The Collections Problem: Agreements Without Payments
Debt collection runs on a simple loop: contact the debtor, negotiate a payment arrangement, collect the money.
The first two steps have been optimised heavily. Diallers reach more people. Analytics predict the best time to call. Scripts are tested and refined. AI voice agents are now handling the conversations themselves — negotiating payment plans, confirming balances, and reaching agreements at scale.
But the third step — collecting the money — is where the loop breaks.
A human agent or AI negotiates a payment arrangement. The debtor agrees to pay £150 today and £150 next month. The agent says: "Great, I'll send you a link to make the payment." Or: "You can pay on our website." Or: "I'll transfer you to our payment line."
The debtor hangs up. The commitment cools. The link sits unopened in an inbox. The website visit doesn't happen. The transfer drops.
Industry data tells the story: the gap between "agreement to pay" and "payment received" is where collections operations lose 30-50% of promised revenue. Every step between the verbal commitment and the actual transaction is a leak.
The fix is architectural, not motivational. Capture the payment during the call, at the exact moment the debtor agrees. No handoff. No link. No second step.
How AI Voice Agents Collect Payments
An AI voice agent handles the entire collections call — from identifying the debtor to confirming the payment. Here's an outbound collection:
AI Agent: "Good morning. I'm calling from ABC Collections regarding an outstanding balance of £312.40 on your account. Am I speaking with James?"
Debtor: "Yes, that's me. What's this about?"
AI Agent: "You have an outstanding balance of £312.40 from your account with XYZ Services, originally due on 15 January. I can help you resolve this today. Would you like to pay the full amount, or set up a payment plan?"
Debtor: "I can do £150 now and the rest next month."
AI Agent: "I can arrange that. £150 today and £162.40 on 18 March. I can take the first payment now. Please enter your 16-digit card number using your phone keypad."
The debtor types their card number. DTMF tones are captured within the PCI-certified payment layer — stripped from the audio stream. The AI never sees the card data. Neither does the collections agency or the original creditor.
AI Agent: "Thank you. I'm processing £150 now."
The payment layer validates the card, routes to the creditor's configured PSP, and processes the transaction. A token is stored for the scheduled second payment.
AI Agent: "Your payment of £150 has been processed. Your next payment of £162.40 is scheduled for 18 March using the same card. You'll receive a confirmation by text. Is there anything else?"
Payment captured at the moment of agreement. Card tokenised for the scheduled instalment. Human involvement: none. PCI exposure for the collections agency: none.
Why This Matters for Collections
The Commitment Window Is Narrow
Behavioural economics is clear: the probability of payment drops rapidly with time after a verbal agreement. A debtor who agrees to pay during a call has the highest likelihood of following through in that exact moment. Every minute, hour, or day between agreement and payment reduces the probability.
Collections operations that capture payment during the call — not after — maximise recovery rates from every contact. The AI agent doesn't say "I'll send you a link." It says "I can take your payment now."
Scale Without Headcount
Collections is labour-intensive. Human agents make calls, negotiate arrangements, and process payments — one at a time. Scaling means hiring more agents, more training, more management, more compliance oversight.
AI voice agents scale differently. They handle thousands of concurrent calls. They work evenings and weekends. They don't need breaks, supervision, or performance reviews. They follow compliance scripts perfectly, every time.
For BPOs handling collections on behalf of multiple creditors, AI voice agents with payment capability transform the unit economics of the business.
PCI Compliance Is a Persistent Liability
Collections environments handle card data constantly. Agents hear card numbers. Recordings capture them. CRM systems store them. Every one of these touchpoints is a PCI surface.
The compliance cost is real: maintaining PCI DSS across a collections contact centre means monitoring agent behaviour, securing recording systems, segmenting networks, and conducting annual audits. For BPOs serving multiple clients, the compliance burden multiplies.
AI voice payments eliminate this exposure. Card data is captured in a PCI-certified environment that's completely isolated from the collections agency's infrastructure. The agency's PCI scope drops to SAQ-A. Recordings are clean. No card data enters any system the agency controls.
Multi-Client, Multi-PSP Reality
BPOs and collections agencies serve multiple creditors. Creditor A uses Worldpay. Creditor B uses Stripe. Creditor C uses a regional processor mandated by their treasury team.
The collections agency can't standardise on one PSP. Each creditor's payments must route to their configured gateway. A PSP-neutral payment layer handles this through a single integration — the AI agent triggers a payment, the layer routes to the correct PSP based on the creditor's configuration.
Payment Plan Automation
Collections isn't just about one-off payments. Most arrangements involve payment plans — agreed schedules of instalment payments over weeks or months.
AI voice agents with tokenisation capability automate the entire plan:
1. Negotiate the arrangement. The AI agent confirms the balance, discusses options, and agrees a payment schedule with the debtor.
2. Capture the first payment. Payment is taken during the call via DTMF. The card is tokenised.
3. Schedule future payments. The stored token is used for scheduled instalment payments. No need for the debtor to re-enter card details.
4. Handle failed payments. If an instalment fails (expired card, insufficient funds), the AI agent can call the debtor proactively: "Your scheduled payment of £162.40 didn't process. Would you like to update your card details now?"
5. Confirm completion. When the balance is cleared, the AI agent (or an automated message) confirms the account is settled.
This loop — negotiate, capture, schedule, retry, confirm — runs autonomously. Human agents only need to intervene for disputed balances, complaints, or complex hardship cases.
Compliance and Regulatory Considerations
Debt collection is heavily regulated. The Consumer Credit Act, FCA debt collection guidelines, OFCOM rules on automated calling, and PCI DSS all apply. AI voice payment systems must satisfy all of them.
FCA Debt Collection Rules
AI voice agents must comply with the same standards as human agents: treating customers fairly, recognising signs of financial difficulty, offering appropriate forbearance, and not applying inappropriate pressure. The AI's conversational logic must include these safeguards — escalating to a human agent when the debtor indicates hardship, providing required regulatory disclosures, and maintaining records of the interaction.
OFCOM and Automated Calling
Outbound AI calling must comply with OFCOM rules on automated and predictive diallers. This includes abandoned call rates, caller ID presentation, and opt-out mechanisms. AI voice agents must identify themselves clearly and provide a way for the debtor to speak to a human if requested.
PCI DSS
Card data captured during the call must be processed within a PCI DSS Level 1 certified environment. The collections agency's systems — diallers, CRM, recording platforms, analytics — must never see raw card data. DTMF tones are stripped from the audio before reaching any system outside the PCI boundary.
Call Recording
Collections calls are typically recorded for quality assurance and dispute resolution. With DTMF tone masking, recordings contain the full conversation but no card data — the payment segment sounds like silence or masking tones where the debtor entered their card number. This satisfies both QA requirements (the conversation is captured) and PCI requirements (no card data in recordings).
The Architecture
Collections AI voice payments use the same foundational architecture as other AI agent payment deployments, with specific requirements for the collections use case:
DTMF Capture with Tone Masking
Card data entered via keypad, captured in PCI-certified environment, tones stripped from audio. Same mechanism as voice payments in any context — but critical here because collections calls are almost always recorded and audited.
Multi-PSP Routing
Each creditor's payments route to their configured PSP. The collections agency integrates once; the payment layer handles routing per creditor.
Tokenisation and Scheduling
Cards tokenised on first capture. Tokens reused for scheduled instalment payments without the debtor needing to re-enter details or receive follow-up calls for each payment.
SMS Payment Link Fallback
For debtors who prefer not to use the keypad, the AI sends a branded payment link via SMS. The debtor pays on their device while the call continues. This is particularly useful for mobile-only debtors who may find keypad entry difficult on a smartphone.
Real-Time Balance Updates
The payment result must feed back into the collections management system immediately. The debtor's balance updates in real time. The scheduled arrangement is recorded. The account status changes from "active" to "arrangement in place."
Inbound vs. Outbound
Outbound Collections
The AI agent calls debtors proactively. This is where the highest volume and biggest opportunity exists. Outbound AI calls can cover thousands of accounts per day — far more than a human agent team.
The call flow: identify the debtor, confirm the balance, negotiate an arrangement, capture payment. All in one call. The AI can adapt its approach based on the account — amount owed, age of debt, previous contact attempts, and the debtor's responses.
Inbound Payment Calls
Debtors call in to make a payment. The AI agent answers, identifies the account, confirms the balance, and processes the payment. This replaces the IVR payment line with a conversational experience — the debtor can ask questions, discuss their balance, and pay in the same interaction.
Failed Payment Follow-Up
When a scheduled payment fails, the AI calls the debtor to collect. "Your payment of £162.40 scheduled for today didn't process. Would you like to update your card details?" This is a high-value automation — failed payment recovery is typically manual and time-sensitive.
What BPOs and Collections Agencies Need
If you're evaluating AI voice payment infrastructure for debt collection:
PCI DSS Level 1 — non-negotiable for any system handling card data. The certification must cover the DTMF capture environment, not just a web form.
Multi-PSP routing — you serve multiple creditors with different PSP relationships. The payment layer must route per creditor without separate integrations.
Tokenisation and scheduling — capture a card once, use the token for scheduled instalments. Failed payment retry logic should be built in.
Regulatory compliance tooling — the AI must support FCA-compliant conversation flows, hardship escalation, and OFCOM-compliant outbound calling.
Recording-safe capture — DTMF tone masking must produce clean recordings with no card data. This is audited.
Real-time integration — payment results must feed back into your collections management system immediately for balance updates and arrangement tracking.
SMS fallback — payment links for debtors who prefer visual checkout or can't use their keypad.
FAQ
Can AI voice agents really negotiate debt payments? Yes. AI voice agents can confirm balances, discuss payment options, negotiate instalment plans, and capture payments — all in a single call. For complex hardship cases or disputes, the AI escalates to a human agent.
Is automated debt collection calling legal? Yes, subject to OFCOM and FCA regulations. AI voice agents must identify themselves, provide required disclosures, offer a route to a human agent, and comply with abandoned call rate limits. The technology is compliant when properly configured.
What about vulnerable customers? AI voice agents should be programmed to recognise indicators of vulnerability or financial hardship — language patterns, tone, explicit statements — and escalate to a trained human agent. This is both a regulatory requirement and an ethical obligation.
How does multi-creditor PSP routing work? The collections agency integrates once with a PSP-neutral payment layer. Each creditor is configured with their preferred PSP. When the AI captures a payment, the layer routes the transaction to the correct gateway based on which creditor the debt belongs to. No per-creditor payment integration required.
What completion rates can we expect? Industry benchmarks for AI voice payment capture in collections aren't yet standardised, but the principle is clear: capturing payment at the moment of agreement consistently outperforms post-call payment links or portal redirects. PolyAI's 75% completion rate across regulated industries provides a reference point.
Can the same infrastructure handle inbound and outbound? Yes. The payment capture mechanism (DTMF, SMS link) is the same regardless of call direction. The AI agent's conversational logic adapts — inbound calls focus on processing an intended payment; outbound calls include negotiation and arrangement.
Automate collections payment capture — at the moment of agreement. Shuttle connects AI voice agents to 40+ payment gateways with PCI DSS Level 1 compliance, DTMF capture, tokenisation for payment plans, and multi-creditor PSP routing.
Book a Call | See Voice Checkout