Collections and Delinquency: How an AI Agent Chases Invoices Without Burning Customers

Collections and delinquency management with AI agents solves a tension that every SME knows: if you don't chase invoices, you don't get paid; if you chase them poorly, you burn the customer. In Spain, the average collection period between companies comfortably exceeds the legal 60 days, and in many SMEs, it stretches to 80 or 90. The problem is rarely due to customer bad faith: it's that following up well requires consistency, judgment, and tact, and the administration team doesn't have time for all three at once. An AI agent does.

This article explains how a collection agent works—what it does on its own, what it escalates to a person, and how it modulates the tone to avoid damaging the relationship—what metrics it moves, and how to set it up. It is part of our AI financial automation guide, where we cover the complete financial cycle.

Why the manual collection process fails

The typical circuit is this: the invoice expires, no one notices until the monthly aging report review, someone sends a generic email ("we remind you that invoice X is pending"), the customer doesn't reply, three weeks pass, another identical email is sent, and finally, the angry manager calls—or worse, no one calls. The failures are structural, not personal:

• Collections start late. Every week of delay in the first reminder reduces the probability of quick payment. With monthly reviews, the first step arrives on average 15-20 days after the due date.
• Everyone is chased the same way. The customer who always pays and just forgot receives the same dry email as the repeat offender. The former is annoyed; the latter is unaffected.
• There is no follow-up. The customer promises to pay "next week" and no one notes the promise or verifies if it was kept.
• No one reconciles in real-time. An invoice is claimed that the customer already paid yesterday, which truly burns the relationship.

How an AI collection agent works

Monitors due dates and acts on day one

The agent reads the invoice portfolio from the ERP every day through native connectors with your systems. It knows what expired yesterday, what expires in 5 days, and what has been pending for 60 days. It can even act before the due date: a friendly reminder 3 days before ("we are re-attaching the invoice, it's due Friday") eliminates the "I didn't receive it" excuse and brings payments forward without any friction.

Segments each customer before writing to them

Here is the difference from classic automated reminders. The agent knows the history: how much that customer bills per year, how many times they've been late, if they have an open dispute, if they usually pay 10 days after a reminder. With this, it decides the strategy: for the distracted strategic customer, a cordial and personal reminder; for the small repeat offender, a firmer sequence with deadlines; for the one with an open dispute, no automatic claims—that goes straight to a human.

Writes like your company, not like a robot

Working as an agent with your company's knowledge—your tone, your payment terms, the context of each account—every message comes out personalized: it references the specific invoice and order, attaches the document, offers a breakdown if there are several overdue, and facilitates payment (link, account number, installment options if your policy allows). The escalation of tone is gradual and designed by you: cordial, firm, formal, and pre-legal action warnings only when you approve them.

Records promises and follows up on them

If the customer replies "I'll pay you on the 25th," the agent records it, waits until the 26th, checks the bank, and acts based on the result: thanks for the payment or resumes the conversation citing the broken promise. This systematic memory is what no overwhelmed team can maintain manually, and it's where much of the collection is won.

Escalates to humans with judgment

The agent shouldn't call your best client on the phone or negotiate discounts. Its job is to resolve 70-80% of the overdue portfolio that only needs consistency, and hand over the remaining 20-30% to the team with the case prepared: debt history, message logs, broken promises, and a recommendation. The difficult call is made by a person, but with all the context in front of them.

What metrics it moves (and by how much)

A collection project is measured by three numbers, and it's worth setting a baseline before starting:

• DSO (Days Sales Outstanding): this is the North Star metric. Reductions of 10 to 20 days are common in portfolios that previously had manual management, and every day of DSO is freed-up cash: in a company billing 5 million, 15 fewer days of DSO is about 200,000 euros back in the treasury.
• Percentage of overdue portfolio: by claiming on day one and with systematic follow-up, the balance overdue by more than 30 days usually drops by half within a quarter.
• Management hours: chasing invoices manually easily consumes 20-40 hours a month in an SME. The agent reduces this to exception management hours.

And there is a fourth, less visible metric: customer complaints for unwarranted claims, which approach zero with daily reconciliation.

Collecting without burning bridges: design rules

The most reasonable concern for any manager is that an automatic system will spoil commercial relationships that took years to build. This is avoided with four design rules:

• Protected accounts list: strategic clients where the agent prepares the message but a person sends it.
• Dispute brake: if there is an open incident or claim, the sequence stops automatically.
• Tone approved by you: you define and approve every template and every level of escalation before activating anything.
• Full traceability: every message sent is recorded and visible to the account manager, who can pause the sequence in one click.

Well-designed, the effect is the opposite of what's feared: customers perceive a professional and predictable administration, and sales reps stop acting as debt collectors, which is probably the task they hate most.

How to get started

The short path: measure the baseline (DSO, overdue, hours), connect the agent to the ERP and the bank, start in draft mode—the agent proposes, a person approves every send for 2-4 weeks—and move to autonomous mode by segments, starting with lower-risk accounts. In 6-8 weeks the system works alone for the majority of the portfolio. This is the same pattern we apply to all AI agents in the finance department: first supervised, then autonomous, always measured.

If you want to know how many days of DSO and how many hours per month you can recover with your real portfolio, request our diagnosis: we analyze your collection process and give you a calculation based on your numbers, not industry averages.

Frequently Asked Questions

Can an AI agent claim invoices without upsetting customers?

Yes, if designed with segmentation and brakes: protected accounts, automatic pause for disputes, and company-approved tone. Most collection complaints don't come from claiming, but from claiming poorly: late, in the wrong tone, or for invoices already paid.

How much can DSO drop with a collection agent?

In portfolios with previous manual management, reductions of 10 to 20 days within 3-6 months are a realistic result. The effect comes from claiming on day one, following every payment promise, and not letting any invoice fall into oblivion.

What does the agent do when a customer doesn't respond to anything?

It executes the defined escalation—reminder, firm message, formal notice—and, once the sequence is exhausted, hands the case to the team with the full history and a recommendation. Serious decisions, such as starting legal action or cutting supply, are always made by a person.

Does it integrate with my current ERP and bank?

Yes. The agent reads invoices and due dates from the ERP and bank movements through native connectors with your systems, so it reconciles payments daily and never claims an already collected invoice. No software change is necessary.