Automation for knowledge work: an AI Copilot that turns tax law into answers, not all-nighters.
Business challenges
Australian tax law isn’t just dense — it’s alive. Rulings shift, instruments multiply, and every opinion needs a citation chain you can defend in a partner meeting. Tax and legal teams bounced between databases, chased conflicting references, and wrote memos slower than the legislation changed. Billable hours went to hunting, not advising. Mid-size firms felt it most: too much change, not enough time, and no safe way to scale expertise.
Our approach to AI in professional services
We didn’t bolt a generic chatbot onto a statute book. We taught a copilot to think like a tax adviser and argue like one. We fine-tuned a domain LLM on millions of annotated rulings, ATO (Australian Taxation Office) guidance, and interpretive decisions — so it speaks the language of Australian tax, not another internet trivia.
We made sure there were no black boxes — every paragraph is linked, every claim traceable. We layered automated cross-checks across three legal databases, added inline footnotes, and shipped it as a web app and API any firm could drop into its existing portal.
Privacy and governance were non-negotiable. All inference runs on AWS CPU instances; on-prem vector stores keep sensitive context inside the firm’s boundary; prompts are field-level desensitised before the model ever sees them.
And the rollout? A new tab in the practice-management system. No downtime, no retraining the firm.
Technologies we used
Backend & Infrastructure
Frontend & Integration
Compliance & Governance
Key takeaways
One thing never changes in engineering: precision beats scale every time. In our projects, we start narrow and train deep in the business domain — that’s how you build trust faster than any general model ever could. We bake transparency into the workflow because citations aren’t decoration, they’re proof. And we don’t ship until privacy is part of the design. Success, for us, is measured in what matters most to the client: time saved, latency reduced, and confidence earned.
