Automating invoice reminders to cut payment time from 24 days
A wholesale packaging supplier chased every overdue invoice by hand from a spreadsheet. We built reminders that run on a schedule and an agent that sorts the replies, with a person approving anything sent to key accounts.
- Client
- A wholesale packaging supplier
- Industry
- B2B distribution, 22 staff
- Engagement
- 5 weeks, then ongoing monitoring
- Published
Note.Client name is anonymized and figures are illustrative until publication is approved.
- 9 days
- Average days to payment, from 24 days
- 61%
- Replies sorted without a person
- 4 hrs
- Owner's chasing time per week, from 11 hours
Business context
A wholesale packaging supplier with 22 staff sells to other businesses on 30-day terms. The owner handled credit control personally, usually on Fridays, from a spreadsheet.
Cash flow depended on those Friday sessions. When the owner was away, invoices went unchased, and the overdue balance grew. The company asked us to automate invoice payment reminders so the process no longer depended on one person.
The problem
Chasing was manual and inconsistent. The spreadsheet listed overdue invoices, and the owner wrote individual emails in Outlook. Some weeks he got through all of them; some weeks he did not.
Replies made it worse. Messages like "paid yesterday" or "wrong PO number on this one" sat in his inbox for days before anyone acted. Good customers who always paid on time sometimes received a reminder the day before their payment cleared.
The owner spent about 11 hours a week on this. Roughly half of that was reading and sorting replies rather than sending anything.
Late payment was not the only cost. The owner could not see which customers were drifting, because the spreadsheet only showed what was already overdue. By the time an invoice appeared there, the relationship was already strained.
What we built
A reminder schedule now runs from the accounting system. Invoices move through polite stages based on how overdue they are, and every email is written in the company's own tone, which we drafted with the owner and the two people who knew the accounts.
An agent reads each reply and sorts it into one of four buckets: paid, dispute, needs a copy of the invoice, or a promise to pay with a date. The buckets appear in a daily summary for one person to review.
We kept a human in the loop for key accounts. Anything above a set balance, and every customer on the named list, waits for a person to click send. We chose that limit because a wrong tone with a major account costs more than a few extra days of payment. This cautious pattern runs through our other AI agent workflows.
The system works only with the accounting records it can read, so invoices with no email address on file still go to the owner. We did not try to guess missing contact details.
Rollout
- Week 1: Reviewed six months of invoices and the owner's actual sent emails.
- Weeks 2 to 3: Built the schedule, the letter templates, and the reply sorting.
- Week 4: Ran alongside the manual process and compared every generated email.
- Week 5: Turned on automatic sending for standard accounts; key accounts stayed manual.
Before and after
| Area | Before | After |
|---|---|---|
| Chasing | Friday spreadsheet session | Scheduled, every business day |
| Replies | Sat in one inbox | Sorted into four buckets daily |
| Key accounts | Same process as everyone | A person approves each email |
| Owner's time | About 11 hours a week | About 4 hours a week |
Results
Average days to payment fell from 24 to 9 over the first quarter. The overdue balance at month end dropped by about a third, which mattered more to cash flow than the speed alone. The agent sorted 61% of replies without a person. The remainder needed judgment, usually a disputed amount or a missing document.
The owner's chasing time fell from about 11 hours a week to about 4, most of it spent on the key accounts and disputes. Invoices unpaid after 60 days fell from an average of 21 to 12.
This is a lighter version of the invoice matching pipeline we built for a freight broker, where the same rule held: keep a person in charge of anything that could damage a relationship.
Lessons learned
- Tone mattered more than timing. Our first draft read like a debt collector, and the owner rewrote it. The reply rate improved after that.
- Automating the schedule was easy. Reading replies was the hard part, and it is where most of the value sits.
- We sent reminders to a few accounts that always pay early, and it annoyed them. The named list now excludes anyone who pays before the due date.
- The agent should never send a final notice. We kept that as a person's job on purpose.
- One template fits most invoices, but a shorter first reminder performed better than the full version. We now lead with the invoice number and the amount.