Every failed payment forces one decision: does this customer get a call, or only a message? It turns on why the payment failed, how much is at stake, and whether the customer can fix it in one step. An expired card needs the customer to act, a bank-side glitch often clears on a retry, and a small amount may not justify a call that costs more than a text. Teams that dial every failure watch cost per recovered payment climb and complaints rise, while teams that call a small, fixable slice let cheaper channels handle the rest. You will be able to triage failures, set the trigger, send the link and prove what you recovered.
Call only the fixable failures
Use this table to triage every failure before anything is dialled.
| Failure reason | What the customer can do | Recommended action |
|---|---|---|
| Expired card | Enter the new card details in one step | Call now, with a payment link |
| Insufficient funds | Top up or switch to another method | Call now if the amount is large. Otherwise SMS only |
| Issuer decline, no reason given | Switch card or contact their bank | SMS only, offering another method |
| Suspected fraud block | Verify with their bank | Route to a human |
| Bank-side technical error | Nothing, it is not their fault | Retry silently |
| Cancelled mandate | Re-authorise, or confirm they meant to leave | SMS with a re-authorisation link. Human follow-up if they ask to leave |
Call when the customer must act and can act in one step. An expired card is the clearest case. Card networks treat it as a hard decline, so retrying will not help and the customer has to supply a new card or method.
Do not call for failures a silent retry will clear. Issuer unavailable and system malfunction errors are soft declines, safe to retry with backoff. A call about a problem that fixes itself in an hour teaches customers to ignore your number.
Insufficient funds is also a soft decline, because the balance may recover by payday. That makes it a judgement call on amount, and any retries should be spaced by hours or days rather than fired back to back.
Keep the AI call away from fraud blocks and disputes. These need identity checks and reassurance, so a human should handle them.
If a failure does not fit one of these rows, default to the cheaper channel and let a person decide whether it earns a call.
Wire the trigger, then the clock
Build the path from failure to first dial as six steps, in this order.
- Emit the event. Your billing system should publish every failure with the decline reason, the amount and a contact number.
- Filter with the triage table. Only "call now" rows continue. Everything else goes to SMS, a silent retry or a person.
- Wait. Hold the call for a few hours so retries and the customer's own fixes can land first. A customer who has already updated their card should never get a call.
- Check suppression. Drop anyone on the do-not-contact list, anyone who has already paid and anyone with an open dispute.
- Dial inside permitted calling hours. Use the customer's local time, not yours.
- Redial on a schedule. Start with a few attempts across morning and evening windows, then test your own best windows. Many people answer on the second or third try.
Add a cheaper trigger before any failure happens: for cards that will expire before the next charge, send a reminder ahead of the due date. A card updated in advance is a failure you never have to triage.
A consumer lending arm we work with calls customers before the due date to check that funds are available, and calls again after a failed debit to offer another payment channel. The second call is the point. A failed-debit call should hand the customer a way to pay, not only remind them that the payment failed.
Campaigns in Insight7 have calling hours, retry rules, a do-not-call list and a test call to your own phone, in beta, and you can see them through our book-a-demo page.
End the call with the link
Build the call around one outcome: the customer has the payment link in their hand before they hang up.
- Open with who and what. Say the caller is an automated assistant and name your company.
- Confirm the person. Ask whether you are speaking to the right customer by first name. Never read out balances or account details until they confirm.
- State the problem neutrally. "A payment did not go through, and there is one quick fix." No blame, no policy.
- Offer the link. Ask to send it now, then wait for the customer to confirm the number or channel.
- Close on the outcome. Handle three cases:
- Pays now: stay on the line until they confirm, then thank them and end the call.
- Will pay later: log a promise with the date, and let the follow-up use it.
- Refuses or disputes: hand to a human. Do not argue or persuade.
Never take card numbers by voice. The link exists so card details go into a secure page. Aim for a call under a minute.
The common mistake is a long policy preamble before the ask. The customer hears a minute of terms and hangs up before any link is sent, which costs the whole recovery.
Buyers worry that automated calls sound scripted and fall apart when the customer goes off script. The handover is the other worry they raise unprompted: is it polite, and what happens with a disputed amount. Test both before you dial real customers. Interrupt the agent, ask who it is, say the amount is wrong and see whether it hands over cleanly.
Send the link while they listen
Send the payment link by SMS while the customer is still on the call. A link that arrives as they speak can be opened at once. A customer told to find the app or portal later has to remember, search and log in, and many will not. Treat this as a hypothesis to test, since it rests on a shorter path to paying. Compare it against SMS-only recovery in your own data.
Design the link so it is safe to send to a phone:
- One link per customer. Use a random, short-lived, server-generated token. Never put card data or account details in the URL.
- Single use. Invalidate the link once it is redeemed. A retry gets a new link.
- Short expiry. Long enough to survive SMS delays, short enough to limit replay risk.
- One tap to the fix. It should open a hosted page showing the amount owed, with the card update or retry on that page. Card details go through a secure form, so raw numbers never touch your systems.
- A short, recognisable text. Use a sender name the customer knows and one line of copy, with no balances or other sensitive data.
If the SMS fails to deliver, the agent should offer another route while the customer is still listening, such as a wallet, a bank transfer or a callback from a person. If the link goes unopened, send one reminder after a set number of hours, and test that delay rather than guessing it.
The SMS can go out through your own messaging or payments stack. Give each link an ID that is also written to the call record, so the two can be joined later. Buyers who run voice and SMS side by side say the synced ID is what matters, not which system sends the text. That shared ID makes the tracking in the next section possible.
Set guardrails before you dial
Settle these rules before the first call goes out.
- Disclose that the caller is AI, at the start of the call. The rules differ by region. In the EU, Article 50 of the AI Act has applied since 2 August 2026 and requires any AI that talks to people to say so from the first interaction. In the US, AI-voice calls already need consent under the TCPA, and a federal rule on stating AI at the start was still not final as of mid-2026. Colorado will require public-facing conversational AI to disclose itself from 1 January 2027. Disclosure is spreading state by state, so build it in now.
- Record consent for every number you call, and store it where the campaign can check it.
- Keep to permitted calling hours and respect the do-not-contact list.
- Give every call a clear route to a human. In a recent Gartner survey, 87% of customers said a company using GenAI for service must offer one. Define that route for disputed and sensitive cases before launch.
Track recovered money, not calls
Build the funnel as a chain: failed, eligible, dialled, answered by a person, link sent, link opened, paid, and recovered value inside a set window such as seven days. Each step shows where customers fall away. Do not count a call as reached if only the AI spoke, so a voicemail greeting never inflates your answer rate. Read the transcripts of answered calls and log the outcome of each.
A recovery rate alone does not prove the call added anything, because many customers would have paid after an SMS or a retry. Hold out a random group that gets only the SMS or your normal dunning flow, and compare recovery rate and recovered value against the called group. A cleaner alternative is a sublist reachable only by voice, so any conversion from it belongs to the call. Do not ask the customer to quote a code mid-call, because people cannot note one down while they are talking. Give each customer a unique link ID instead, so payment events join back to the call.
Start with a small pilot on a sample. Buyers do not want to pay to find out whether it works. They want a short, low-risk test and payment only after a measured lift.
Then set the money against the cost. Track cost per recovered unit against the amount recovered, because a finance lead needs a number to carry upstairs. Measurement is the gap: in one Gartner review of 432 AI customer-service use cases, 42% could not be measured at all.
Add the complaint and opt-out rate as a guardrail. If recovery rises while opt-outs climb, you are buying payments with goodwill.
Related reading: 7 Signs Your Email and SMS Onboarding Sequence Needs a Voice Channel, How to Notify Customers About Price Changes and New Fees With AI Voice Calls and AI Voice Agent Use Cases for Customer Onboarding, Activation and Retention With Examples.
Judge it by money recovered
If you run collections, billing or retention, the number that earns this programme its budget is the one finance can find in the bank. Calls dialled and links sent are only activity.
Give your team one rule they can repeat: call only when the customer can fix it in one step, send the link before they hang up, and count only money that landed.

