources flow to the likely-payer segment, legal escalation flows to confirmed non-payers, and hardship pathways flow to identified-vulnerable customers. This is resource allocation done with data instead of gut feel.
The 80/15/5 Reality of Collections Portfolios
- 80% of debtors are routine. They will pay with light nudging. Dedicating a human to every call destroys margin.
- 15% need genuine negotiation. Payment plans, restructuring, forbearance.
- 5% need legal escalation or insolvency routing.
The entire point of intelligent collections is routing each case into the right path as early as possible. Traditional operations route by portfolio age and balance. Those are weak signals compared to what the actual call reveals.
What the AI Listens For
Intent Language
"I want to pay" vs "I'll try to pay" vs "I'll see what I can do" - these are very different commitment levels. The AI classifies intent strength on a scale rather than a binary.
Capacity Markers
Income mentions, employment status, other debt mentions, household composition. These feed an affordability position that is much richer than a credit-bureau-only view.
Commitment Quality
"I can pay GBP 150 on Friday" is specific. "I'll send something soon" is not. Specificity in commitment correlates strongly with follow-through.
Emotional Engagement
Debtors who engage emotionally with the situation - acknowledge the debt, express concern, respond to empathy - are more likely to pay than disengaged or hostile debtors. This is not magic, it is observed pattern in collections psychology.
Follow-up Behaviour
Did the customer pick up the second call? Did they respond to the SMS? Did they make the first promised payment? Each signal updates the propensity score.
Stat block: AI propensity modelling
- 80/15/5: Typical allocation across routine, negotiation, and legal.
- 20-40%: Lift in right-action rate when allocation uses call-derived signals vs portfolio age only.
- 100%: Share of calls contributing data to the model.
How the Allocation Actually Works
After every AI voice call, the propensity score updates. Customers in the highest propensity bands receive lighter-touch follow-up (SMS reminders, self-service payment links). Customers in the middle band receive structured negotiation calls. Customers in the lowest band and with dispute or hostility signals route to legal review earlier, saving months of wasted outbound attempts.
Crucially, vulnerability signals override propensity. A high-propensity debtor showing vulnerability cues routes to forbearance even if the payment signal is strong. See vulnerability detection for the rule.
Why Traditional Scoring Falls Short
Traditional collections scoring uses credit bureau data, loan performance history, and portfolio age. These capture a slice of the debtor's situation but miss the most informative signal: what the debtor actually said on the last call. AI voice agents fix this because the call is the data source.
Court Timing Implications
Propensity scoring feeds directly into court escalation timing. Low-propensity, dispute-free debtors should escalate to legal before they disappear from the reachable population. See court timing optimisation.
Bottom Line
Working every case with equal effort is a broken allocation model. Propensity-to-pay modelling from call-level signals routes each debtor into the path that matches their actual situation. Human specialists work cases that genuinely need negotiation. Legal resources escalate only on real legal cases. Routine payers self-serve. See related: emotional consistency, ROI framework, why 85% go unanswered.
Call Sarah on +1 (332) 241-0221 or book a consultation.
Frequently Asked Questions
Is propensity scoring a high-risk AI system under the AI Act?
It depends on use. Scoring that influences creditworthiness decisions can be high-risk. Scoring that drives resource allocation inside a collections operation typically is not. See our AI Act guide.
Does the model improve over time?
Yes. Each call outcome feeds back into the model. Portfolio-specific patterns emerge within 4-8 weeks of sufficient volume.
How does this interact with GDPR Article 22?
The scoring informs human decisions about escalation. It does not make decisions with significant legal effect autonomously. Human intervention rights apply to any significant decision.
Can we use our existing credit-bureau score alongside the AI model?
Yes. The call-derived signals are additive to your existing scoring stack, not a replacement.
What happens when the model disagrees with an experienced collections manager?
The manager overrides. The model surfaces signals; humans retain judgement. Over time, calibration between model and experienced managers converges.