Xephyr has been named a finalist in the Professional Services category of the 2026 Financial Review BOSS Most Innovative Companies, for Xavier, our analytics agent.
This year's list says a lot about where Australian business is heading. In her introduction to the 2026 list, the Financial Review's special reports editor Penny Durham noted that nearly every winning project uses artificial intelligence to solve a problem. She also noted that this comes at a time when public mood about AI has shifted from excitement to scepticism, and in some places to resentment.
We think the scepticism is fair. Most of it comes from AI that gives answers nobody can check, running on data that has left the building. Xavier was built to fix exactly that problem, and we are proud the judges recognised it.
The innovation: an analytics agent that replaced our own work first
For five years, Xephyr has delivered data and analytics work for some of Australia's largest organisations, including Wesfarmers, Telstra and Bupa. Much of that work followed the same pattern. A business leader asks a question. An analyst writes the query, checks the result, builds the chart and sends it back. A single question can take two to four hours.
In 2024 we made a call: large language models would make most routine analytics automatable within a few years. We could wait for someone else to automate our work, or we could do it ourselves. We chose to build the product that would replace our own billable hours.
That product is Xavier.
Xavier lets business users ask questions in plain English and get accurate, governed answers from their own data. It turns the question into a query, runs it against the organisation's data platform and returns the answer with a full audit trail. Every answer shows the original question, the query it wrote, the data it used and how it reached the result. Nothing is a black box.
Built for trust, not just speed
Three design decisions set Xavier apart.
| Decision | What it means for our clients |
|---|---|
| Sovereign by design | Xavier runs inside the client's own environment, on-premise or in their private cloud. Their data does not leave, and their existing access controls still apply. |
| Trained on real domain knowledge | Xavier arrives configured with five years of consulting knowledge: metric definitions, business rules and validated query patterns from real enterprise work. |
| Model-agnostic and cost-efficient | Xavier runs on a range of models, including open-weight models from NVIDIA. This has cut inferencing costs by 4 to 5 times compared with deployments that use frontier models only. |
These choices took longer to build. A cloud-only product would have been faster and cheaper for us. But our mission is to give enterprise leaders trusted answers from data they control, and you cannot honour "data they control" with a product that moves the data somewhere else.
The results
Before any deployment goes live, we test Xavier against a structured set of 100 questions, from easy to hard, and check every answer against known-correct results on the client's production data.
In validated testing with a major Australian telecommunications company, Xavier achieved:
- 98.8% accuracy across the 100-question evaluation
- Questions that took 2 to 4 hours now answered in under 3 minutes, an estimated 90% or greater reduction in analyst time for recurring reporting
The use case was call quality reporting, a high-volume area where network engineering teams had spent significant time writing and checking queries by hand. The client's engineers reviewed Xavier's logic at each stage, from pilot through to production.
Changing how we work, without leaving people behind
Building Xavier meant changing our own business. We redesigned our workforce with our team, not around them. We defined three roles for an AI-native consultancy (Navigator, Pathfinder and Engineer) and reskilled our people into them. Data analysts became data and AI consultants. Software engineers became the backbone of our agentic delivery function. No roles were made redundant.
Our team now uses AI across our own delivery every day, with productivity gains that range from 25% on complex work to more than 400% on templated tasks. We use what we sell.
All of this sits inside an ISO 27001 certified security framework, independently audited, which covers how we build, deploy and run Xavier.
What's next
Xavier is moving from answering questions to acting on them. Our roadmap includes multi-step workflows, memory that learns from feedback, and deeper integration with the systems our clients already use. We are also working with hardware partners to bring fully integrated sovereign AI to organisations with strict data residency needs.
"Being recognised alongside this year's finalists means a lot to our team. The question for enterprise AI is no longer whether it can answer a question. It is whether you can trust the answer, and whether your data stayed where it belongs. That is the problem we set out to solve."
Lawrence Puang, Founder and CEO, Xephyr
Thank you
Thank you to our team, who built Xavier while still delivering for our clients every day. Thank you to the clients who tested it with us and held it to a high standard. And thank you to our partners, including NVIDIA through its Inception Program, for backing our work.
Congratulations to all the 2026 winners and finalists, and to the Financial Review and KPMG for 15 years of recognising Australian innovation.
Want to see how Xavier works on your data? Book a conversation with our team
Read the full list: Meet the Most Innovative Companies of 2026, The Australian Financial Review