By Karthik Sivakumar, AI Enabler
The promise of AI is that it can transform our work, but taking advantage of that opportunity in a complex environment isn’t always straightforward. My name is Karthik, and I would like to share a story about my experience in using AI in the public sector and how that has inspired me to help others navigate from experimentation to adoption.
Introduction: The public service learning curve
At the age of 40, moving from the private sector to the Australian Public Service (APS) was a significant transition. I came from a service delivery environment, a world of operational metrics and rapid customer outcomes. Stepping into a budget policy officer role at the Department of Finance, the change in pace was matched only by the change in language.
I was suddenly navigating a complex fiscal landscape, deciphering everything from new policy proposals and portfolio budget statements to the intricate internal processes and rigorous timelines that drive the delivery of the budget and MYEFO cycles.
The experiment: Building the 'Budget Bot'
The opportunity to close this gap arrived during the caretaker period(Opens in a new tab/window). Taking advantage of what is traditionally a quieter period for budget teams, I used the downtime to experiment with a way to research faster and learn more effectively. My project was to build what I nicknamed 'Budget Bot(Opens in a new tab/window)'.
Using NotebookLM(Opens in a new tab/window), a tool with a key feature crucial for improving accuracy, I could 'ground' the AI strictly in the documents provided. Unlike open chatbots that search the entire internet, NotebookLM answers questions based only on the source files you upload. To power my bot, I used 10 years of publicly available budget and MYEFO papers.
To ensure my experimentation remained fully compliant with Finance’s internal guidance on the appropriate use of generative AI tools, I relied exclusively on publicly available, open‑source material, primarily published Budget and MYEFO papers. Finance’s Artificial Intelligence (AI) Transparency Statement(Opens in a new tab/window) offers a clear overview of how the department governs the use of AI, and I aligned my approach with these internal expectations.
Features in action: Bringing the budget papers to life
Budget Bot became my personal tutor, turning ten years of static PDF files into an interactive knowledge base. It allowed me to instantly bridge the 'institutional memory' gap, helping me decode government jargon and see the long-term story behind the numbers. This knowledge and understanding often take years to develop.
As a newcomer to the complexities of federal budgets, this tool made a real difference. It transformed static PDF documents into dynamic learning aids through several powerful features:
- Audio overviews: I converted dry economic summaries into engaging 'podcasts' where 2 AI hosts discussed macro-economic trends and concepts. Listening to these on my commute helped me absorb the tone behind the budget papers before I even reached my desk.
- Timelines: I was able to use my bot to trace the funding history of specific programs across a decade of documents. This allowed me to trace and visualise exactly when funding started and stopped for different programs. Anyone who has tried to research a program’s history through years of budget papers will know how notoriously difficult that can be.
- Briefs and mind mapping: The tool didn't just read the text; it mapped the architecture of the budget. It helped me visualise the relationships between disparate policy measures and quickly identify which specific section of the budget papers held the key details. This turned the sprawling archive into a structured, interconnected web of insights.
Since my initial experiments, the platform has evolved rapidly. My bot now includes video overviews, infographics, data tables, and slide decks, alongside multi-language support and diverse report formats like case studies, concept overviews, study guides and quizzes. With this type of technology, public servants who are new to the service, or to a particular area, can become effective much faster. For me, that meant I could make meaningful contributions to policy discussions and budget processes because I understood the terms and context much better.
It takes a village: From personal project to collaborative pilot
My director and assistant secretary were incredibly supportive, encouraging me to move my bot beyond a personal experiment. They connected me with other agency advice units and other internal stakeholders to test it, provide feedback, and validate its outputs.
The journey gained further momentum when I presented my experience at the AI Co-Lab, a cross-sector initiative where APS agencies, the private sector, and academia collaborate to explore effective adoption of innovative ideas. Sharing my story there was a major turning point. The feedback was overwhelmingly positive, and it sparked a wider conversation about other high-value use cases, such as an 'entitlements finder' or a 'senate estimates notebook'.
This experience proved that generative AI, when grounded in high-quality data, can significantly enhance productivity. It also reinforced how important it is to ensure that we have humans involved in designing, verifying, and using AI.
Strengths and limitations: The reality check
My experiments with my 'Budget Bot' offered clear takeaways. Its primary strength was demonstrating the productivity gains and high degree of accuracy that can be achieved by grounding a generative AI tool in a curated set of high-quality data.
Strengths: Velocity and direction
Beyond just improving accuracy, the speed of information retrieval was transformative. My bot didn't just find facts; it offered ideas on how to structure reports and acted as an encouraging, non-judgemental and knowledgeable assistant. It could instantly direct me to the most relevant document for deep‑dive research.
Limitations: Context and foundations
This project also underscored a critical requirement: humans are essential. Human oversight is essential for verification and to address potential context gaps that AI tools might miss. My bot does not replace the foundational knowledge required to navigate the fiscal landscape. You still need to understand the distinct purposes of budget papers 1 to 4 and MYEFO to interpret answers correctly and appreciate the nuances.
While generative AI can accelerate analysis and spark new perspectives, every output it provides still requires careful human review. In line with Finance’s internal guidance, AI is always used with human oversight, and staff remain fully accountable for the decisions and actions that AI tools support.
The roadblock: Hitting the 'scaling wall'
Despite the success of my experiments, I eventually hit a wall. My ambition was to transition the Budget Bot from a personal tool into a formal pilot, democratising access for the entire division. I wanted to empower my colleagues to leverage these features for their own work streams whether that meant accelerating general research, navigating complex finance processes, or reviewing detailed proposals with greater efficiency.
To support this work, I followed Finance’s internal protocols for using generative AI, completing both the AI use case summary and the AI risk self‑assessment before proceeding. This ensured the activity was assessed as a low‑risk use case and approved for use with unclassified information. Finance outlines this approach in its AI Transparency Statement(Opens in a new tab/window), which describes how the department applies whole‑of‑government policy, maintains internal oversight of AI use cases, and ensures risks are identified and managed before any AI tools are adopted.
However, I quickly learned that having a validated use case and funding approval is different from having the AI enterprise infrastructure to support it. Some of the challenges I faced:
- Access and accessibility: The 'single‑player vs multiplayer' dilemma. While I could share my bot for colleagues to view and use the chat interface, with a subscription tied to an individual account there was no easy way for teammates to use the full features or build their own projects.
- Continuity and lifecycle: The question of key person risk. If the knowledge base lived under my user profile, what would happen to that institutional asset if I moved roles? We lacked a mechanism for team-based ownership.
- Governance and trust: While I understood the security boundaries, scaling meant trusting a wider cohort of public servants. As I explored the possibility of scaling this experiment beyond a small, contained use case, I found that the level of governance, controls and authorisation naturally increased. From my perspective, navigating these enterprise frameworks at scale required more care and coordination to ensure everything aligned with the appropriate protocols.
The operational reality
Beyond the technical hurdles, there was the simple reality of capacity. As a budget policy officer, my primary role is defined by strict, time-bound deliverables. Solving these enterprise-level infrastructure gaps requires significant investment meeting with IT security, and governance stakeholders to design new pathways. I found myself in a catch-22: I had a tool that saved time, but I lacked the time to navigate the organisational complexity required to scale it while managing the pressures of my core role.
Even with approval to proceed, I found it difficult to move further because I was working in a space without pre‑existing pathways. Being a first mover meant taking on the full-time investment of navigating governance, security and approval processes from the ground up. These processes are important but building a new route through them while also managing the demands of my core role made progression challenging. The upside is that this groundwork will make the process easier for the next person who wants to explore a similar use case.
Conclusion: From user to enabler
This experience ultimately shaped my next career step. I realised that what I experienced was part of being early in an emerging space, not a barrier, but an opportunity. Individual innovation can spark powerful ideas, and by contributing to the early groundwork, I can help build the AI pathways that will turn future “experiments” into sustainable team capabilities.
I have now transitioned into the AI Delivery and Enablement (AIDE) function within the Department of Finance. AIDE was established to operationalise the AI Plan for the Australian Public Service, moving the APS from ad-hoc experimentation to coordinated, scalable impact.
My new role sits within the acceleration function. Our mission goes beyond helping individual policy officers; we are focused on scaling high-priority AI use cases, across multiple agencies, which deliver tangible improvements to government service delivery and policy outcomes for the Australian community.
This entails fast-tracking the uptake and re-use of successful AI solutions, so agencies don't have to reinvent the wheel across the APS. We will be collaborating closely with range of public servants to distil insights, normalise the use of AI and overcome barriers to scalability.
In the Acceleration team, I am using my experience with my Bot to help the APS scale AI solutions, ensuring that future APS innovators find a bridge, not a wall.
More information
View the read-only version of the Budget Bot(Opens in a new tab/window).
Register for my next Co‑Lab session(Opens in a new tab/window).
I would love to hear from others who have been experimenting to learn what’s worked, what
hasn’t, and where you might have encountered roadblocks. Reach out at aide@finance.gov.au