Skip to main content
Automation•8 min read

Dunning: Chasing Invoices Without Losing the Customer

November 8, 2026•By Ajan Kanagalingam

An invoice is sixty days late. Nobody has chased it, because chasing feels rude and everyone is busy. When somebody finally does, the answer comes back in four minutes: we never received it, send it to accounts payable rather than to the person who signed the contract.

Separate the causes before the tone

Keep the ageing report. Age, balance and customer history all matter for prioritising. What ageing cannot tell you is why payment is blocked, and that is the column most systems do not have. Sixty days late because the invoice went to the wrong address is a different problem from sixty days late because the customer is short of cash.

The first needs a corrected invoice and an apology. The second needs a conversation about a payment plan. The third, genuine refusal over a disputed deliverable, needs the person who owns the relationship and possibly a look at what the contract says.

One sequence sent to all three gets the easy money late and the hard money never.

Why it is lateWhat actually fixes itSafe to automate
Never reached the right personResend, to accounts payableYes
Missing a reference their system needsReissue with the PO numberYes
Sitting in an approval queueA nudge to the named approverYes, once
Customer short of cashA call and a payment scheduleNo
Disputing the workThe account owner, and the contractRoute it, and pause reminders on that invoice
Cause unknownAsk what is preventing paymentAsk, yes. Guess, no

Stopping the sequence on a disputed invoice is a rule that routinely gets missed. A customer who has raised a genuine query and then receives three escalating reminders learns that your systems are not listening, and that is a harder thing to repair than a late payment.

Start before the due date

One useful message arrives before anything is late. A short note a few days ahead, confirming the amount and the date and asking whether anything is needed to process it, catches the administrative failures while they are still free to fix.

It also changes the relationship of the whole sequence. You are not chasing, you are confirming, and a customer who has already acknowledged the invoice has far less room to be surprised by it later.

After that, four touches is usually enough: the day it becomes overdue, about a week later with a specific date requested, a call from a person, and then whatever your escalation actually is. Every message should contain the invoice number, the amount, the due date and a way to pay in one click, because a reminder that requires the customer to go and find the invoice is a reminder that gets postponed.

Where AI earns its place

The drafting, first. Writing four well-judged messages per customer segment is a half-day job that keeps getting deferred, which is how a business ends up sending one bluntly worded template for years. Generating a consistent set, in a tone you actually approve, removes the reason that never happened.

Then the sorting. Grouping an overdue list by likely cause rather than by age is pattern work: which invoices lack a PO reference, which customers have a history of needing a resend, which accounts have an open support ticket that might be a dispute. The signals are already in your systems and nobody has time to cross-reference them, which is the same gap described in getting your receivables under control.

One category has to exist for any of this to be safe, and it is the one nobody builds: cause unknown. An email saying we are looking into it is not evidence of a cash flow problem, and a support ticket is not evidence of an invoice dispute. Without an explicit unknown, a suggested cause becomes a recorded fact, and the next reminder is sent on the strength of a guess.

That points at the division that actually matters here, which is not that AI writes and people chase. AI proposes a likely blocker from the evidence it can see and shows that evidence. Workflow rules decide what may be sent, checking recorded payments, agreed payment dates and open disputes first. A named person owns the exceptions. Approved templates bound the tone.

Payment-behaviour patterns are the quietly valuable one. A customer who reliably pays after the third reminder is a pattern that belongs in your cash forecast, and a pattern that is getting worse is worth attention on its own. The two uses are not alternatives, which is the point of a cash flow forecast that is worth trusting.

What to keep off the automation

Anything that threatens. Legal action, interest charges, suspension of service and referral to a collections agency all carry commercial and sometimes legal weight, and they need a person who knows what the account is worth and what the contract permits.

The failure mode here is specific and worth naming. An automated sequence that threatens a client of eight years over an invoice their finance team had already queried costs more than the invoice. Build the stop condition at invoice level rather than account level. Pause generic reminders on a disputed invoice and route it to a named owner, instead of freezing an entire account because somebody opened an unrelated support ticket. Test that rule before the sequence goes live.

Tone also needs a human read. Ask for a firm reminder without defining the boundaries and you have left the tone to the model, with firmness as the only instruction it was given. Read the final message as the customer, not as the sender, which is the discipline in keeping customers while fixing processes.

The upstream fix

Before improving the chasing, count how many of your overdue invoices had something wrong with them. Missing reference, wrong recipient, unclear line items, terms that were never agreed in writing.

That number is uncomfortable for a lot of businesses, and it means part of the collections problem is an invoicing problem wearing a disguise. Measuring it first is the gap analysis that tells you whether to build a sequence or fix a template.

Frequently Asked Questions

What is dunning?

The process of chasing an invoice that has passed its due date, usually as a sequence of reminders that escalate in tone and seriousness. The word sounds archaic and the process is the difference between an invoice getting paid in forty days and getting paid in ninety, which for a small business is the difference between comfortable and borrowing.

Why do invoices go overdue?

For several reasons at once, and the mix matters more than any single cause. In UK Department for Business and Trade research from 2024, based on 300 telephone interviews, 36% of surveyed businesses attributed paying suppliers late to administrative errors, 31% to disputed invoices and 23% to technical issues including invoices getting lost, while 20% overall pointed to their own customers paying late. Those are shares of businesses rather than of invoices, respondents could give more than one reason, and it is UK data. The useful takeaway is that disputes are close behind admin failures, so a process built as though lateness means reluctance to pay will mishandle a large share of cases.

How many reminders should a sequence have?

Four is usually enough, spread so the first arrives before the due date rather than after. A pre-due nudge, a short note on the day it becomes overdue, a firmer one about a week later naming a date, and then a call from a person. Each should ask for something specific, and at least one should ask whether there is a problem rather than just repeating the amount.

Where does AI genuinely help?

Drafting the sequence to a consistent tone, personalising each message with the actual invoice details so it is easy to act on, grouping overdue accounts by likely cause rather than just by age, and flagging patterns a person would miss, such as one customer who always pays after the third reminder. The repetitive drafting is the part that stops happening when everyone is busy.

What should AI never do here?

Send anything that threatens, mentions legal action, interest charges or stopping service, and any generic reminder about an invoice that is under dispute. Those messages have commercial and sometimes legal consequences, and they need a person who knows the relationship. An automated sequence that threatens a long-standing client over an invoice their own finance team queried is an expensive thing to automate.

How do you know the process is working?

Days sales outstanding, and the proportion of invoices paid without any reminder at all. The second number is the more useful one, because it measures whether your invoicing is clear rather than whether your chasing is persistent. A business that improves its invoice accuracy usually finds a chunk of its collections problem disappears.

Get paid sooner without souring the relationship

We build dunning sequences that sort by cause, stop on disputes, and leave the difficult conversations to a person.

Related Articles

Automation

Warehouse Automation Without Buying Robots

September 12, 2026Read more →
Automation

How to Automate GST/HST on Canadian Invoices With AI

February 22, 2026Read more →
Automation

Work Orders: The Information Gap That Costs You

November 4, 2026Read more →
AK
Ajan Kanagalingam
Founder & ChatGPT Consultant, ChatGPT.ca

Ajan leads the ChatGPT.ca team: 200+ custom GPT builds and automation projects for 50+ businesses across 20+ industries. Based in Markham, Ontario. PIPEDA-compliant solutions.

Stay ahead of AI in Canada

Weekly case studies, new tools, and ROI playbooks for Canadian SMEs. One email, zero spam.