Key takeaways
- AI auto loan collections pilots almost always succeed — they run on early-stage accounts where volume is high, conversations are structured, and borrowers are still reachable
- At 60 to 90 days past due, the borrower profile changes entirely: broken promises, hardship cases, disputes, and borrowers who have already spoken to multiple agents
- Research across 22 million cases found that borrowers break promises made to AI more than to humans on accounts requiring active negotiation — not an argument against AI, but the analytical case for stage-gated deployment
- The right deployment boundary is the Notice of Intent to Repossess — AI handles the pre-NOI volume, human collectors handle the late-stage negotiations that require judgment and authority
- A subprime auto lender using Prodigal started on 33% of their pre-charge-off portfolio and expanded to 100% after validating results: month 1 at +6%, month 2 at +27% in the 30-day bucket, month 3 at +8% overall
AI auto loan collections pilots almost always look good.
Early-stage accounts are the best possible testing ground because the volume is high, accounts are still curable, conversations are structured, and borrowers are generally willing to engage.
But the situation at 60 to 90 days past due is a different question entirely.
Why Pilots succeed and deployments struggle at scale
Early-stage AI auto loan collections works because the work is defined. A borrower 15 to 30 days past due is typically reachable, often surprised by the delinquency, and willing to discuss a resolution. The conversation has a predictable shape, so AI handles this well.
At 60 to 90 days past due, the borrower profile changes. Many accounts have already been contacted multiple times, some have made promises that did not hold, some are in genuine hardship, and some have disputes. The conversation no longer has a predictable shape.
A generic AI auto loan collections platform deployed uniformly across all delinquency stages will perform strongly in the early bucket and struggle in the later ones, because the deployment was not designed to reflect how the work changes by stage.
Why pilots succeed and scale deployments struggle
What changes when AI auto loan collections moves beyond the pilot stage
Account status
Still curable. Borrower is reachable and often surprised by the delinquency.
Conversation shape
Predictable. Cure math, payment arrangement, commitment captured.
Borrower profile
First or second contact. Generally willing to engage.
AI performance
Strong. The work is defined, repeatable, and scalable.
Account status
Multiple prior contacts. Some promises already broken. Hardship cases present.
Conversation shape
Unpredictable. Disputes, complex hardship, legal risk assessment needed.
Borrower profile
Fatigued, skeptical, or in genuine financial crisis. Higher stakes of mishandling.
AI performance
Degrades without stage-gated deployment. Human judgment required here.
The solution is not a better AI platform. It is a deployment structure that matches AI to the stage where it performs best and preserves human capacity for the work that requires it.
What the research says
A 2022 study across 22 million debt collection cases found that borrowers break promises made to AI agents more than promises made to human collectors on accounts requiring active negotiation.
This is frequently misread as a reason to avoid AI in auto loan collections. It is not. It is the basis for deploying it correctly.
Borrowers at 15 to 45 days past due are not in active negotiation, they are resolving a curable delinquency through a structured conversation. AI handles this at scale with consistency and compliance that human agents cannot match at volume.
Borrowers at 60 to 90 days past due, past the Notice of Intent to Repossess, are in a different situation. The account requires human judgment to read the borrower's circumstances, exercising authority on concessions, assessing whether a commitment will hold. This is where human collectors add irreplaceable value.
The research does not say AI fails in collections. It says AI and human collectors are not interchangeable across all stages, so the deployment structure should reflect that.
The right deployment structure
Stage-gated AI implementation in auto finance means deploying AI where it performs best and preserving human capacity for the work that requires it.
Stage-gated deployment
AI handles the volume. Humans handle the judgment.
15–30 DPDAI
30–45 DPDAI
45–60 DPDAI + triage
NOIBoundary
60–90 DPDHuman
90+ DPDHuman
Outbound contact and inbound responses across the full early-stage book
Accurate cure calculations on every call from the live account
Payment arrangement capture and commitment logging
Extension routing and hardship triage
After-hours coverage capturing payments outside business hours
Complex negotiations requiring authority and concession decisions
Hardship assessments where structured workout plans are needed
Accounts with disputes or legal complexity
Situations where reading emotional signals determines the outcome
Repossession decisions and late-stage escalation
This is not a limitation. It is how a well-designed deployment maximizes the contribution of both.
proAgent handles outbound and inbound contact across the full early-stage book, cure calculations, payment arrangements, commitment capture, and hardship triage. It detects payment intent signals, adapts to each borrower's situation, and routes complex cases to human collectors when the conversation requires judgment that goes beyond a structured resolution.
proCollect orchestrates the digital outreach strategy, optimizing which channel reaches each borrower, when to contact them, and what message matches their profile, so proAgent is working the accounts most likely to resolve.
What good AI implementation in auto finance looks like
The champion-challenger rollout is the right structure for AI implementation in auto finance. It starts on a defined portion of the portfolio, typically 30 to 50%, runs both approaches simultaneously, and validates results before expanding.
A subprime auto lender working with Prodigal started on 33% of their pre-charge-off portfolio. The remaining 67% ran on the legacy process. Results were measured month over month.
Phased rollout
How to expand AI auto loan collections without guessing
Start on a defined portion of the portfolio, measure results at each stage, then expand based on evidence.
+6%
Lift in payments collected through stronger text and email engagement
67% still on legacy process
+27%
Lift in the 30-day past due bucket as AI models refined on live data
Comparing champion vs challenger
+8%
Overall lift in payments. Results consistent enough to expand to 100%.
Expanded to 100% pre-charge-off
Starting on a portion of the portfolio removes the risk of full deployment before the model is validated. Results at each stage justify expansion. The lender controls the pace.
The lender subsequently expanded to 100% of pre-charge-off accounts and began preparing to apply the same approach to post-charge-off recovery.
This is what AI implementation in auto finance done correctly produces:
- Measurable improvement at each stage
- Validated before expansion
- With the deployment structure matched to how the portfolio behaves by delinquency stage.
Five questions before you commit to a deployment
- Has the vendor deployed AI auto loan collections in production at scale, not in a pilot, on a portfolio with similar subprime composition to yours?
- How does the platform perform on 60 to 90 day accounts specifically? Ask for data, not a demonstration.
- What is the integration approach for your loan origination system and servicing platform, and who owns it after go-live?
- What does the phased rollout structure look like, and what milestones are contractually defined before expansion?
- How does the platform handle escalation to human collectors, and how is that routing logic configured for your operation?
A vendor who cannot answer these with specificity has not deployed at the scale or complexity your portfolio requires.
Frequently asked questions
Frequently asked questions
Deploying AI in auto loan collections: common questions
What does an RFP for AI in auto loan collections need to cover that a standard software RFP misses?
A standard software RFP evaluates features, integrations, and pricing. An AI collections RFP needs to go further on three dimensions. First, production evidence — ask for live deployment data on portfolios with similar subprime composition, not pilot results or demo performance. Second, domain specificity — ask how the model handles subprime auto-specific scenarios like partial payment disputes and cure calculation, not generic collections conversations. Third, model governance — ask how the model is monitored for performance drift, how errors are flagged, and who is accountable when the AI produces a wrong output. Prodigal's deployment model includes a dedicated implementation team that stays with the account through ongoing optimization, with performance reviewed continuously rather than at contract renewal.
How much does my operations team need to be involved to get the AI agent configured correctly?
Your operations team's involvement is primarily in the first few weeks — sharing how your current workflows are structured, what your compliance requirements are, and what your collections strategy looks like by delinquency stage. This is how the AI agent gets customized to your specific operation, not just deployed on a generic template. After initial configuration, ongoing involvement shifts to reviewing performance dashboards and providing directional feedback as results come in. The vendor's implementation team owns configuration changes. There is no SFTP file management or ongoing technical lift required from your team after integration is complete.
If the AI agent makes a compliant error on a consumer call, who is liable — the lender or the vendor?
Regulatory liability for consumer-facing communications sits with the lender, not the vendor. This is true regardless of whether the contact was made by a human agent or an AI system. This makes the compliance configuration of the AI agent a direct operational risk for the lender, not a vendor problem to manage independently. The right vendor relationship includes clear documentation of how compliance guardrails are configured, who authorized each configuration, and a complete audit trail for every contact attempt — so that in the event of a regulatory inquiry, the lender can demonstrate exactly what the system was programmed to do and why.
What reporting is available to prove compliance in the event of an audit or regulatory inquiry?
Prodigal provides full audit-grade documentation for every interaction including complete call logs with timestamps, duration, and disposition; consent verification records tied to each contact attempt with the API check result logged; 100% call recordings and transcripts; interaction-level compliance and quality scorecards; opt-out audit trails with timestamp and channel; and configuration change logs showing who changed what and when. Performance reporting gives leadership drill-down by channel, delinquency stage, or segment across contact attempts, RPC rate, payment rate, dollars collected per agent hour, AI containment rate, and transfer rate. Role-based access, scheduled email summaries, and CSV export are all available.
How do I evaluate whether an AI vendor's performance claims are real and not just pilot results?
Ask for production data, not pilot data, and ask for it specifically on subprime auto portfolios. Pilot deployments run on early-stage accounts in controlled conditions and almost always perform well — that is not the same as sustained performance across a full subprime book including 60 to 90 day accounts at scale. Ask the vendor how long their current subprime auto deployments have been live, what percentage of the portfolio each covers, and what the results look like beyond month one. A vendor with genuine production experience will have this data readily available. One who cannot provide it is telling you something important about the maturity of their deployment.
.svg){kind=icon}
.png)
.webp)
.webp)
.jpg)
-min.avif)
.avif)
.avif)
.avif)
.avif)
.avif)