Case study · Accounting
Document-to-ledger automation for an accounting practice
Receipts, bills and bank statements flow in by email, get OCR'd and classified, and post to the right ledger account in QuickBooks or Xero, with a human approval gate on anything outside the model's confidence band.
The challenge
What we had to solve.
Bookkeepers were spending the majority of their week on data entry, opening attachments, retyping line items, classifying expenses and reconciling against statements. Nothing about that work was billable at a senior rate.
The firm had tried off-the-shelf OCR tools but none integrated cleanly with their ledger stack, and none handled the messy reality of Canadian small business paperwork: bilingual receipts, hand-corrected invoices, statements that mix CAD/USD.
Our approach
How we tackled it.
We built an ingestion pipeline that watches a per-client inbox, OCRs attachments, runs a classifier trained on the firm's actual category history, and produces a posting suggestion with confidence scores.
Above 90% confidence, the posting goes straight to a draft ledger entry. Between 70%-90%, it goes to a fast-review queue that a bookkeeper clears in under 5 seconds per item. Below 70%, the document goes to a human-review tray with the model's best guess shown but not applied.
The system writes back to QuickBooks Online and Xero via their respective APIs, and every action is logged for audit. Nothing posts to a client's ledger without an explicit approval chain.
We trained the classifier on 18 months of the firm's historical data and tuned the confidence bands during a two-office pilot before rolling out firm-wide.
Deliverables
What we built.
Specific, named outputs, not vague "strategy".
Inbox ingestionPer-client email inbox monitor, attachment OCR, vendor + amount + date extraction.
Category classifierFirm-trained ML model that suggests ledger account, GST/HST treatment and project tag.
Review UIBookkeeper queue, keyboard-driven approval, audit trail and per-client confidence tuning.
Ledger syncTwo-way write-back to QuickBooks Online and Xero with idempotent posting and reconciliation.
ReportsWeekly time-saved per bookkeeper, classification accuracy trend, exception backlog.
Outcomes
What it returned.
- 22 hours per week reclaimed across the bookkeeping team in the pilot offices, verified by time-tracking comparison.
- Classification accuracy stabilized at 93% above the confidence threshold within 6 weeks of training.
- The firm reallocated reclaimed hours to advisory work that bills at 4x the rate of data entry.
- Audit trail held up cleanly against PIPEDA + provincial CPA requirements, which the firm had been worried about.
The takeaway
What we learned.
Accounting automation isn't about replacing bookkeepers, it's about deleting the most boring parts of their week and freeing them to do the work clients actually need. The confidence-band design was the key: full automation where it's safe, human review where it isn't, no surprises in either direction.
Services used
How we got it done.
Related work