AI voice calls work well once someone answers. Latency is no longer a problem, and simple phone menus are handled fine. The real constraint is connect rate, which is exactly the constraint your human AR rep already faces. Use voice as a late stage escalation channel on significantly overdue invoices, not as your primary follow-up. SMS gets more responses for less effort.
The AI calling question comes up on almost every demo, and it always arrives with a specific flavor of skepticism.
"I have used AI calling tools on the sales side. There is this latency in the conversation. Does that happen here?"
It is a fair thing to be suspicious about. Plenty of people tried AI voice agents in 2024, sat through the half second pause before every reply, and concluded the technology was not ready. That impression has stuck longer than the problem itself has.
So here is the honest version, including the part that does not flatter the feature.
The good news: the conversation itself works
The latency problem is largely solved. Current voice models handle interruptions, natural pauses, and genuine back and forth well enough that most people on the receiving end of a collections call do not clock it as AI. We have customers whose clients have had complete conversations about an invoice without any idea they were not speaking to a person.
Simple phone menus are handled fine. Press one for accounts payable is not a difficult problem. Deeper or stranger dial trees, and systems that require speaking a name into a directory, are less reliable, though this comes up less often than you would expect. Most B2B accounts payable contacts are reached on a direct line or extension, not through a main switchboard.
Context is the part that genuinely differentiates an AI collections call from a generic voice agent. The system already knows the invoice number, the amount, the due date, how overdue it is, what else the customer has outstanding, and what was discussed in previous emails. When the customer asks "which invoice is this" or "how much do we owe in total," the answer is available immediately. That is a better experience than a temp reading off a spreadsheet.
The bad news: nobody picks up
Here is the part vendors tend to leave out of the demo.
The limiting factor in AI calling is not the quality of the conversation. It is that the call has to be answered, and most of the time it is not.
You do not get anything close to a ninety percent connect rate on collections calls. You get a fraction of that, and the reason is not technical. A customer with a sixty day overdue invoice who has ignored three emails is not sitting by the phone hoping to hear from you. They are avoiding the call on purpose. An AI dialing the number does not change their motivation.
This is worth sitting with, because it reframes what the feature is actually for. AI calling does not solve the connect problem. It solves the cost of the connect problem. Your AR rep spends the same time dialing, waiting, and getting no answer as an automated system does, except the automated system is not spending a salary doing it and does not get discouraged after the fifth unanswered call.
The conversion is not "AI calls work better than human calls." It is "the calls that were not happening because nobody had time now happen."
What about voicemail?
Leave them, but do not build your expectations on them.
Collections voicemails have low response rates regardless of who leaves them. The customer ignoring your calls is generally ignoring them deliberately, and a message does not alter that calculation. A voicemail is worth leaving for the record, and occasionally it reaches someone who genuinely missed the call and calls back the same day. Neither of those is a reason to adopt the channel.
If a vendor pitches AI voicemail drops as a growth lever, be skeptical.
Where voice actually belongs in the sequence
The practical answer, based on how our customers configure it, is late.
Most set voice to trigger once an invoice is meaningfully overdue, commonly somewhere between 60 and 90 days, after email and SMS reminders have already run. One customer starts calls at day 70. Before that point the sequence is email with some SMS.
The logic is straightforward. On a slightly overdue invoice, a phone call reads as disproportionate, and customers notice the mismatch. On a ninety day invoice, a call reads as normal collections practice, because it is. The channel carries an escalation signal, and using it early spends that signal for nothing.
There is also a simple risk calculation. On a genuinely delinquent account, you have less relationship left to protect. Testing the channel there costs you very little and tells you quickly whether it works for your customer base.
| Days overdue | Channel | Why |
|---|---|---|
| Before due date | Gentle nudge, measurably improves on time payment | |
| 1 to 30 | Friendly, documented, low friction | |
| 15 to 60 | Email plus SMS | SMS lifts response rate substantially |
| 60 to 90+ | Add voice | Escalation signal, different channel breaks the pattern |
| 90+ | Voice plus internal escalation | Loop in the account owner or project manager |
SMS quietly outperforms voice
If we are being honest about the data across our customer base, SMS is the channel that punches above its weight. It gets read, it does not require the customer to be free at a particular moment, it generates more replies than voice for less cost, and it is far less intrusive.
Multi channel sequences beat single channel sequences consistently, but that is not because every channel contributes equally. It is because different customers respond to different things, and hitting the same person with a second modality breaks the pattern of ignoring the first. Adding SMS to an email sequence usually produces a bigger lift than adding voice does.
Voice earns its place in specific situations: high value invoices where the amount justifies the effort, accounts that have gone quiet across every other channel, and situations where you need an actual conversation to find out why payment has stalled. Outside those, it is the more expensive way to get a smaller result.
For a fuller picture of how the channels stack, see multi channel payment chasing.
The things worth checking before you turn it on
If you are evaluating AI calling as part of an AR platform, these are the questions that separate a usable feature from a demo trick.
Does the call have invoice context? If the AI cannot answer "how much do we owe in total" or "which invoice is this," it is a robocall with better diction. Context is the whole value.
What happens when the customer says something unexpected? A dispute, a request to speak to someone, a claim that they already paid. The right behavior is to capture it, end the call gracefully, and surface it to your team, not to improvise a commitment on your behalf.
Can you exclude specific customers? You will have accounts where a partner or executive owns the relationship and no automated call should ever go out. Being able to suppress those individually is not optional.
Where does the call land afterward? A call that happened and was never logged is worse than no call. The transcript, the outcome, and any promise to pay should sit against the invoice alongside the email and SMS history.
Are failures visible? Wrong numbers, landlines that cannot receive SMS, and disconnected lines are extremely common, and they are usually a symptom of stale customer data rather than a platform problem. You want those surfaced, because they are worth fixing at the source.
The realistic expectation
AI calling is a good feature that gets oversold. It will not transform your collections on its own, and any vendor implying otherwise is describing a version of the world where customers answer the phone.
What it does is add a channel that was previously too expensive in staff time to use consistently, at the point in the sequence where a different channel actually changes the outcome. Combined with email and SMS running reliably underneath it, that contributes to the same result our customers see across the board.
TDG Inc reduced manual follow-ups by 80% and cut DSO by 15 days within three months. Another Yonovo customer took DSO from 65 days to 41. Across our customer base, automated follow-ups recover an average of 15 hours per week.
The calls are part of that. They are not the reason for it.
If you want to hear what one sounds like against a real overdue invoice, book a demo and we will run it live. It is a more useful test than any description.
Frequently Asked Questions
Do AI voice calls sound robotic to customers?
Not in the way they did two years ago. Current voice models handle natural pauses, interruptions, and back and forth conversation well enough that most people on a collections call do not identify it as AI. The awkward latency gap that made earlier AI calling tools obvious has largely closed. The remaining tell is usually an unusual question that falls outside the invoice context, not the voice itself.
Can an AI collections call handle a phone menu or dial tree?
Simple menus, yes. Instructions like press one for accounts payable are handled reliably. Deep or unusual menu trees, or systems that require speaking a name from a directory, are less consistent. In practice this matters less than expected, because most B2B AP contacts are reached at a direct line or extension rather than through a main switchboard.
Should AI calls leave voicemails?
They can, but expect little from it. Collections voicemails have low response rates whether a person or an AI leaves them, because a customer avoiding your calls is usually doing so deliberately. A voicemail is worth leaving for the record and for the occasional customer who genuinely missed the call, but it should not be the reason you adopt the channel.
When in the sequence should AI calls be used?
Late. Most Yonovo customers configure voice to start once an invoice is significantly overdue, commonly around 60 to 90 days, after email and SMS have run. Early reminders are better served by email and SMS, which are cheaper, less intrusive, and preserve the customer relationship. Voice is an escalation signal, and it loses that meaning if you use it on day five.
Is SMS more effective than AI voice calls for collections?
For most invoices, yes. SMS reaches the customer without requiring them to be available at a specific moment, gets read quickly, and generates a higher response rate for less cost. Voice outperforms it only in specific situations, mainly high value invoices, unresponsive accounts, and cases where a real conversation is needed to establish why payment has stalled.
Will AI calls damage customer relationships?
They can if you use them too early or too often. A call on a slightly overdue invoice reads as disproportionate, and customers notice. Used as a late stage escalation on genuinely delinquent accounts, calls are treated as normal collections practice. Most platforms also let you exclude specific customers from calling entirely, which is worth doing for accounts where a partner or executive owns the relationship.
