title: "Tax-return first-draft automation for a mid-tier AU accounting practice" dek: "A 2023 build that took individual tax returns from 75 minutes to 14 minutes per file. Practice put through 38% more returns the following financial year." sector: "Professional Services" client: "Mid-tier AU accounting practice · ~22 accountants" engagement: "Audit → Pilot" duration: "12 weeks" year: "2023" outcome: "Individual return prep: 75 min → 14 min · 38% more returns processed in FY24" solution: "Document-intelligence pipeline (PAYG, ATO pre-fill, deduction substantiation) writing first-draft returns directly into Xero Tax with accountant review queue." timeSaved: "~61 minutes per return · A$0.34 cost per return drafted" visual: "none" cardFigure: "workflow" timeMetric: "61 min" timeMetricLabel: "saved / return" costMetric: "A$0.34" costMetricLabel: "cost per return" speedMetric: "5.4×" speedMetricLabel: "faster prep" publishedAt: "2023-08-14" keywords:
- tax return automation Australia
- Xero Tax AI
- accounting practice automation
- ATO pre-fill processing
The problem
A mid-tier AU accounting practice — twenty-two accountants, around 6,400 individual return clients — was running its annual tax season the way most AU accounting practices run theirs: with weeks of overtime and a stack of paperwork that grew faster than headcount.
A typical individual return took 75 minutes of accountant time: 20 minutes reading client-supplied documents (PAYG summaries, deduction receipts, share-trading statements), 30 minutes typing those into Xero Tax, 15 minutes reconciling against the ATO pre-fill, and 10 minutes of final review.
The Practice Manager wasn't looking for AI. She was looking for throughput. Hiring two more junior accountants would have absorbed margin without solving the seasonal-peak problem.
What we did
A two-week scoping engagement followed by ten weeks of build. The deployed system:
- Accepted client documents through the practice's existing client portal (no new tooling for clients)
- Ran a document-intelligence pipeline that extracted PAYG totals, deduction line items, dividend statements and capital-gains events
- Pulled the matched ATO pre-fill data through the practice's existing connection
- Reconciled the extracted client documents against the pre-fill (flagging mismatches)
- Wrote a first-draft return directly into the practice's Xero Tax workspace
- Routed every draft into an accountant-review queue with confidence scores per line item
The accountant reviewed the draft, corrected anything below threshold, added any deductions the system hadn't seen evidence for, and lodged. The system never lodged. Every return was reviewed and signed by a qualified accountant.
The outcome
| Before (FY23) | After (FY24) | |
|---|---|---|
| Time per individual return | ~75 minutes | ~14 minutes (including review) |
| Returns processed in season | ~5,900 | ~8,150 (+38%) |
| Accountants added | n/a | 0 |
| Reconciliation mismatches caught by system | n/a | 312 (saved an estimated A$47K in client-relations issues post-lodgement) |
| Cost per return drafted (model + infra) | n/a | A$0.34 |
| Senior accountant complaints about season-end fatigue | "Significant" | "Almost none" — Practice Manager, internal staff survey |
The practice didn't drop their fees. The 38% throughput gain became margin. The Practice Manager reinvested some of it in a paid week off for the team after lodgement deadline — first time in nine years.
We used to lose two accountants every year to burnout. We didn't lose one in FY24. That's the metric I actually care about.
— Practice Manager, mid-tier AU accounting practice
What we'd do differently
Build trust-mode early. Accountants resisted the system for the first three weeks — they didn't trust the drafts. We should have shipped a "show me the source of every figure" view from day one rather than week four. Once that landed, adoption was immediate.
Tier the routing. Simple returns (PAYG-only, no investments, no business income) could have gone through a faster path with a shorter review. We treated all returns identically; in retrospect a tiered confidence threshold would have saved another 4–5 minutes per simple file.
What we didn't do
We didn't replace Xero Tax. We didn't auto-lodge anything. We didn't train a custom model on the practice's client data. We didn't deploy any tool that takes a tax position without a qualified accountant signing it off.
This was 2023 work. The pattern — AI does the typing, the qualified human does the judgement — is now how every regulated-industry build we ship is shaped.