Realisation is a symptom, not a diagnosis. The useful question is not how far it fell but which of four leaks it fell through, because the fixes point in opposite directions.
Key takeaways
Your realisation rate is dropping for one of four reasons, and they need different fixes: the work was priced below what it costs you to deliver; the work delivered was wider than the work quoted; the work was billed so late that the client had stopped associating it with value; or the work was billed properly and never collected. AI is useful here because it can read time narratives, engagement scopes and the receivables ledger together and tell you which of the four you actually have — which is the one thing a monthly realisation percentage on a dashboard never tells you.
That is worth saying plainly because most firms respond to a falling realisation rate by raising standard rates, which fixes only the first leak and makes the third and fourth worse. If the problem is that August work is invoiced in November, a rate rise increases the size of the number you eventually write down.
Start by separating the two events that both get called a write-off. A write-down happens before the bill goes out: recorded value at standard rates is reduced to the amount the client is asked to pay. A write-off happens after: an amount you have already invoiced, already recognised as revenue and already remitted tax on is abandoned. Firms that keep one combined figure cannot diagnose anything, because the two have almost nothing in common operationally.
The mechanism is unglamorous and that is the point. Time narratives are free-text; engagement letters are documents; the ledger is structured. A language model can read the narratives on a written-down job and propose a cause code against the four categories above, quote the sentences it relied on, and leave the code in a review queue. Over a few hundred closed jobs that produces something a spreadsheet cannot: a distribution of causes by job type, by partner and by client.
It can also do the tedious cross-check that nobody does by hand — comparing the deliverables named in the signed engagement letter against the deliverables described in the time narratives, and flagging jobs where the second list is longer than the first. That is scope drift made visible while the work is still open, rather than at the write-down meeting six months later.
What it must not do is set the code and close the loop unattended. A cause code is an accusation about a partner’s pricing or a manager’s scope control, and it should be ratified by a person before it becomes an input to anyone’s year-end conversation.
This is where practices lose real money by filing the two events in the same bucket. Under paragraph 20(1)(p) of the Income Tax Act, a taxpayer may deduct debts owing to it that are established to have become bad in the year and that were included in computing income for that year or a preceding one. Unbilled work that you decided not to invoice was never included in income, so reducing it is not a bad-debt deduction — it simply never became revenue. An invoiced account that goes uncollected is a different animal, and Treadstone Law’s note on writing off bad debts explains how the deduction is approached in an Ontario setting.
The GST/HST side is the one most often left on the table. Section 231 of the Excise Tax Act lets a supplier that has written a bad debt off in its books deduct, in determining net tax, the tax portion of the amount written off — provided, under subsection 231(1.1), that the tax was included in the net tax reported for the period in which it became collectible and that the net tax reported was remitted. Subsection 231(4) puts a limit on how long you have: the deduction must 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. Treadstone Law’s article on HST recovery on bad debts walks through the practical version.
Two practical consequences follow. First, the date you write a debt off in your books is not a bookkeeping formality — it starts a clock. Second, if you later recover part of it, subsection 231(3) requires you to add the corresponding amount back. Both are exactly the kind of thing a machine should be watching for and a person should be signing off.
A four-partner Ontario accounting practice reports realisation on a single monthly line and watches it slide over three quarters. Rates go up in the spring; the line keeps sliding. The firm then codes eighteen months of closed jobs by cause, using AI to propose the code from the narratives and a partner to confirm it.
The distribution is lopsided in a way nobody expected: personal tax work is close to standard, and the slide is concentrated in a single corporate-year-end job type where the engagement letter names three deliverables and the narratives routinely describe six. That is scope drift, not a pricing problem, and the fix is a change-in-scope clause and a manager instructed to invoke it — not a rate card.
The numbers to measure in your own firm are the ones you already hold and can count before and after: the median days between last time entry and invoice date, and the proportion of closed jobs whose narratives describe deliverables absent from the engagement letter. Both are countable. ‘Hours saved’ is not, which is why it makes a poor business case.
Whatever you decide to change, the record has to survive. Section 230 of the Income Tax Act requires every person carrying on business to keep records and books of account in a form that lets the tax payable to be determined, and subsection 230(4)(b) sets a general retention floor of six years from the end of the last taxation year to which the records relate; subsection 230(4.1) adds that records kept electronically must be retained in an electronically readable format. Treadstone Law on corporate record retention is a useful plain-language companion.
If you are a law firm, the same analysis runs into a professional-conduct rule worth reading before you change anything. Under rule 3.6-1 of the Law Society of Ontario’s Rules of Professional Conduct, a lawyer must not charge or accept any amount for a fee or a disbursement unless it is fair and reasonable and has been disclosed in a timely fashion, and rule 3.6-3 requires the statement of account to detail fees and disbursements clearly and separately. A large write-down is often the price a firm pays for having skipped the ‘disclosed in a timely fashion’ part.
If what you are actually trying to fix is your own firm’s month-end close rather than the economics of client work, that is a different job and it lives on the accounting automation page. This article is about a practice delivering to a book of clients.
Should realisation be measured per job or per client?
Both, and the difference is the finding. A client whose individual jobs each realise acceptably but who generates a stream of small unbilled requests will look fine per job and poor per client.
Can AI decide what to write down?
No. It can assemble the file, quote the narratives and propose a category. Reducing a client’s bill is a partner’s decision and it is the partner who signs the account.
Is a write-down evidence the fee was unreasonable?
Not in itself, but it is evidence worth reading. The conduct rules ask whether the fee was fair, reasonable and disclosed in time; a pattern of write-downs on one job type usually means the estimate given at the start was not the estimate the work required.
A 30-minute call is enough to tell you whether AI pays for itself here.
