AI is useful at three points in a firm’s billing run: assembling the draft bill from work in progress, drafting the narrative that explains it, and chasing what goes unpaid. What it must not touch is the statutory content of the invoice itself, because your client’s ability to recover the tax you charged depends on it.
Key takeaways
At a practice billing a book of clients, the monthly run has four stages. WIP assembly pulls together time, disbursements and fixed-fee milestones for each client. Review is where a partner decides what actually gets billed. Production turns the approved amounts into a compliant invoice. Collection is everything that happens after it is sent.
AI is genuinely useful in assembly, in the narrative part of production, and in collection. It is not useful in review, because review is the professional judgment that the whole exercise exists to protect. And it must not be trusted to determine tax treatment, because that is a legal characterisation of the supply, not a formatting question.
Note the boundary with a firm’s own bookkeeping. Invoicing your clients as a practice is a delivery process; running your own ledger is a different job, covered on our AI in accounting operations page.
If you are a GST/HST registrant making a taxable supply that is not zero-rated, section 223 of the Excise Tax Act requires you to indicate to the recipient either the consideration and the tax payable in a manner that clearly indicates the amount of the tax, or that the amount charged includes the tax. If you show a rate, you must also show the total tax or the total of the rates.
Your client’s side of that is stricter, and it is the reason a badly built invoice template creates real friction. The Input Tax Credit Information (GST/HST) Regulations prescribe what a registrant must obtain before claiming an input tax credit, in three tiers by total amount: under $100, the supplier’s name, the invoice date and the total; $100 or more and under $500 adds the supplier’s registration number and the tax amount or rate; $500 or more adds the recipient’s name, the terms of payment, and a description of each supply sufficient to identify it.
That last requirement is the one that catches professional firms. "Professional services rendered" is not a description of each supply sufficient to identify it, and on a bill of any size it is your client’s problem before it is yours. Practical guidance on the registration-number side is set out in whether your HST number has to appear on your invoices.
Keep the supporting records. Section 286 of the Excise Tax Act requires records sufficient to determine your liabilities, kept in Canada in English or French, retained for six years after the end of the year to which they relate, and retained in an electronically readable format if you keep them electronically.
WIP assembly. Pulling time entries, disbursements and milestones into a proposed bill per client, flagging matters with WIP but no billing activity, and spotting entries that sit on the wrong client. This is matching work, and it is the least judgment-loaded part of the run.
Narrative drafting. Turning fifty time entries into three paragraphs a client will read, in your firm’s voice, against the engagement letter’s description of the work. Drafted by the model, edited by the person who signs.
Collection. Xero’s Canadian invoicing product describes scheduled payment reminders, notification when an invoice is viewed or paid and a view of unpaid invoices, with GST and HST applied automatically to the invoice. Sequencing reminders, drafting the escalation email and grouping debtors by age are all safely automatable; deciding to escalate is not.
Query handling. When a client disputes a line, retrieving the underlying entries and drafting the explanation is fast and mechanical. Whether to hold the line or write it off is a partner decision.
An illustration, not a case study. Suppose a four-partner practice bills roughly 180 clients monthly. Assembly took a manager three days, most of it chasing partners for narratives, and invoices left on the 14th. Two recurring errors cost real money: compilation clients whose engagement letters had been superseded were billed at old rates, and out-of-pocket disbursements sat in a spreadsheet that was reconciled quarterly.
They changed three things. Draft bills are now assembled automatically on the first business day with a narrative already drafted from the time entries. Partners review in a queue rather than by email. Disbursements are matched to clients at the point of payment rather than at quarter-end.
The invoice template itself was rebuilt once, by hand, so that every bill carries the firm’s registration number, a per-supply description and clearly stated tax. That part was never automated, and should not be — it is the piece of the document that has to be right every time.
For accounting practices, the type of engagement drives both the narrative and the scope. the Auditing and Assurance Standards Board distinguishes audit, review and compilation engagements, and each carries different responsibilities that the engagement letter states and the bill should reflect.
A model drafting a narrative does not know which engagement it is describing unless you tell it. Feed it the engagement letter, not just the time entries, and the descriptions stop drifting into language that overstates what you did.
If your bills routinely need explaining, the problem may be upstream of the invoice entirely — see drafting a fee narrative clients accept.
Can AI decide the GST/HST treatment of a supply?
No. Whether a supply is taxable, zero-rated or exempt is a legal characterisation. Configure it once per service line and have a person confirm it.
Can it send the invoice without a human?
Technically yes. In a professional practice it should not, because the invoice is the firm’s statement of what it did and what it is owed.
How long do we keep the backup?
Six years after the end of the year to which the records relate, under section 286 of the Excise Tax Act, and in an electronically readable format if you hold them electronically.
A 30-minute call is enough to tell you whether AI pays for itself here.
