AI implementation success starts with getting the foundations right
While Australia’s race to embrace AI is accelerating, many organisations are discovering that ambition alone does not translate into business value. Shreshta Shyamsundar, Distinguished Technologist at Infosys, explains why successful AI adoption depends on investing in the foundational capabilities that underpin the technology.
Beneath Australia’s enthusiasm for AI lies a sharp reality check: As per a recent report by ADAPT, 78% of boards frame AI as a strategic priority, but only 24% actually have AI-ready data architectures to pull it off. The report underscores the fallout, with 72% of Australian CDAOs stating that they are yet to realise measurable ROI. The reasons? Unreliable data, governance gap, and difficulties in scaling value.
Closing this divide calls for prioritising core AI foundations ahead of the next wave of deployments. At a time when lifting productivity has become a national economic priority, getting these foundations right is becoming as much a business imperative as a technology one.
Build the data foundation
The tendency for enterprises to deploy advanced models before optimising the broader data estate often creates a critical point of failure. For instance, pilot projects that perform flawlessly in controlled settings often collapse under live operational conditions. Ultimately, data preparation dictates the success of the entire lifecycle, and the quality of everything that follows depends on it.
Recent survey data indicates that these data foundations are slowly maturing. In the retail industry, for example, just 12% of Australian retailers feel fully confident that their customer and product data foundations are ready for AI-led use cases. However, critical operational data remains heavily siloed across legacy architectures, complex operational networks, and fragmented cloud environments, most notably within capital-intensive sectors such as mining, banking, and healthcare.
Engineer the context layer
Getting this data ready for AI deployments demands far more than cleaning and structuring: it requires organisations to confront years of deferred decisions around lineage, access control, labelling, and the harder cultural work of agreeing on what the data means. This is the tough, unglamorous infrastructure work that turns promising pilots into production-grade systems. Data teams must update standard operating procedures to maintain data accuracy and pipeline hygiene over time.
The discipline that increasingly separates AI programs that hold up in production from those that quietly fail is context: what a system is given to reason over, before it generates anything. But context alone is not enough. In a regulated enterprise, the context an AI system acts on must be verified and traceable to its source, not merely retrieved.
That is the difference between a model that produces a fluent answer and one whose answer can be trusted – because every element it relied on, whether a business definition, a regulatory rule, or a customer record, is grounded in a known, current, governed source rather than assembled from whatever happened to be nearby. In banking, that means a risk model reasoning over the regulatory definition in force today, provably; in healthcare, that means clinical context that is complete and current at the point of decision.
Make governance a runtime property
Context engineered this way, grounded, structured and verifiable, lifts output quality and stakeholder trust together, because they are the same problem.
Most governance failures at scale come from treating governance as something added after the fact: explainability reports and audit logs assembled once decisions have already been made. That model breaks the moment agents begin to act on their own, because the record arrives after the action it was meant to control.
The organisations scaling AI safely are moving governance into the runtime itself – accountability enforced at the point an action is taken, not reconstructed afterwards. In practice, that means every agent action traces back to the human authority that permitted it: an unbroken line from decision to accountable person. As enterprises deploy fleets of semi-autonomous agents, that line is what stops human authority and machine action from quietly drifting apart.
Scale with an AI-first framework
Governance built this way is not a compliance cost bolted onto the architecture; it’s the control plane that makes scaling possible in the first place. As Australia’s AI governance landscape continues to evolve, organisations that embed accountability into day-to-day AI operations will be better positioned to meet growing regulatory and stakeholder expectations.
With data, context, and governance foundations in place, execution speed becomes the decisive variable. An AI-first operating framework can help organisations connect these three foundations across AI and agent deployments while keeping human oversight intact.
Platforms such as Topaz from Infosys offer one example of how these principles can be put into practice. That is what lets organisations move past managing isolated use cases and towards integrated capabilities that scale across multiple business functions.
The organisations positioned to capture much of the significant AI-led growth are those investing heavily in foundational capabilities today, well ahead of the next enterprise technology cycle. Their structural advantages and productivity gains will inevitably accumulate, sector by sector, across the broader economy.
