AI voice agents work in lending. But only when they are pointed at the right calls.
The strongest commercial returns come from high-volume, low-judgement interactions: routine servicing queries, payment reminders, balance and settlement enquiries, and early-stage arrears outreach. Complex forbearance conversations, customers showing signs of financial difficulty, complaints and disputed balances all stay with human specialists. That boundary is not a compliance footnote; it is the commercial logic of the whole model.
The correct frame for ROI is not headcount reduction. It is freed agent capacity, faster first contact, better out-of-hours coverage, and higher throughput on routine work so your specialist teams can focus on the calls that genuinely require human judgement.
Key takeaways
- AI voice agents pay off first in high-volume, repeatable servicing and routine arrears interactions, not in complex hardship or dispute handling.
- The FCA's Financial Lives 2024 survey found that 52% of UK adults show at least one characteristic of vulnerability, which makes explicit escalation design non-negotiable, not optional.
- The commercial gain is faster coverage and freed specialist capacity, not indiscriminate cost-cutting.
- Start narrow: one bounded call type with clean escalation rules, then expand once controls are proven.
The practical filter is three questions: Is this call high volume? Is the outcome predictable from the information the customer provides? Does it require agent discretion beyond following a deterministic policy rule? If the answer to the first two is yes and the third is no, it is a strong candidate.
The table below maps the split for most lending operations.
|
Good first AI voice use cases |
Keep human-led |
|---|---|
|
Repayment date and amount queries |
Complex affordability conversations |
|
Outstanding balance and settlement figure requests |
Customers showing signs of financial difficulty or vulnerability |
|
Payment confirmation and receipt queries |
Nuanced forbearance and hardship arrangements |
|
Application and decisioning status checks |
Complaints and disputed balances |
|
Standard ID and verification steps |
Litigation-related matters |
|
Simple payment arrangement setup (policy-driven) |
Any call where the customer deviates from the expected path |
|
Pre-arrears payment reminders and outreach |
Complex repossession or enforcement discussions |
|
Standard product and rate change notifications |
Customers requiring mental health or financial wellbeing support |
Some interactions sit in the middle. A payment reminder that starts routine can quickly surface genuine financial difficulty. A settlement enquiry can become a complaint. The right design is not to categorise these as either safe or unsafe in advance; it is to build the AI to detect deviation early and hand off cleanly.
The practical rule: if the agent handling this call would need to exercise discretion, apply policy judgement, or make a decision that affects a customer's financial position in a meaningful way, it should not be contained by AI. The moment the call requires anything beyond retrieving information or confirming a deterministic policy outcome, escalation should be immediate and seamless.
A motor finance lender we worked with drew this boundary clearly: AI took routine repayment queries and simple arrangement confirmations; specialist agents handled everything involving changed circumstances or hardship. That split, not the technology itself, was what made the business case defensible.
The economics of AI voice in lending rest on a simple principle: routine conversations handled at a lower marginal cost, with coverage extended beyond staffed hours, free your specialist agents to work higher-value interactions.
Indicative figures in the market suggest AI voice can handle routine calls at somewhere in the region of 20p to 30p per minute, compared with five to eight times that for a fully loaded live agent minute. Treat those numbers as directional; actual cost depends on telephony infrastructure, orchestration layer, integration complexity, and how your escalation model is designed. Any vendor quoting a single universal rate should be pressed for a full total cost of ownership model before you build a business case around it.
McKinsey's research on digital-first collections found that well-designed programmes can reduce collections costs by at least 15%, with stronger resolution and repayment performance than traditional approaches. The mechanism is the same as the voice agent model: move lower-complexity work to lower-cost channels, and free specialist capacity for the interactions that need it.
The real ROI frame The gain is not just lower cost per contact. It is:
- More attempts worked in the same staffed period
- Faster time to first contact on early-stage arrears (one mortgage servicer reduced first contact from over six days to just over one day with automated outreach)
- Higher containment on routine servicing, which reduces queue pressure for specialist agents
- Arrangement throughput that does not depend on staffing levels or shift patterns
Vanity metrics like deflection rate look good in a board deck but tell you little about commercial value. The measures that matter for a lending business case are:
If your pilot cannot move at least two of these metrics meaningfully, the scope is probably too narrow or the integration too shallow to demonstrate real value.
Yes, within clearly defined boundaries. This is the section most lenders approach with the most caution, and rightly so. Arrears and collections calls touch customers who may be financially stressed, vulnerable, or both. The answer is not to avoid automation here; it is to be precise about which part of the arrears journey AI can handle responsibly.
The following interactions are well suited to AI voice when escalation rules are explicit and tested:
The escalation design is not optional. It is the mechanism that keeps AI deployment in arrears commercially viable and operationally responsible. For the full governance framework on vulnerable customer handling and Consumer Duty obligations in automated servicing, see our Consumer Duty guide for contact centres and our FCA voice agent compliance overview.
The boundary between routine and complex is where most implementations succeed or fail. A well-designed handoff, with full conversation context passed to the specialist, is worth more than any efficiency gain from containment.
Voice is where most lending automation starts, but the strongest operational gains come when routine servicing works coherently across channels. Customers do not stay in one channel: they might receive an SMS reminder, click through to a web journey, and then call to confirm. If those touchpoints are disconnected, the customer restarts the interaction each time and your agents absorb the duplication.
A mortgage servicer we worked with was running collections outreach across phone, SMS, and web with no shared context between them. Customers were being contacted multiple times on the same matter through different channels, with no visibility across the operation. Consolidating onto a coordinated platform reduced duplicate handling and gave the collections team a single audit trail across all touchpoints.
The operational benefit is not just efficiency. A single platform view improves outcome coding, escalation handoff, and the audit trail regulators expect. It also makes it easier to identify customers who are cycling through multiple channels without resolving their query, which is often an early indicator of difficulty.
For guidance on platform selection across these channels, see our AI virtual agent vs chatbot comparison.
For many lending teams, the fastest ROI is not replacing daytime agents. It is capturing the contact volume that currently falls through the gaps.
Customers do not arrange their financial concerns around your staffed hours. A missed call at 8pm on a payment query can become a missed arrangement, an escalated arrears case, or a complaint. AI voice can hold that contact window open without the cost of extended shift cover.
Where out-of-hours and overflow coverage pays off:
The staffing argument is straightforward: extending coverage through AI costs significantly less than the equivalent live agent hours, and for routine interactions the quality of resolution is comparable. The risk is not in offering the service out of hours; it is in offering it without proper escalation routes for customers who need more than a routine response.
The implementation failures in this space almost always come from starting too broad. The right approach is bounded, measurable, and expandable.
Do not start with forbearance, repossession, or anything that requires agent discretion. The risk is not the technology; it is deploying it before you have the controls, the integration, and the confidence to expand responsibly.
To explore what a bounded first deployment looks like for your lending operation, speak to us about a cost-modelling session or book a readiness review with our CX and contact centre team.
Which lending calls actually pay off to automate? The strongest first use cases are repayment queries, balance and settlement requests, payment reminders, application status checks, and simple arrangement setup where policy rules are deterministic. Complex affordability, forbearance, vulnerability, complaints, and disputes should remain human-led.
Can AI handle collections and arrears calls? Yes, for routine interactions: pre-arrears outreach, simple promise-to-pay capture, standard arrangement setup, and payment confirmation. Any indication of financial difficulty, distress, vulnerability, or dispute must trigger immediate escalation to a specialist agent with full conversation context passed across.
Where should a lender start with AI voice? Start with one bounded, high-volume routine call type with clear intent and a tested escalation path. Prove the controls, integration, and handoff quality before expanding. Avoid complaints, hardship, and complex forbearance until the operational foundations are solid.