Can AI actually help with late payments?
AI for accounts receivable helps with the part that is failing, which is almost never knowing who owes you money. Your accounting system already knows that. What fails is the follow up: the reminder that should have gone at day eight and went at day thirty, the second reminder nobody sent because the first one felt awkward, and the customer who was never called because the person who should call them was on a job.
A machine is very good at the consistency and very bad at the judgement. Build to that division and it works. Ignore it and you will annoy good customers.
What is the real cost of chasing badly?
Cash sitting in someone else's account, and a slow erosion of what your payment terms mean.
Terms only exist if they are enforced predictably. A business that chases at day eight every time trains its customers to pay around day eight. A business that chases whenever someone remembers trains its customers that the terms are a suggestion, and the average days to payment drifts out year after year without anyone deciding it should.
The second cost is worse and harder to see. The owner ends up doing it, at night, because nobody else wants to, and that is the most expensive hour in the business spent on the least valuable task in it.
What should a machine do here?
It should do everything up to the point where tone and commercial judgement start to matter, and then stop.
- Watch the ledger. Notice what has aged past terms without anyone running a report.
- Sequence the follow up. First reminder, second reminder, escalation, on the schedule you set rather than the schedule anyone remembers.
- Draft in your voice. A reminder to a fifteen year customer should not read like a reminder to a first time buyer who has already gone quiet.
- Attach the evidence. Invoice, purchase order, delivery docket, so the reply is not "send it again", which is where half of these conversations die.
- Reconcile. Match payments as they land and stop chasing people who have already paid, which is the fastest way to lose a customer over an eleven dollar shortfall.
- Escalate to a person. Flag the accounts where the pattern has changed, because a good payer who has suddenly gone slow is telling you something a reminder will not fix.
What must stay with a person?
A person keeps anything that changes the relationship or the commercial terms, without exception, and that line does not move.
A machine may not decide to put an account on hold, agree a payment plan, apply or waive a fee, threaten escalation, or make the phone call to a customer who is genuinely in trouble. Those are commercial and sometimes human decisions with consequences a reminder email does not have.
The reason is not squeamishness, it is that these decisions are irreversible in the way that matters. You can unsend nothing. That is why our builds put an approval gate at the last reversible step, a principle we set out in human in the loop AI.
Will customers be annoyed by automated reminders?
Only if the reminders are wrong, generic, or arrive after they have paid. Correct and timely reminders are not resented, they are expected.
What people actually object to is being chased for something already settled, being addressed as an account number after a decade of trading, or receiving three identical messages because two systems both think they own the follow up. All three are build failures, not automation failures, and all three are avoidable if the machine reads the same ledger the money lands in.
A useful test: would a competent, polite bookkeeper have sent this message today, in these words, to this customer? If yes, send it. If you hesitate, that is a gate.
How do you know it worked?
Take a baseline before anything changes, then measure the same numbers after.
Three are enough. Average days to payment. The value of receivables past terms at month end. And the hours a person spends on follow up in a week. Write those down before the machine starts, because once it is running you cannot go back and collect them, and without them the review becomes an argument about whether things feel better. That is the failure mode we describe in why AI projects fail.
The counting matters more than the machine. A build that saves four hours a week and shifts payment by six days is easy to justify. One that does neither should be turned off, and you can only know which you have if you took the baseline.
Where it fits with the rest of the admin
Receivables is rarely the only lane. It usually sits next to quoting, job records and reporting, and the value compounds when those talk to each other, because a machine that knows the job was completed and signed off knows the invoice is defensible.
If the back office is where your time is going, the admin disappears is the shape of that work, and Penny is the crew member who keeps the finance picture current. When you want to talk about your own ledger, get in touch.
Before you change anything about your terms or your escalation practice, the ACCC guidance for business and business.gov.au are the right references, and a conversation with your accountant is cheaper than a mistake.