Example workflow — based on patterns we've seen, not yet a deployed case study.
Accounting & bookkeeping practice · BAS prep + supplier reconciliation
The problem. Quarterly BAS prep means downloading bank statements, MYOB/Xero exports, and supplier invoices that arrive as PDFs, scanned receipts, email attachments, and the occasional fax. The bookkeeper spends two solid days per client reconciling line items — most of it is matching, not judgement.
What we'd build. A workflow that watches the client's inbox (or a forwarding alias), parses supplier invoices on arrival, matches them against MYOB/Xero bills already in the system, flags mismatches, and drafts the GST coding for the bookkeeper to confirm. Everything is held in a review queue — nothing posts to MYOB until the bookkeeper ticks it.
Realistic outcome. Two days of reconciliation down to half a day, on a typical 100-supplier quarter. The bookkeeper still does the judgement calls; the AI does the typing.
Example workflow — based on observed patterns at owner-operator factories.
Small manufacturer · supplier-email parsing + job updates
The problem. A 12-person job shop in regional Victoria takes 30–50 supplier emails a day — price changes, ETA updates, 'back-order on the M12 bolts.' The production manager reads each one, decides whether it affects any open job, then walks out to the floor to tell whoever needs to know. Real cost: the things that don't get walked out on time.
What we'd build. An agent that reads each supplier email, cross-references the open job list, identifies which jobs are affected (by component code, by supplier, by delivery date), and posts a structured update — 'Job #4471 — M12 bolts delayed 4 days, affects Friday's assembly' — into the existing whiteboard channel (Slack, Teams, or a wall-mounted tablet). No invoices touched. No POs raised. Just the read-and-route work.
Realistic outcome. Most days the production manager opens the channel instead of opening 50 emails. Some days the agent's read is wrong and the manager corrects it. The wrong reads are useful — they show what the agent missed and what to fix next sprint.
