1. Context is the variable that determines whether an AI system produces useful output. What does “providing AI with the right context” require from an infrastructure and data architecture standpoint?
Providing AI with the right context starts with knowing where relevant business information lives, whether it can be trusted and how it relates to other information. In most organisations, data is spread across operational systems, data warehouses, documents and real-time feeds. Bringing it together matters, but so does preserving those relationships and making clear what each piece of data means.
That requires reliable pipelines, consistent business definitions, current information, clear metadata and controls that determine who or what can access it. Organisations also need to understand how data was created and changed, so an answer can be traced back to an authoritative source.
For more complex systems, context includes the instructions an agent receives, the tools it can use, its memory and the actions it is permitted to take. Retrieval, context-window management and evaluation therefore become engineering disciplines, rather than exercises in writing better prompts.
This is why we introduced Genie Ontology. It acts as an automatic, continuously learning context layer that builds a semantic map of how your business actually works and ensures that AI agents are grounded in an enterprise’s unique enterprise truth rather than making blind assumptions. The model is only one part of the system. The quality of the data, architecture and controls around it will determine whether the output is useful, reliable and safe to act on.
2. AI cost management is becoming a board-level conversation as GPU spend scales faster than expected returns. What does responsible AI cost governance look like in practice?
Responsible AI cost governance starts with visibility across the teams, users, models, providers and business processes driving spend. Organisations also need to distinguish between experimentation, production workloads and activity that consumes resources without delivering a measurable business outcome.
AI spend is becoming ungovernable. Traditional tools are not enough because AI costs can vary according to token usage, model selection, agent experimentation and workflow retries. A monthly invoice tells leaders what was spent, but not necessarily why. Leaders need reporting that connects usage to specific people, environments, use cases and outcomes. This level of visibility and governance requires a centralised control point, like Databricks’ Unity Gateway, to enforce budgets, alerts, approval thresholds and hard token limits. Model routing can also help direct a task to the most appropriate model rather than defaulting to the most expensive option. Sandboxed environments can support experimentation, while production systems require tighter controls around access, performance and spend.
The objective is to create enough visibility and discipline for experimentation to continue without becoming an unmanaged liability. Cost governance should not simply police AI use. It should help teams make better decisions about where AI is delivering value, where it needs refinement and where resources should be redirected.
3. Governance and oversight of AI systems is something every organisation says it prioritises, and very few have operationalised it. What does AI governance require, and what’s the gap between what most enterprises have built?
Similar to cost governance, AI governance also begins with visibility. Organisations need to know which models, agents and tools are in use, what data they can access, what actions they can take and who is accountable for the outcome. Without that information, it is difficult to assess risk or enforce policy consistently. However, the reality is most enterprises govern AI assets separately, leading to fragmented systems and limited visibility.
Traditional controls only determine who accesses a model. But as AI workflows span multiple models, agents, and MCPs, it becomes difficult to understand what happened during a specific interaction. Teams often need to piece together logs from multiple systems to troubleshoot failures or audit AI activity.
To operationalise AI governance at scale, enterprises need a centralised control plane. This is why we introduced Unity Gateway. It delivers unified governance, providing one place for cost and security controls, covering Databricks models, external models, agents, MCP, and skills.
Built on Unity Catalog, it extends governance to runtime interactions. Administrators can enforce Contextual Service Policies in real time to block PII exposure or unsafe content. Furthermore, unified tracing captures model interactions and MCP tool activity in a single governed telemetry layer, providing end-to-end visibility into how workflows execute across services.
4. Natural language interfaces are changing how business users interact with data, reducing the dependency on data teams for analysis. How far along is that shift in practice, and where does the natural language interface still fall short of what business users need?
Natural-language interfaces are becoming useful for a growing range of questions, especially where the underlying data is well-structured and the business definitions are clear. For example, with tools like Databricks Genie One, a marketing or operations team can now ask for a performance trend in everyday language and get a starting point without waiting for a new dashboard or a data team to write a query.
Our customers are already applying this in practical ways. Airwallex uses Genie to help non-technical users analyse metrics such as campaign ROI, reporting 3-4x faster time to insight. Atlassian is bringing plain-English data insights into its Rovo AI assistant, and the City of Melbourne uses Genie to help explore near-real-time pedestrian data across more than 700 datasets.
However, it’s important to understand the shift isn’t a replacement for data specialists but merely changing where and how their time is spent. The quality of the answer still depends on whether the data is current, definitions are consistent and the system understands the organisation’s approved metrics. A user can ask about “revenue” or “active customers”, but the system needs to know which metric, timeframe and data source those terms refer to.
Natural-language interfaces without enterprise context also struggle when a question is poorly defined, the data is incomplete or the answer requires judgement rather than retrieval. They can produce a plausible response without fully understanding the decision the user is trying to make.
How we see this playing out in organisations today is through a redistribution of roles. Business users get faster access to governed analysis, while data teams focus on the models, definitions, permissions and quality controls that make self-service reliable.
5. Real-time data and AI are being positioned as the foundation for personalised business experiences. But real-time is an infrastructure commitment, not a feature toggle. What has to be true about an organisation’s data architecture before real-time AI can deliver on that promise?
For real-time AI to deliver personalised experiences, an organisation’s underlying data architecture must fundamentally shift. Today, moving data between operational databases and analytical systems introduces unacceptable latency.
The rise of agentic software development is driving a massive increase in the applications created, and every new app needs a system of record. Today, the new user is the AI agent, not the database administrator. In fact, our data shows agents now create over 80 percent of databases. These agents need operational databases that can respond in milliseconds while operating under the exact same governance as analytical systems.
This is why we introduced Lakebase, a serverless Postgres database built specifically for AI agents. By separating compute and storage, it helps teams build applications directly on Databricks using the same data foundation and Unity Catalog security as their analytics. It entirely eliminates the architectural tax of keeping operational and analytical data separate.
By processing real-time transactions and AI workloads in a single governed environment, organisations can finally power truly personalised business experiences without traditional infrastructure bottlenecks.
6. “Responding to changing business needs” implies an AI system that can adapt as conditions change. What does that adaptability require, and how do organisations maintain it without constant re-engineering?
Adaptability starts with recognising that AI systems cannot be treated as static models. Business conditions change and customer behaviour evolves, therefore AI systems need access to the right context, data, tools and permissions required to respond to those changes.
This means separating the underlying data and business logic from any single model or application. When data is governed centrally and business definitions are clearly documented, teams can update the information an AI system uses without rebuilding the entire system. The same applies to the tools and workflows an agent can access.
This is where model choice also matters. Organisations need to use the most appropriate models for each task, rather than relying on a single model, as this can create a locked-in approach and limit flexibility, cost efficiency and the ability to adjust as performance, requirements or available models change.
Organisations also need strong monitoring and evaluation. They need to understand how a system is performing, identify when its responses are no longer reliable and compare outcomes against business metrics. The goal isn’t to constantly re-engineer AI systems. Instead, create an architecture where the context, policies, and workflows can be updated in a controlled way as needs evolve.
7. The gap between organisations that are delivering intelligent business experiences and those still in proof-of-concept is widening. What distinguishes the ones that have moved from pilot to production?
The organisations moving from pilot to production are the ones that connect AI to a clear business priority and build the foundations required to operate it reliably, at scale.
They first define the outcome and how success will be measured before determining the technology required. That may be improving customer service, detecting fraud, reducing operational costs or helping employees make faster decisions. They then treat AI as an enterprise system rather than an isolated model. This means bringing together quality data, business context, security, governance, monitoring and reusable components, allowing teams to build new use cases without starting from scratch each time. It also means giving teams the control to deploy AI responsibly while managing cost and performance as usage grows.
We’re seeing this in practice with customers such as NAB and Bupa. NAB’s Ada platform replaced a legacy Teradata environment and now supports hundreds of production use cases across customer engagement, financial crime detection and regulatory reporting. Bupa used the Databricks Data + AI Platform to build a governed customer 360, which now powers a personalisation engine for faster, more relevant campaigns and underpins ML models that spot emerging health risks like prediabetes to enable proactive, personalised care. In both cases, AI is connected to core business processes rather than operating as a standalone demonstration.
The organisations still in proof of concept often have a promising demonstration. However, they lack clarity around ownership, data access, risk, cost or ongoing performance. Production-ready AI needs to have all of those pieces to work together and improve over time.
8. Regional investment from global technology companies often follows customer demand rather than leading it. What is Databricks committing to in ANZ in terms of local infrastructure, talent, and partner ecosystem?
Databricks’ investment in ANZ reflects both the strong demand we are seeing from organisations across the region and our long-term confidence in the market’s potential.
In May, we announced an investment of approximately US$300 million (approximately AU$420 million) across Australia and New Zealand over three years. This commitment spans infrastructure, people, skills and the broader customer and partner ecosystem. It will also support the adoption of Genie, Lakebase and Unity Gateway in the region.
We are continuing to invest in our local team through a new Sydney headquarters, which will quadruple our existing footprint and include dedicated training facilities. It will provide a hub for collaboration between our people and customers, including City of Melbourne, Airwallex, Atlassian, NAB, Telstra and Queensland Health, and partners.
We are also investing in the skills required to help organisations move from experimentation to production, in accordance with Australia’s National AI Plan. Over the next five years, Databricks plans to help upskill 100,000 learners across Australia and New Zealand through bootcamps, hackathons, customer enablement programs and multi-city learning festivals. This includes initiatives designed to help non-technical users work with data and AI more effectively.
9. What do the next twelve months look like for Databricks in ANZ in terms of product availability, customer expansion and regional capability?
Over the next twelve months, we expect to empower more ANZ organisations to move from evaluating individual AI capabilities to production at scale.
Bio:
Attributed to Adam Beavis, Vice President and Country Manager for ANZ at Databricks
Capabilities including Genie, Lakebase, and Unity Gateway will play an important role in this implementation shift and are already generally available in the region. Alongside our growing team in ANZ, our new capabilities help organisations develop governed AI agents, build responsive applications and make data more accessible for day-to-day decision-making.
These are also the central themes we are exploring at this year’s Databricks Data + AI World Tour in Sydney. We are seeing leaders across ANZ become increasingly focused on where AI can make a tangible difference, how it can be integrated into existing workflows and what is required to scale it responsibly. The question is becoming less about what AI can do in theory and more about where it can create lasting value.
Over the coming year, we are excited to see organisations across industries applying these capabilities to high-value use cases. Doing so will require them to connect AI to clear business needs, build on trusted data and strong governance, manage costs as deployments scale and work with the right partners.
