Australian accounting firms handle some of the most sensitive data a business holds: payroll, tax positions, trust distributions, and the personal finances of directors and their families. Open source AI promises full control over where that data lives, and the current crop of open models is capable enough to be taken seriously. The catch is that running them properly brings ongoing duties a busy practice needs to plan for rather than discover halfway through BAS season. This guide covers where AI actually helps, what self-hosting involves, what it costs, and a safe order in which to adopt it.
Where AI earns its keep in an accounting practice
The useful applications are not exotic. They sit close to the daily work of partners, managers, and graduates, which is exactly what makes them worth pursuing.
Drafting client correspondence, engagement summaries, and follow-up emails from file notes
Sorting, tagging, and describing transaction data before it reaches a reviewer
Preparing first-pass working notes and variance commentary for a senior to check
Answering staff questions from internal procedure manuals and checklists
Summarising long ATO correspondence and client document bundles
None of these replace professional judgement, and none should. They remove the repetitive drafting and triage that surrounds judgement, which in most practices is where the recoverable hours quietly leak away.
What running open source actually involves
Self-hosting an open model keeps client data inside infrastructure you control, which is a real advantage. It also makes the firm the operator of that infrastructure, permanently. Before committing, be clear about what the role includes.
Securing client data to the standard the Privacy Act expects, with evidence you can produce on request
Keeping the model server patched, monitored, and backed up
Logging access for professional standards and audit purposes
Maintaining availability through busy reporting periods, when the system matters most and staff have the least spare time
Re-evaluating the model itself every few months as new versions ship
That last point gets missed. Open model releases move quickly, and a model chosen in March can look dated by September. Someone in the firm has to own the upgrade cycle, test outputs after each change, and confirm nothing has drifted in quality.
The numbers for a mid-size Australian practice
Costs vary, but the shape is consistent. A firm of 15 to 40 staff that self-hosts a capable model should budget for GPU hosting at $2,500 to $3,500 a month, a security review before any client data touches the system, and a meaningful slice of a senior person's time to keep it all running. Add it up and $60,000 a year is a realistic figure for doing the job properly, before anyone has drafted a single letter.
A managed Claude deployment usually lands differently. A typical build for the same firm runs $20,000 to $45,000 to design and implement, with monthly usage costs in the hundreds rather than the thousands of dollars, and no specialist hire. The data-handling controls move into the design: which categories of client data may be sent, what gets de-identified first, and what stays out entirely.
Self-hosted: control over data location, but the firm carries security, uptime, and staffing
Managed Claude: faster to deploy, contractually defined data handling, no infrastructure to run
Hybrid: a narrow internal task on an open model, with client-sensitive work on Claude
For most practices the hybrid question is the honest one. There are tasks where open source genuinely fits, usually internal and narrow. The mistake is letting that narrow case justify hosting everything.
One more line item deserves attention: insurance. Professional indemnity insurers in Australia are starting to ask how firms supervise AI-assisted work. A documented review process and a clear record of which model handled which data make that conversation short. An undocumented self-hosted setup makes it long, and potentially expensive at renewal time.
A safe sequence for adoption
Accounting practices do best when they introduce AI in order of risk, not in order of excitement.
Start with internal drafting and summaries that never leave the firm
Classify client data early so the right work goes to the right model
Add client-facing outputs only once review controls have caught real errors
Document what was reviewed and approved, so the compliance story writes itself
This sequence builds staff confidence and a clean compliance record before client data is ever involved, which is exactly the order a regulator and a managing partner both prefer.
Making the call
If your firm has genuine infrastructure capability and a narrow internal use case, open source can earn a place. If the goal is reliable drafting, triage, and review support across the practice without adding an operations burden, a managed Claude build gets there sooner and with a clearer compliance story.
We help Australian accounting firms make this decision with the numbers on the table, defaulting to Claude for anything client-sensitive and using open source only where it clearly fits. Book a brainstorming session and we will map it for your practice.
