AI for Bookkeepers in Canada: 5 Client Jobs to Hand to Claude, and 3 to Keep

Month-end for one business is a chore. Month-end for 55 of them, all landing in the same ten days, is a practice. If you run a small bookkeeping firm, you already know where the hours go: not into the ledger, which QuickBooks Online and Xero already categorize from the bank feed, but into everything around it. The questions email to the client who bought lumber on the company card. The memo to the CPA. The GST/HST worksheet that has to be right before anyone signs.

AI for bookkeepers is mostly about that surrounding work. This guide covers five client jobs a small Canadian practice can hand to Claude this month, the three that stay with a human, and the settings to change before a single client file goes anywhere near an AI tool.

What AI for bookkeepers is good at

Your software already does the pattern-matching. Bank rules, receipt capture through Dext or Hubdoc, category suggestions on the feed: that layer is what most "AI bookkeeping software" pitches are selling, and you own it already. Claude and ChatGPT don't compete with it. They do the work a junior would do if you had one. Read a messy export and say what's odd. Turn a list of uncertain items into a polite client email. Assemble a year-end package with a cover memo. Write the close procedure down so the next hire follows it.

The catch is memory. A fresh chat doesn't know client A's chart of accounts, or that client B's $210 charge lands every month and is fine. You re-explain, it re-guesses, and the time you saved leaks back out. A skill fixes that: one file holding the client's categories, rulings and regular vendors, so the assistant starts from month twelve instead of month one. Our guide to Claude skills for small business covers what a skill is and how to install one on any plan, Free included. The five jobs below use the ones a practice needs.

Five client jobs to hand to Claude this month

1. The monthly categorization pass

Say you run a three-person practice in Windsor and a landscaping client drops a quarter of card statements on you at once. Install the Bookkeeping Copilot skill, keep one Claude Project per client so their rulings don't bleed into each other, and paste the client's chart of accounts and regular vendors into the setup interview. Then paste the export.

It categorizes what it recognizes, marks the rest UNSURE with a reason, and names the anomalies: a software charge that appears on two consecutive days, fuel up 22% on the same job count, a subscription that crept from $49 to $79. Its rules forbid inventing a number, and every response ends by telling the reader to have their accountant verify anything important. In a practice, that closing line is a feature. Your client sees it too.

2. The questions email

The UNSURE list is the email you were going to write anyway. Paste it back and ask for a client-facing note: one line per item, amount and date inline, phrased so the client can answer yes or no from their phone at a job site.

For a one-off cleanup with no setup, the Expense Categorizer prompt does the sort on its own. Paste two or three months of lines at once, name the categories your practice uses, and get back a categorized table with duplicates and personal-looking spend flagged rather than buried. Three months beats one, because creeping subscriptions only show up across a longer window.

3. GST/HST period prep

Sales tax is where a practice most wants help and where AI most needs a leash. The GST/HST Filing Prep prompt has the leash written in: it organizes, totals and flags, and it marks everything "for accountant review". It won't decide whether an input tax credit is eligible and it won't produce a final figure for the return.

Run it two weeks before the period closes. The flag list surfaces the missing receipts and the one mystery transaction while there's still time to chase them, and the reconciliation against what the software says catches the mismatch you'd otherwise find at filing.

4. The year-end handoff to the CPA

Most practices hand a client's year to an accountant, and the thread then runs for weeks: can you resend the aging, what's the $340 in March. The Accountant Handoff Pack skill knows the client's province and software, and builds the package before anyone has to ask. A checklist of what's gathered and what's missing. The categorized export with the UNSURE items called out. A sweep for duplicates and missing receipts. A ranked list of questions worth professional time, and a one-page cover memo in the accountant's terms.

Run it a few days before the handoff date rather than on it. The gap is where the missing receipt gets found and the duplicate subscription gets cancelled, so what goes out is a package the accountant can process in one sitting.

5. The close itself, as a routine

If you want the same close on every client file, the Monthly Finance Close workflow is four steps in a fixed order: confirm the exports are complete before anything gets analyzed, categorize and flag, produce a plain-English review the owner can read, then package for the accountant. Save each month's review and paste it in as next month's baseline; the anomaly check gets sharper with history.

For a practice this doubles as the SOP. A new junior can run the four steps on day one and produce something you'd recognize.

Using Claude for bookkeeping: the practice setup

Claude runs custom skills on every plan now, including Free, so trying this costs nothing. Four settings matter more than the plan does.

• One Project per client. A Project holds the client's files and instructions, and the skill loads inside it. The Free plan allows five Projects, so real client volume lands you on Pro.
• Turn off model training. On the Free, Pro and Max plans, Settings > Privacy has a toggle called Help Improve our AI models. It's on unless you switch it off. With it on, Anthropic can train on your chats and keep them for up to five years; off, retention drops to 30 days. ChatGPT's equivalent is Improve the model for everyone under Settings > Data controls. If you'd rather work under commercial terms that rule out training entirely, that's what the Team plan is for.
• Strip identifiers before pasting. Account numbers, SINs, and the names on payroll files. The skill needs dates, descriptions and amounts; it doesn't need to know whose they are. Payroll records and a sole proprietor's books are personal information under PIPEDA, and you're the one accountable for where they go.
• Keep the CPA's chart. In the setup interview, paste the accountant's category list instead of letting the skill invent one. Books categorized the way the accountant thinks save a correction round at tax time.

Three jobs that stay with a human

• Anything that decides what gets filed. Input tax credit eligibility, whether an expense is deductible, how to treat a shareholder loan. The tools above organize and flag, then stop. That stopping point is what keeps your engagement letter honest.
• Payroll remittances. Source deductions come from the CRA's payroll calculator or your payroll software, never from a chat model. A confident wrong number here turns into a penalty with the client's name on it.
• The ledger of record. QuickBooks or Xero stays the book. AI output is a draft you post from, and the totals get checked before they're posted. Assistants still make things up, and a skill's rules reduce that without removing it.

One more, which isn't a job: banking credentials go into nothing. The export is enough.

Where to start

AI for bookkeepers works best one client at a time. Pick the one whose month-end you dread most, install the Bookkeeping Copilot, and run one close with it before you change anything else. The rest of the finance items sit in the finance category, and the Canadian tool-by-tool comparison is in AI finance tools for small business in Canada. For a scored map of where AI pays off across the whole practice, the free AI Opportunity Audit covers seven areas of the business in one sitting.