Playbook · invoice data entry automation ontario
Nobody in your office should be retyping a supplier invoice
What it costs to extract supplier invoices into QuickBooks or Xero, the CRA rules deciding where records live, and the validation step nobody skips twice.
Lasse Pettersen
Accounts payable in a small business is one person, a folder of PDFs, and an hour on a Thursday. The work is not difficult. It is a person acting as an interface between an email attachment and an accounting system, which is the most expensive way to move data ever invented and the only one that gets tired.
This is the automation with the clearest arithmetic on this site, because both sides of it are already measured. You know how many supplier invoices arrive. You know roughly what the person entering them costs an hour. Multiply, and the decision usually makes itself in either direction.
Count it first
Three numbers, and they take one evening to get:
- Invoices per month. Count them in the inbox rather than estimating. The estimate is always low.
- Minutes per invoice, including the chasing. Opening the attachment, reading it, keying it, coding the line items, filing it, and the separate five minutes spent later finding it again.
- The loaded hourly cost of whoever does it. Wage plus the employer’s share, or the bookkeeper’s rate if the work is outsourced.
A firm processing 120 invoices a month at 12 minutes each is spending 24 hours a month on data entry. At $35 an hour loaded that is $840 a month and $10,080 a year. A build at the low end of the range pays back inside a year and keeps paying.
Below about 40 invoices a month the arithmetic usually does not work, and the honest answer is to say so. That case appears in which jobs to automate first alongside the other builds that look attractive and are not.
The pipeline, step by step
One: the trigger
A watched label in the mail system, a watched folder, or both. Only attachments get processed, and the original message is kept.
The step everybody skips is deduplication at the front. The same invoice arrives twice more often than anyone expects, once from the supplier and once forwarded internally by a colleague being helpful.
Two: extraction
The document goes to a vision model, which returns structured data: supplier, invoice number, date, due date, currency, subtotal, tax, total, and the line items.
The rule that makes this safe is one instruction: return null rather than guess. A field the model cannot read confidently comes back empty and goes to a human. A model asked to always produce a number will always produce a number, and it will be wrong on exactly the invoices that matter, which are the unusual ones.
Three: validation, before anything is posted
Nothing reaches the accounting system without passing every one of these:
| Check | What it catches |
|---|---|
| Line items sum to the subtotal | A misread digit, the most common extraction error |
| Subtotal plus tax equals the total | A tax line read from the wrong row |
| Invoice number not already posted for that supplier | The duplicate payment, which is the expensive one |
| Supplier matches a known vendor record | A changed bank detail on a lookalike invoice |
| GST or HST registration number present where required | A missing input tax credit claim |
| Total inside the normal range for that supplier | The decimal point in the wrong place |
Anything that fails goes to a person with the original document attached and the extracted data beside it, so the review takes 20 seconds rather than reopening the whole task.
Four: the entry
The bill or expense is created in the accounting system with line items coded to accounts, the original PDF attached to the entry, and a link back to the source message.
Five: the confirmation
A short message per successful entry and a loud one per failure. Monthly, the run counts are reviewed: how many processed, how many needed a human, and which suppliers cause it. That last list is where the next improvement always is, because the failures are never spread evenly.
The Canadian rules that shape this build
This is where a workflow template written for another market quietly becomes wrong.
Documentary requirements for input tax credits. To claim a GST or HST input tax credit, the supporting document has to carry specific information, and the requirement escalates with the value of the supply: the lightest tier under $30, more required from $30, and more again at $150 and over, including the supplier’s registration number and, at the top tier, the recipient’s name and the terms of payment. An extraction pipeline that drops the registration number has not just lost a field. It has broken the claim the invoice exists to support.
Where the records live. The Income Tax Act requires books and records to be kept in Canada unless the Canada Revenue Agency gives written permission otherwise, and electronic records must remain in an electronically readable format for the retention period. That turns storage location into a build constraint. It is one of the reasons the cheapest available tool is sometimes not an eligible tool, and the same logic that governs personal information in PIPEDA and your AI tools applies here for a different statute.
Retention. Six years from the end of the last tax year the records relate to, as a general rule. A pipeline that reads a PDF and discards it has converted a filing job into a compliance problem. Every build here stores the original next to the posted entry.
None of this is tax advice. Confirm your own position with your accountant or with the CRA before it decides your architecture.
What it costs
$3,000 to $6,000 to build, inside the $2,500 to $12,000 range published on the workflow automation page. Running costs of roughly $30 to $90 a month at small business volumes, most of it the vision model calls.
| Driver | Cheap end | Expensive end |
|---|---|---|
| Supplier formats | A dozen suppliers, consistent layouts | Long tail of one-off suppliers and scanned paper |
| Line item coding | One account per supplier | Line by line coding across a real chart of accounts |
| Approval | Post and notify | Threshold based approval with a second signer above a value |
| Accounting system | A modern cloud system with an API | A desktop system, or one with no write access |
| Purchase orders | None | Three way matching against orders and receipts |
The last row is a different project. Three way matching is worth doing and it is not what this build is.
What it must not do
- Never post an entry that failed validation. The whole value of the pipeline is that the accounting system stays trustworthy. One bad automated entry costs more confidence than a hundred good ones earn.
- Never let the model decide the account code on an invoice it has not seen before. New supplier, human codes it once, the mapping is stored, and every later invoice from that supplier follows the mapping.
- Never pay on the strength of the extraction. Extraction creates the bill. Payment stays a human decision with its own approval, because invoice fraud in this country arrives as a plausible PDF with changed bank details.
- Never discard the original. See the retention rule above.
Where this sits against the rest
This is the back office half of the same argument the front office half makes. If the problem is that enquiries go unanswered rather than that invoices go unentered, build lead response automation first, because revenue beats labour in a small firm. If the diary is the leak, the build is appointment reminder automation.
There is a related but different job in a trades business, which is getting your own invoices out to your customers faster after the work is finished. That one is the fourth section of the four automations that pay in a trades business, and it changes when you get paid rather than what it costs you to record what you owe.
Statistics Canada put Canadian business AI use at 19.2% in the second quarter of 2026, with data analytics the most common application at 36.6%. Document extraction sits below that in the rankings and above it in obviousness, which is the usual shape of the gap this site exists to close. A bookkeeping firm in Collingwood, a machine shop in Cambridge and a contractor in Northern Ontario are all doing the same 12 minutes per invoice, in Canadian dollars, by hand.
If you have the three numbers at the top of this page and they beat the build, you do not need an assessment to proceed. If you cannot get them, or if the same hour is being spent on four different tasks and nobody has separated them, that is exactly what the $999 assessment counts.