Chronic late payment is a pattern, not an event — and patterns are invisible in a system that files everything by invoice.
Key takeaways
Your accounts receivable ledger already holds the answer; what it lacks is a stable definition of ‘late’ and a way to see the pattern per client rather than per invoice. Compute days from invoice date to cash received for every settled invoice, weight by value so a $200 invoice does not outvote a $20,000 one, and aggregate by client and by year. AI is useful for the next step: clustering the resulting shapes so that chronic-late clients separate from clients who were late twice during one bad quarter, and reading the correspondence to identify the invoices that were late because they were disputed.
That third category is the one that makes or breaks the exercise. A client who withholds payment because your invoice was wrong is not a credit risk; treating them as one produces a report that a partner will correctly ignore.
With that in place the segmentation is straightforward, and it is worth doing before you buy anything: chronic and large, chronic and small, occasional, and disputed. Each of the four leads somewhere different — a deposit or a retainer, tighter terms, no change at all, and a conversation about your own invoicing respectively.
The first is the limitation period. Treadstone Law explains the basic Ontario limitation period and a companion answer applies it to breach of contract. A client whose habit is to pay at nine months is a client whose oldest unpaid invoice deserves a date in a diary, not a note in a file.
The second is a tax deadline most practices miss entirely. Section 231 of the Excise Tax Act allows a supplier that has written off a bad debt in its books to deduct the tax portion in determining net tax, subject to the conditions in subsection 231(1.1); subsection 231(4) requires the deduction to be claimed in a return filed within four years after the day the return was due for the reporting period in which the debt was written off, and subsection 231(3) requires an add-back if the debt is later recovered. Treadstone Law’s article on HST recovery on bad debts explains it in practice, and a related answer covers the income-tax side.
Put together, the two clocks turn a late-payment report from an interesting chart into a work list with dates on it, which is the only version that changes anything.
Mostly it does not, and knowing why is worth two minutes. The Office of the Privacy Commissioner notes that business contact information — an employee’s name, title, business address, telephone number or email address — collected, used or disclosed solely for the purpose of communicating with that person in relation to their employment or profession is not covered by the Act. Analysing when your corporate client’s accounts-payable contact answers emails sits outside PIPEDA on that basis.
It changes where your client is an individual or a sole proprietor. Then the payment history is personal information about a person, and the ten fair information principles in Schedule 1 of the Act apply — including identifying purposes, and limiting use, disclosure and retention. The OPC also notes that Alberta, British Columbia and Quebec have their own private-sector privacy laws. Treadstone Law on whether PIPEDA applies to your business, what counts as personal information and customer data collection cover the practical questions for a business.
The practical rule that follows: if you would be uncomfortable telling a client you keep a behavioural score on them, do not keep one. Keep the facts — days to pay, disputes raised, plans agreed — and let a person draw the conclusion.
Canada’s voluntary code of conduct on advanced generative AI asks organisations to commit to outcomes including accountability, transparency, and human oversight and monitoring of systems after deployment. It also states plainly that the code does not change existing legal obligations, giving PIPEDA as an example. For a segmentation model that decides who gets chased and who gets a deposit demand, the operative commitment is the oversight one: someone should be reading what the model produced, on a schedule, and be able to explain a client’s ranking.
A Nova Scotia practice with about ninety active clients has a general sense that ‘everyone pays late’. Weighted days-to-pay by client tells a different story: the median is comfortable, and two clients account for most of the money-days. One is a large institution whose payment run is monthly and predictable, which is a terms problem solved by invoicing four days earlier. The other is genuinely slow and had two disputed invoices in the period, both about scope.
So the report produces three actions, none of which is a policy change: move one client’s invoice date, take a deposit from another, and rewrite the scope paragraph in one engagement letter template. The measures kept afterwards are weighted days-to-pay by client, and the number of invoices disputed on scope — because the second is the one the firm can actually fix.
Should we fire chronic late payers?
Sometimes, and that is an owner’s decision made with the margin in front of them. A client who pays at 120 days at full rate may be worth more than one who pays at 30 at a discount.
Can the model set credit terms automatically?
It can propose them and show the evidence. Changing a client’s terms is a commercial decision and a communication, and both need a person.
What about the firm’s own payables?
Different job, different page. Your own supplier ledger belongs on the accounting automation page; this article is about a book of clients.
A 30-minute call is enough to tell you whether AI pays for itself here.
