AI debt collection system: outbound voice calls with payment capture
An outbound calling system that contacts customers with overdue balances, follows collections stages and takes card payments during the call.
The problem
A telecom provider needed to contact large numbers of customers with overdue balances without growing its human calling team. The calling had to stay compliant with collections expectations, and the business wanted payments captured directly on the call rather than through a separate follow-up step.
Challenges
Calling only when permitted
Collections contact is bound by rules on timing and conduct, so the system had to respect permitted hours, weekends, public holidays and do-not-contact flags on every record. Each dialling decision is checked against those status flags first, so a busy batch can never override the limits the business has set.
Voicemail versus a live person
A collections call reaches a voicemail box as often as a person, and the two need different handling. The system had to tell them apart, then either continue the live conversation or leave a compliant message, without treating every unanswered call as a failed attempt.
Repeated files without duplicate accounts
Overdue account files arrive on a schedule and often contain accounts we have already seen. Ingestion had to upsert records rather than insert them, so an account is never duplicated across runs and never contacted twice simply because it appeared in two files.
How the system works
Lead ingestion
Overdue account files arrive over FTP on a schedule. An automation parses each file and upserts the records into Supabase, so the same account is never duplicated across runs.
Call scheduling engine
A cron-based scheduler checks each record's status flags, such as active today, do-not-contact and permitted time windows. It then batches outbound calls into allowed hours and skips weekends and public holidays.
Staged conversation
An ElevenLabs voice agent follows scripts matched to the collections stage, from an early reminder through to active enforcement. Each stage has its own tone and its own required disclosures.
Voicemail detection
The system works out whether a person or a voicemail answered. It either continues the live conversation or leaves a compliant voicemail message.
SMS fallback
When a call goes unanswered, or as an extra touchpoint, an SMS with a secure payment link is sent automatically.
In-call payment collection
The caller is moved into a keypad-based payment flow linked to a payment processor, so a card payment can be taken with no human involved.
Multi-brand routing
One platform serves several brands. The lead source decides which scripts, voice and caller ID are used for each call.
Post-call logging
Webhooks return the transcript and outcome, which are logged against the lead record. Account status is updated to contacted, promised to pay, disputed or escalated, and that status feeds reporting and later scheduling.
Design decisions
Keypad entry for card payment
When a customer wants to pay during the call, we hand over to a keypad payment flow tied to a payment processor instead of asking for card numbers aloud. This keeps sensitive values out of speech recognition and lets a payment complete without any person involved.
Scripts matched to the collections stage
Each stage of the collections process has its own script, tone and required disclosures. We chose this over one generic conversation because an early reminder and active enforcement call for very different wording, and the business needs each to stay within its approved language.
Brand chosen by lead source
One platform serves several brands, and the lead source decides the scripts, voice and caller ID used. Keeping that decision in the data rather than in separate builds means adding a brand is configuration, and the customer hears the brand they recognise.
Result in production
Collection calls run at scale without a matching increase in headcount. Every call is logged and every account has a tracked status that informs the next scheduling cycle.
Collection calls can run at scale without the calling team having to grow in step with the number of accounts.
Every call is logged with its transcript and outcome, so the business can see exactly what happened on each account.
Tracked account status feeds the next scheduling cycle, so later contact follows what actually happened rather than a fixed list.
What we would do differently
We would move call scheduling out of cron and into a queue-backed worker earlier. It scales more cleanly as call volume grows and makes retries and rate-limit handling far less painful.
Who this pattern fits
Although we built it for a telecom provider, the pattern suits any organisation running high-volume, compliance-sensitive outreach from a list, including utilities, subscription services and other recurring-billing businesses. The stack is stable, can be self-hosted where needed, and is manageable by a small team.
Tools used
ElevenLabs
Speech synthesis for the voice agent, chosen for latency and call quality.
Twilio
Telephony carrier for inbound and outbound calls, SMS and call recording.
Twilio Pay
PCI-compliant keypad card capture during live calls.
Braintree
Payment processor for card capture and reconciliation.
n8n
Workflow runtime and orchestration, with version-controlled logic.
Supabase
PostgreSQL database, authentication and storage; the source of truth for records.
FTP ingestion
Scheduled intake of account files from partners.
Cron scheduling
Batch jobs, reconciliation and daily roll-ups.
Related
Frequently asked questions
Is an automated collections system compliant with regulation?
The system runs approved scripts, respects opt-outs, records consent and passes edge cases to a person. It supports your compliance process but does not replace it. Decisions about what is permitted remain with the business.
How does a customer pay during the call?
The voice system hands the caller into a keypad-based payment flow linked to a payment processor, so card details are entered on the phone rather than spoken aloud. If a call goes unanswered, an SMS with a secure payment link can be sent instead.
Can this approach be used outside telecom?
Yes. It fits any setting where list-based outbound dialling, voicemail detection, payment capture and structured record-keeping matter, such as utilities or subscription billing.