New Zealand's accounting press has spent the past year telling build stories. Acuity, the CA ANZ magazine, reported MGI+MORE developing an assistant called Rosie that now works across annual and management accounts jobs, and Wallace Diack building frankie, which downloads client reports and pre-fills workpapers, reportedly saving around seven hours per staff member per week. The reported numbers are the firms' own, but the direction is hard to argue with: the practices that automated preparation work are not going back.
The less-told part of every build story is the build. Someone in the firm became a part-time product manager. Someone maintained the thing when the accounting platform changed its export format. And the firm now owns software whose documentation is one enthusiast's memory. Those firms proved the value; they also paid the pioneer tax. In 2026 the same outcome is available off the shelf.
Build versus buy, honestly tallied
Time to live: custom bots take months of iteration; a Claude Cowork deployment is typically running inside a week
Cost: an internal build easily consumes $20,000 of staff time before it is reliable; the fixed-fee setup is $3,500 AUD plus subscriptions
Maintenance: your bot breaks when Xero or the IR portal changes; a platform agent inherits fixes from the vendor
Key-person risk: the builder resigns and the bot becomes archaeology; skills written in plain language survive staff turnover
Capability ceiling: a scripted bot does the one workflow it was built for; a general agent picks up new workflows by being given a new procedure
The one place build still wins is a workflow so unusual that no general tool models it. Almost nothing in a compliance practice qualifies: chasing documents, pre-filling workpapers and reconciling ledgers are the most standard problems in the profession.
What the off-the-shelf path looks like
Claude Cowork is the desktop agent version of Claude. It reads scoped client folders, connects to Xero and email by name, and follows skills: written procedures encoding the firm's workpaper template, chase email tone and materiality lines. No code, no hosting, no integration project. The skills are the same institutional knowledge the NZ builders hard-coded, except they are plain documents any senior can edit.
First workflows for an NZ firm: the client document chase, workpaper pre-fill from the Xero file, and a morning pulse across provisional tax and filing dates
Year-end fit: the 31 March cycle, IR filing under the extension-of-time regime, and the February-to-April crush are exactly the season the automation flattens
Privacy: scoped folders and named connectors map cleanly against Privacy Act 2020 principles, and nothing trains on firm data by default
Why the NZ stories matter on both sides of the Tasman
Australian firms should read the NZ build stories as market intelligence, not local colour. New Zealand practices moved first for structural reasons: smaller teams feel the capacity squeeze harder, the 31 March year end concentrates the pain into one brutal quarter, and CA ANZ has been unusually active in putting AI adoption in front of members through its resource centre and Acuity coverage. The playbook that emerged, automate preparation, keep judgment human, measure in hours per staff member per week, transfers directly to an Australian practice staring down its July-to-October lodgement program. The tooling generation has simply moved on: what took MGI+MORE a development effort in the build era is now configuration. Firms on either side of the Tasman starting today inherit the proof without repeating the project.
The frankie test
Seven reported hours per staff member per week is a useful benchmark precisely because a sceptical partner can test it. Pick two preparers, run one month of workpaper pre-fill and document chasing through an agent, and time it against the control pair. At Australian-equivalent charge-out rates of $160 an hour, five staff each recovering even four hours a week returns over $12,000 of capacity a month. If your own numbers come in at half the Wallace Diack figure, the setup cost disappears inside the first quarter anyway. That is the correct spirit in which to read any vendor claim, including ours: run the month, keep the timesheet, decide on your own data.
For the firm that still wants to build
Some practices have the appetite and the talent, and the builders clearly enjoyed it. The middle path is building on top of the platform rather than beside it: custom skills and MCP connectors on Claude give you the bespoke behaviour without owning the infrastructure. That is a weekend of tinkering, not a product programme, and it de-risks the experiment to nearly nothing. Book a brainstorm call or start with the fixed-fee Cowork setup, which deploys the same stack for NZ firms as for Australian ones, and we will happily compare notes with whoever on your team wanted to build Rosie.
