Which Collections Calls to Automate First

Collections is not one job

Ask an operations lead which part of collections they would automate and the answer is usually "the calls". But an accounts-receivable process is not a single kind of call. A courtesy reminder placed before a due date and a conversation with someone who has just lost their job are both collections calls in the sense that they sit in the same queue, and almost nothing else about them is the same: not the script, not the authority required, not the cost of getting it wrong.

That distinction is the entire decision. Automate the wrong stage and you have built something that will, sooner or later, say a scripted sentence to someone in genuine difficulty. Automate the right stage and you get calls made that nobody was ever going to make by hand.

The question is not whether

Nobody sensibly asks any more whether a voice agent can hold a collections conversation. The useful question is which stage of your ledger it should touch first, and where it must stop and hand over. Those two answers are the deployment.

Why the bottom of the ledger never gets a call

Every collections team runs on a constraint that rarely gets written down: a person makes a fixed number of calls in a day, and the ledger is longer than that. Something has to decide who gets called. In practice that something is the size of the balance.

Prioritising by balance is a rational rule and it has a predictable side effect. The small accounts at the bottom of the list get a letter, then an email, then nothing. Many of them are not refusals at all. They are people who put the payment behind something else and never got a prompt at the moment when acting on it was easy. The tail of a ledger is not the uncollectable part. It is the uncalled part.

That gap is the only honest reason to reach for an automated reminder stage. Not that a machine collects better than a good collections officer, because it does not. That it makes the calls nobody had the capacity to make.

The stages that automate well

The stages that suit a voice agent share three properties. The script barely changes between accounts. The volume is high enough that nobody works the whole of it by hand. And the outcome is a fact to record rather than a judgement to make.

Pre-due courtesy calls. A reminder placed before the due date, which is not collections at all. It is a service call that stops the account ageing in the first place. Cheap to run, trivial to script, and almost universally skipped, because a team working overdue accounts has nothing left over for accounts that are not overdue yet.
The first reminder after the due date. The work here is reaching the person, confirming they know, and agreeing a date. It is close to identical across thousands of accounts, which is exactly the shape of work that automation handles well and that people find deadening.
Broken arrangements. Someone promised to pay and the date passed. The person already agreed once, which makes this a different conversation from a first reminder, and it is one of the most often skipped: it is uncomfortable to make, and the team has already moved on to newer receivables. An agent makes it on the day it falls due, in the same words, every time.
Stale contact data. Some share of any ageing ledger is not people refusing to pay but numbers that no longer reach anybody, and until the file is worked you do not know how big that share is. Finding out is worth doing before you buy another round of contact data against those records.

Notice what those have in common. None of them asks the agent to decide anything. They ask it to call, to ask, to listen, and to write down the answer in a form somebody can query later.

The calls that must stay with people

These limits are not a roadmap of features arriving next quarter. They are design decisions, and the reason to write them down before the first call is that each one carries a cost that is not measured in recovery rate.

Hardship. Someone describing illness, redundancy or a household in crisis needs a person with discretion, and needs one immediately. The agent's job is to recognise it, stop, and route the case with the transcript attached so that nobody has to start the story again.
A charge under dispute. When a customer says the amount is wrong, the call has stopped being a collections call and become a service investigation. Continuing the collections script past that point is bad practice, and where debt collection is regulated it is usually a breach as well.
Anything that needs authority. Write-offs, restructures, extensions beyond your published options, settlements. These are decisions with a mandate behind them. An agent can capture the request precisely and route it. It should never be the thing that grants it.

There is a fourth limit that is less about judgement and more about hygiene: payment credentials. A card number spoken to an agent is a card number sitting in a transcript, and no recovery rate justifies that. The correct behaviour is a transfer to the payment channel you already run, and a record that the transfer happened.

The plainest test

If a stage needs the agent to weigh someone's circumstances, or to grant something, it is not a stage to automate. If it needs the agent to call, ask a fixed question and record the answer, it probably is.

Write the rules before you write the script

The order matters more than it sounds. An agent built script-first inherits its boundaries by accident, from whatever the prompt happens to say that week. Built rules-first, the same document that constrains it becomes the rubric its calls are scored against, which is what turns a written policy into something enforceable.

Contact hours, and how often a single account may be called. Usually set by regulation, sometimes by your own policy, and always the first thing anyone auditing the operation asks about.
Who may be contacted. The account holder, and nobody else: not a colleague, not a relative, not whoever answered the number.
Mandatory disclosures. What the agent must say about who it is and who it is calling on behalf of, at the point in the call where your market requires it said.
The stop conditions. What happens when someone disputes the amount, asks not to be called again, or says something that signals hardship. Each needs a scripted response and a defined destination.
The handover points. Which situations end the automated call, who receives it, and what travels with it.

That document is also the answer to the question a compliance officer actually asks, which is rarely "is it accurate?" and almost always "can you show me how it behaved on the call I am asking about?"

What to measure, and against what

The temptation is to compare the automated stage against nothing at all, on the grounds that the accounts it works were not being called before. That produces a flattering number and teaches you nothing you can act on.

Take a segment, not the ledger

One ageing bucket, small enough that you can read every transcript it produces in the first weeks. An automated collections stage that nobody reads is how a tone problem becomes a complaint.

Compare like with like

Contact rate, arrangement rate and the value of those arrangements, on that segment, against the same segment worked the way you work it today. If the segment genuinely was never called, say so out loud and measure against what it recovered unattended.

Watch the arrangements, not the promises

The number that matters is not how many people agreed to pay. It is how many of those agreements were kept, which you only learn a cycle later, and which is the only figure that tells you whether the calls were persuasive or merely polite.

Run the per-minute arithmetic before any of it, so you know what the comparison is against. Automated calls are billed by conversation time, which changes which accounts are worth calling rather than only changing what a call costs.

Where this leaves the decision

A voice agent belongs at the early, repetitive, high-volume end of an accounts-receivable process, making the calls that capacity has always rationed away. It belongs nowhere near the calls that require somebody to exercise judgement about a person's circumstances. Teams that hold that line recover money from a part of the ledger that was written off by default.

Teams that do not hold it get the one outcome a collections operation cannot afford, which is a complaint with a recording attached. The line is not difficult to draw. It just has to be drawn before the first call, in writing, by somebody with the authority to draw it.

Frequently asked questions

Which parts of a collections process can an AI voice agent handle?

The early, repetitive, high-volume stages: courtesy calls before the due date, the first reminder after it, following up arrangements that were not kept, and working through contact data to find out which records still reach anybody. What those have in common is that the script barely changes between accounts and the outcome is a fact to record rather than a judgement to make. The agent calls, asks, listens and writes the answer into a field.

Which collections calls should never be automated?

Three kinds. Anything involving hardship, where someone is describing illness, redundancy or a household in crisis and needs a person with discretion straight away. Anything where the customer disputes the amount, because at that point the call has become a service investigation rather than a collections call. And anything requiring authority: write-offs, restructures, settlements, or extensions beyond your published options. The agent should recognise all three, stop, and route the case with the transcript attached.

Can an AI voice agent take a payment during the call?

It should not, and that is a design decision rather than a missing feature. Card numbers, security codes and bank credentials spoken to an agent end up in a transcript. The correct behaviour is to transfer the customer to the payment channel you already run, or send the payment link by whatever means you have configured, and record that it happened. Keeping payment instruments out of the conversation keeps them out of the recording as well.

How do you keep an automated collections call inside your compliance rules?

By writing the rules before the script rather than after it. Contact hours, how often one account may be called, who may be contacted, any disclosure your market requires, what to say when someone disputes the amount or asks not to be called again, and where the call must be handed to a person. Those rules become the agent's instructions and also the rubric its calls are scored against, so the same document that constrains the agent is the one that proves how it behaved.

Does the customer know they are not speaking to a person?

That depends on the rules you set, and in markets where a disclosure is required it should be scripted, mandatory and checked on every call rather than left to the model. Building a collections agent designed to be mistaken for a human where the local rules say otherwise is a risk taken on the operator's behalf, not a feature. On the conversation itself, a well-built agent answers fast enough and handles interruption well enough that it sounds like a call rather than a recording.

Is an automated call worth making for small balances?

That is exactly where the arithmetic changes. A human call has a floor cost that makes chasing small balances unprofitable, which is why teams prioritise by balance and why the tail of the ledger goes uncalled. An agent is billed by conversation time, so the question stops being whether the balance justifies a person's time. Put your own volumes and balances through the calculation rather than taking anyone's word for it.

How is this different from an autodialler playing a recorded message?

A recorded message broadcasts and waits. It cannot answer a question, take a date, or tell the difference between somebody disputing the charge and somebody who simply forgot. An agent holds the conversation: it answers from the material you gave it, agrees a specific date and amount, and classifies how the call ended. The difference shows up in the arrangement rate rather than the contact rate, and in having a ledger you can query instead of a list of dial attempts.

What should we measure to know whether it worked?

Contact rate, arrangement rate and the value of those arrangements, measured on one defined segment against the same segment worked the way you work it today. Then, a cycle later, the figure that actually matters: how many of those arrangements were kept. Agreement rates alone reward calls that were pleasant rather than calls that were effective, and the gap between the two only becomes visible when the promised date arrives.