Every Australian accounting firm has heard the AI transformation pitch, and most partners have learnt to discount it. So here is the version with the discounting already applied: a 30, 60 and 90 day roadmap for a five-partner suburban firm, with the failure modes included, and an honest claim at the end. Sixty-hour weeks becoming forty is the trajectory across a season, not week one, and anyone promising week one is selling something.
The starting maths
A five-partner firm running 15 staff through a fifteen-week season at five assembly hours per person per week burns roughly 1,100 hours on work that produces no advice: chasing documents, keying data, reconciling, drafting the same emails. At a blended $170 charge-out that is more than $180,000 of capacity per season. That number is the budget for this whole exercise, and it makes a $3,500 fixed-fee setup plus Claude subscriptions look like what it is: a rounding error against the problem.
Days 0 to 30: the chase and the pulse
Deploy Claude Cowork with scoped client folders, Xero and email connectors, and a written one-page AI policy
Switch on the document chase: live outstanding-items lists per job, drafted reminders staff approve before sending
Switch on the morning pulse: deadlines, blocked jobs, overnight arrivals, sorted before the office opens
Pick one senior as the owner; tools without an owner quietly die
The chase is first because it pays back immediately and builds trust: nobody mourns losing that job. By the end of the month documents are arriving earlier and the mental load of remembering has moved into software.
Days 30 to 60: regulated work, measured honestly
BAS first pass: GST reconciliation and draft workpapers with exceptions lists, reviewed by the same seniors as before
Payroll reconciliations and STP variance checks on the employer book
Timed side-by-side benchmarks on at least ten live files, review time included, written down
The benchmark discipline matters more than the tooling. Firms that skip it end up arguing from impressions in the partner meeting; firms that keep it have a number, and the number decides the rollout.
Days 60 to 90: workpapers, letters, and the boring institutional bits
Year-end workpaper preparation from the Xero file and folder, in the firm's own template
Engagement letters and routine client correspondence drafted for review
Skills tuned to the firm's materiality lines and known client quirks
The AI policy updated from practice, and consent language added to the engagement letter template
What the partners should be doing while this runs
The roadmap fails if it is treated as an IT project. The partner group has three jobs across the 90 days. First, protect the review standard publicly: staff take their cue on whether AI-prepared work is second-class from what partners say in the corridor, not from the policy document. Second, decide in advance what the recovered hours are for, because capacity without a destination becomes early finishes for a fortnight and then quietly refills with low-value work. The good destinations are advisory conversations, the tax planning season that always gets squeezed, and simply not burning out the two seniors who carry every July. Third, talk to clients about it before they ask. The firms that frame AI adoption as a service improvement, faster turnaround and earlier warnings, control the story; the firms that stay quiet get asked whether the fees should be lower now.
The money, restated plainly
Costs across the first season: $3,500 setup, Claude seats for the team, perhaps twenty senior hours of tuning and benchmarking. Return: a conservative half of the assembly layer, call it $90,000 of capacity at the blended rate, plus the amendments that never happen because the exceptions surfaced early. No revenue growth is assumed anywhere in that arithmetic. If the benchmarks at day 60 do not support continuing, stop; the sunk cost is smaller than one month of the problem.
The failure modes nobody puts in case studies
Three patterns kill these rollouts. First, skills never tuned to the firm's actual templates, so output needs rework and staff conclude the tool is more trouble than it saves. Second, quiet bypassing: one sceptical senior processing files the old way and the benchmark data silently corrupting. Third, review discipline slipping once trust builds, which works right up until the file where it does not. Each has the same antidote: a named owner, written benchmarks and the review step treated as non-negotiable, exactly as it would be for a new graduate.
What forty hours actually looks like
By day 90 the realistic picture is preparation time halved on the automated job types, documents arriving weeks earlier, and the season's overtime concentrated on genuine judgment work instead of assembly. The sixty-hour weeks do not vanish; they stop being spent on work a machine does better, which is what makes the next season shorter than this one. Start with the fixed-fee Cowork setup at $3,500 or book a brainstorm call and we will build the 90-day plan against your actual job mix, benchmarks included.
