What was done
AdaptAI built AutoLedger for a firm with 40+ SMB clients. It pulls transactions from QuickBooks Online, classifies them against Canadian tax categories with a GPT-4 prompt, and flags low-confidence items for human review.
Outcome
The developer reports roughly 5 hours saved per bookkeeper per week, month-end close reduced from three working days to under four hours, and eight new clients taken on without hiring. The firm is not named.
Evidence: Vendor case study (client unnamed). Figures are as reported by the source and have not been independently audited by AgentsNorth.
Prompt Reconstructed by AgentsNorth
The source didn't publish its prompt. This is our reconstruction of a starting prompt for a similar workflow, not the company's actual prompt.
Categorize each transaction below into the client's chart of accounts. For each, return: account, GST/HST/PST treatment (ITC eligible yes/no), confidence 0–1, and a one-line reason. Anything under 0.8 confidence or involving meals, vehicles or home office goes to a 'needs review' list.