← Back to blog
LD
Laurent Duplat — Founder, Vocalis AI Published May 19, 2026 · 9 min read · Late payment recovery

It is 3:47 pm on a Wednesday. In a 60-employee SME, the accounting manager opens her client receivables spreadsheet. Thirty-four overdue invoices, eleven of them more than 60 days late. She picks up the phone to call client #1, hits voicemail. Client #2: "call back next week, I'll check with accounts". Client #3: "haven't received your invoice". By 5:30 pm she has made six calls, collected zero payments, and 28 more cases are waiting. Tomorrow she starts again. It is a full-time job that consists of plugging a leak with a bucket.

Late payment recovery is one of the last great forgotten productivity reserves in French SMEs. The tools exist — automated emails, SMS, electronic registered letters — but none reproduce the most effective channel: the human voice. Until the arrival of AI voice agents capable of conducting a respectful, compliant and persuasive dunning conversation at scale. This article breaks down the complete workflow: what works, when, and why switching to voice as early as D+15 tips the recovery rate from 38% to 78%.

The key finding: on 100 invoices unpaid at D+30, an SME that only sends emails recovers on average 38% of the amount. The same SME equipped with a multichannel email + SMS + AI voice call workflow recovers 78%. The difference — 40 points — represents, for a business invoicing 2 million euros/year with 4% unpaid, about 32,000 euros of cash saved every year.

SME unpaid invoices 2026: 12.4 billion euros/year unrecovered

Consolidated figures from the Banque de France and the Observatory of Payment Delays paint a brutal picture: average client payment delays remain at 44 days in France, against a European average of 34 days. SMEs — those with the least reserve cash — are the first victims. Out of 1.2 million SMEs, about 25% report that payment delays have threatened their survival in the past three years.

The hidden structure of unpaid invoices

Not all unpaid invoices are alike. The typical distribution observed on a panel of B2B SMEs is revealing:

This distribution is critical: 90% of unpaid invoices are recoverable amicably, provided you act fast and well. The drama for SMEs is that they spend most of their time on the 10% of hard defaults — out of ignorance, fear, or because internal tools do not allow industrializing simple reminders.

12.4 bn€annual unpaid invoices in French SMEs
44 daverage client payment delay in France (vs 34d EU)
90%of unpaid invoices recoverable amicably

Multichannel dunning workflow (email/SMS/voice/formal notice)

Effective dunning is not a one-off action but an orchestrated sequence combining several channels at precise moments. The golden rule: progressive escalation of channel and tone, without ever breaking the commercial relationship before the last resort.

D+1 — Automated polite email

The day after the due date, an automated email goes to the accounting contact. Neutral tone, short format, clear subject line: "Invoice #2026-1847 — payment reminder". No accusation, just a notification: "Unless we have made a mistake, your payment has not reached us". This email alone recovers 35 to 52% of unpaid invoices according to industry studies. Marginal cost: zero. Yet 40% of SMEs do not send it systematically.

D+8 — Email + SMS reminder

One week later, second contact via dual channel. SMS bypasses email inflation in B2B inboxes. Ultra-short format: "Hello, your invoice of [date] remains pending. Please get back to us: [link]. Best regards, [SME]". SMS has a 98% open rate vs 23% for B2B email.

D+15 — AI voice call

This is where the workflow changes dimension. The AI voice agent makes a targeted call, usually in early afternoon (slot where accounting departments are reachable). It introduces itself, identifies the invoice, listens to the reason for the delay, proposes a solution (immediate payment by link, payment plan, missing supporting document). Average conversation: 2 min 40. Major psychological effect: the client understands their invoice is being seriously tracked, without being aggressed.

D+30 — Automated formal notice

If the three previous steps yielded nothing, automatic triggering of an automatic formal notice by electronic registered letter (eIDAS compliant). Mandatory mention of late payment interest (ECB rate + 10 points) and the 40 euro fixed compensation (art. L441-10 French Commercial Code). This is also the stage where, if necessary, the complete automation of the case is triggered on the litigation side.

Why this sequence works: each step filters out easy payers. At D+1, you recover oversights. At D+8, the distracted late payers. At D+15, the cash flow tensions (who negotiate a plan). At D+30, only the real disputes and defaults remain. Human time is only mobilized on these 10-15% of complex cases — instead of being diluted across 100% of receivables.

The optimal moment to switch to voice (D+15)

Why D+15 and not D+5 or D+30? The question is not trivial. Too early, the call seems aggressive and damages the commercial relationship. Too late, the case has already been buried under other priorities on the client side. Behavioral data from our SME panel indicates an optimal window between D+12 and D+18.

The psychological trigger of the phone

An email can be ignored, archived, forgotten. An SMS can be read and forgotten within a minute. A phone call, even brief, creates a social obligation to respond. The accounting decision-maker who picks up feels committed to providing an explanation. This mechanism — well studied in social psychology — is why the voice call has a conversion rate 4.2 times higher than email alone for equivalent amounts.

The key role of AI in scalability

The historical problem with voice dunning is its cost and time. An accountant can make 30 to 50 dunning calls per day, 60% of which go to voicemail. At 35 euros loaded hourly cost, each useful call costs between 8 and 14 euros in human time. For 200 monthly unpaid invoices, that is between 1,600 and 2,800 euros of hidden cost — often invisible in the P&L because it is diluted in salaries.

The AI voice agent changes this equation. It places 100% of calls (never sick, never on leave), spreads them intelligently across the day, does not get discouraged, stays consistent in tone. More importantly: it frees the accountant for complex cases — where humans really make the difference (plan negotiation, dispute management, key account relationship).

"Before, I spent 60% of my time on basic dunning and 40% on real cases. With the AI agent handling D+8 and D+15 automatically, it is reversed. I now handle disputes and key accounts in depth, and our DSO has dropped from 52 to 38 days in four months."

— Florence M., chief accountant of a 60-employee industrial SME, Lyon

To go deeper on the financial impact, see our analysis Reduce DSO with voice AI.

Respectful and effective AI voice call script

An AI voice agent for late payment recovery is neither an aggressive robot nor a telemarketer. It is a polite accounting assistant following a strict framework — the one that maximizes payments while preserving the customer relationship.

Standard structure of an AI dunning call (D+15)

  1. Identification (10 sec): "Hello, [First Name Last Name] speaking from the accounting department of [issuing SME]. I am calling about invoice #2026-1847 dated [date]. Do you have a few minutes?"
  2. Verification (15 sec): "This invoice, amounting to 4,850 euros, was due on [date]. Can you confirm you have received it?"
  3. Active listening (30-60 sec): depending on the answer, conditional branches (invoice not received → immediate resend by email; under validation → deadline; cash flow issue → payment plan proposal; dispute → transfer to sales manager)
  4. Solution proposal (30 sec): "I can send you a secure payment link immediately by SMS. Or if you prefer, we can agree on a 2 or 3-month payment plan. What do you prefer?"
  5. Confirmation and trace (15 sec): "Perfect, noted. You will receive confirmation by email within the minute. If you have any questions, you can call me back at [number]. Have a good day."

Compliance rules to encode in the agent

To compare this AI approach with traditional approaches, see our analysis Collection agency vs AI and our complete file on debt collection by AI.

Concrete craftsman case: 8 invoices recovered out of 10

Not all examples come from industrial SMEs. A carpenter craftsman established in Côte-d'Or, 4 employees, 680,000 euros annual revenue, deployed the workflow in February 2026. Before intervention: an unpaid rate at D+60 of 9.8% of revenue, or about 66,000 euros of permanent outstanding receivables. His problem was not volume (40 to 50 client invoices per month) but time: between job sites and management, he literally did not have one hour per week to devote to dunning.

The sequence deployed for this craftsman

Custom configuration adapted to his pace:

Results measured over 3 months

On 134 client invoices issued between February and April 2026 of which 10 went unpaid at D+15:

The most unexpected effect for him: sleep quality. No longer having to mentally postpone reminders changes the general mood. It is a benefit rarely quantified but often cited by craftsmen and very small business owners who automate these anxiety-inducing tasks.

Key takeaway: deploying an AI workflow for late payment recovery is not only justified for SMEs of 50+ employees. From a dozen unpaid invoices per month and a business volume above 250,000 euros/year, the investment in time saved + receivables recovered is largely positive. See our 7-day deployment guide for practical terms.

Frequently asked questions about automated late payment recovery

When should the first reminder for an unpaid invoice be triggered?

From D+1 after the contractual due date. A polite automated email the day after the due date, without accusations, raises the spontaneous payment rate from 35 to 52% according to Atradius studies. The longer you wait, the lower the recovery probability: by D+90, it drops below 22%.

Is an AI voice call to follow up with a client legal in France?

Yes, provided three rules are respected: do not hide the artificial nature of the agent if the client asks, respect legal hours 8am-8pm working days, and respect the right to object. The Vocalis AI voice agent records refusals and automatically stops reminders to remain GDPR and Consumer Code compliant. The European AI Act, progressively applicable, reinforces these transparency obligations — better to anticipate them.

What is the difference between dunning and amicable recovery?

Dunning (D+1 to D+30) is a commercial reminder preserving the customer relationship: email, SMS, polite call. Amicable recovery (D+30 to D+90) involves a formal notice with registered letter and mention of legal consequences. Beyond D+90, you move to litigation or order to pay. See Debt collection by AI for phase details.

How much does an unpaid invoice really cost an SME?

Far more than the face amount. For a 5,000 euro unpaid invoice: cash tied up for 60-90 days, 3 to 8 hours of internal accounting time for manual reminders, bank fees if overdrawn, risk of doubtful debt provision. The Observatory of Payment Delays estimates the total hidden cost at 15-25% of the nominal amount of each unpaid invoice. On an SME with 50,000 euros of annual unpaid invoices, that represents between 7,500 and 12,500 euros of invisible losses that never appear in the accounts — unless modeled.