Payroll is due. One timesheet is missing, another disagrees with the supervisor's record, and two entries look almost identical. Across the desk, an invoice waits for approval while a customer says they have already paid the last one.
Most of that work is finding out what happened before anyone can finish the numbers. AI can help here. It can collect the details, compare documents, draft explanations and put open questions beside their sources. The person who knows the business still decides what gets posted, approved and paid, and calculations still need a reliable method.
Eight places it could help
| Work | Information it needs | Useful preparation | Decision that stays with a person |
|---|---|---|---|
| Documents and inbox | Bills, receipts and correspondence | A sorted list with missing details and links to the originals | Where each item belongs and what needs attention |
| Invoice preparation | Agreed scope, rates and approved work | Draft line items with their sources and open questions | Whether to invoice, and the accounting and tax treatment |
| Receivables follow-up | A dated ledger export and recent messages | Questions to resolve and possible reminder drafts | Payment status, dispute response and what to send |
| Expense preparation | Receipts and existing categories | Suggested categories, with uncertain items set apart | Correct category and treatment |
| Hours checks | Submitted time and approved actual-work records | Differences, missing submissions and possible duplicates | Correct hours and payroll submission |
| Reconciliation support | Statements and accounting exports | Possible discrepancies tied to source rows | Accepted matches and the finished reconciliation |
| Monthly reports | Checked reports and operating notes | A comparison, possible explanations and open questions | What the changes mean and what to do |
| Purchasing | Requests and approved supplier terms | A comparison with missing details shown | Supplier choice, commitment and payment |
Possible uses to test. What each AI tool can reach varies by product and plan.
Before payroll: a list of questions
Made-up example. The business, records and figures are invented to show the method. They are not a client result or output from a real AI run.
An office manager has a submitted-hours file and a supervisor-approved log of actual work for the same week. Each row carries an employee ID, a day and a source reference. The files contain five situations:
| Record | Submitted hours | Actual-work log | Prepared finding | Question for the reviewer |
|---|---|---|---|---|
| A / Monday | 6.5 | 6.5 | The entries agree | Check the matching source references |
| B / Tuesday | 5.0 | 4.5 | Difference of 0.5 hours | Which entry reflects the work done? |
| C / Wednesday | Missing | 3.0 | No submission supplied | Get the missing submission |
| D / Thursday | 4.0 | Missing | No supporting log supplied | Get the actual-work record |
| E / Friday | Two entries of 4.0 | One entry of 4.0 | Possible duplicate or incomplete log | One period of work or two? |
The half-hour difference on row B is arithmetic, and a checked spreadsheet formula can calculate it. AI helps by preparing the question and pointing to both records.
Row C is the trap. Treating the missing submission as zero hours changes its meaning. Filling it from the roster swaps time scheduled for time worked. The reviewer needs the actual record first. Row E is similar: two matching entries may be an error or two pieces of poorly recorded work, and deleting one would hide the question.
Why a convincing total can mislead
The submissions add to 23.5 hours, counting both E entries and leaving C open. The available actual-work entries add to 18.0 hours, with D's log missing. A summary reporting "5.5 excess hours" treats two incomplete sets as complete. The useful output lists the differences and missing records separately and gives no approved total until the reviewer settles them.
Before invoicing: four items, four next steps
Made-up example. The business, records and figures are invented to show the method. They are not a client result or output from a real AI run.
A service firm has an approved scope and rate card, completion approvals, a receivables export saved Monday afternoon and newer customer messages. One instruction to "prepare invoices and chase anything overdue" would hide four different decisions.
| Item | What the supplied information says | Prepared next step |
|---|---|---|
| A | Work is approved and the agreed rates are available | Prepare line items for review; check it was never invoiced |
| B | Monday's export shows an unpaid invoice; a Tuesday message says payment was sent | Confirm receipt and allocation before any reminder |
| C | The customer disputes part of the scope | Give the owner the disputed item and the agreement |
| D | Completion approval is missing from the files | Ask for the approval; treat the work status as open |
For item A, the approved scope covers two planning sessions at CAD 175 each and six preparation hours at CAD 95. The draft shows its arithmetic:
| Approved work | Calculation | Draft amount, CAD |
|---|---|---|
| Planning sessions | 2 × 175 | 350 |
| Preparation | 6 × 95 | 570 |
| Subtotal before GST/HST | 350 + 570 | 920 |
Before issuing it, the reviewer checks the scope, approval, prior billing and GST/HST treatment.
Item B is where a draft is most likely to fill a gap. Monday's export and Tuesday's message belong side by side. A draft that marks the invoice paid because of that message has turned a promise of payment into received cash. The next step is to confirm receipt and allocation. An automatic reminder would ignore what the customer already said.
Once an item is confirmed unpaid and undisputed, AI can draft the reminder. The person responsible decides when to send it.
When your accounting software already does it
Your accounting system may already handle much of this. QuickBooks Online, for example, suggests matches for bank transactions and asks you to review and post them. Its reconciliation guide describes comparing your transaction records against your bank or credit card statement. An AI explanation of a difference is no substitute for that reconciliation.
When a task follows clear rules, a good invoice template, timesheet form or checked formula may save more effort than an assistant.
AI is more useful where the preparation crosses a message, a document and a set of figures. Fluent wording is still no accuracy check. Microsoft's Copilot in Excel FAQ asks you to review, edit and verify anything Copilot creates before you rely on it, and to avoid using it for decisions in sensitive areas such as finance. Tax positions, accounting adjustments, payment instructions and payroll approval stay with your accountable people and systems.
Make the preparation easy to find and finish
A first test can use a few selected exports. Keep the period, the update date of each file and the source references with the output, so checking does not become another search. Use only information approved for the chosen tool. Canadian privacy authorities' generative AI principles ask for legal authority and appropriate purposes when personal information is collected and used; AI and sensitive information covers what to keep out.
Product and guidance details checked November 2026.
Start with a period you can check
Pick a completed period with a reviewed final result. Prepare the questions list or invoice draft from its inputs and compare it with what you know happened. Did every source row survive? Were missing records visible? Did the draft invent a match, miss a duplicate or treat a payment message as received cash? Count preparation, review and correction time together; how to tell whether AI is actually helping sets out the full comparison.
If payroll prep or invoicing keeps eating your week, book a free 30-minute call. We'll look at where the time goes and what could be prepared before you review it.


