Singapore’s new capability programmes and production experience from Australian banks and insurers point to a practical test for AI investment: whether the people responsible for a system get the knowledge and practice to operate it well.

At Singapore’s Data + AI World Tour on 16 September 2026, Minister for Digital Development and Information Josephine Teo identified three obstacles to deeper AI adoption: skills shortages, integration with existing workflows and access to good data. An organisation has to deal with all three, often within the same project.

That has consequences for the people buying the technology. The team expected to maintain the data, question an answer or approve a change needs to be involved while the system is being built. Leaving its preparation until the handover meeting creates a dependency that another software purchase may do little to resolve.

Singapore and Sydney hosted Databricks World Tour events on the same day. Singapore’s announcements and Sydney’s published financial-services agenda offer two useful perspectives on that problem. One makes the connection between implementation and workforce capability explicit. The other sits alongside evidence from banks and insurers already putting AI into daily use.

Singapore: build capability into the project

Databricks announced plans to train 20,000 more people in Singapore over three years. More revealing than the target is the delivery arrangement announced alongside it. The EDB-supported Enterprise AI Transformation programme would provide participating companies with a Databricks-sponsored Forward Deployed Engineer, working with local partners on deployment and an operating model for scaling subsequent AI work independently.

These are commitments across enterprise sectors, with results still to be demonstrated. But the design deserves attention. The questions asked during development can shape what people learn, while those who will operate the system can challenge decisions before they become expensive dependencies. That gives a buyer a reason to plan and budget for learning alongside implementation.

For public bodies, the work also has to fit the agency’s responsibilities. Databricks’ GovTech case study describes an AWS-based platform migration, fine-grained access controls through Unity Catalog and an engineering-team-led plan to train users through Databricks Academy. It is a vendor-published account, but it gives a concrete example of platform change accompanied by preparation of the people who will use it.

Singapore’s public-sector data obligations also have their own basis. Government guidance distinguishes the Public Sector (Governance) Act and government instructions from the private-sector PDPA framework. Agencies set the data-security requirements for their suppliers. A useful learning exercise therefore needs the agency’s approved data, access arrangements and operating procedures, or a suitable substitute for sensitive material.

This is where a capable local team earns its place. It can work with the people who authorise access, understand how decisions move through the organisation and stay involved as the service enters operation. Regional specialists can add depth, provided the work and access are agreed. More offshore capacity does not, by itself, grant permission to move government data or enter government systems.

Sydney: useful AI brings operating responsibilities

Sydney’s published agenda included a financial-services forum on governed data and AI foundations, with speakers listed from NAB, Great Southern Bank and Lendi Group. Published customer evidence shows why those questions matter.

In June 2026, NAB reported that Databricks Genie and Genie Code were live within its analytics community. Early use was saving two to four development days per use case. Broader business access was still being piloted, with training and governance supporting the rollout.

The distinction between those two stages is useful. Helping experienced analysts work faster and opening a conversational tool to a wider workforce create different demands. Someone still has to judge whether the question was well formed, whether the underlying data is appropriate and whether the answer is fit for the decision being made.

In a published Databricks presentation, Suncorp reported around 18 in-house generative AI applications in production, used by more than 3,000 staff. The presentation also set out a wider claims-automation design that retained human review, with monitoring addressing different needs across business, engineering and governance teams.

For an insurer, a useful answer is only part of the job. Staff need a route for handling incomplete information, engineers need to investigate failures and managers need to understand when a service should be restricted or recovered. APRA’s CPS 230 requires covered institutions to manage operational risk, maintain critical operations through disruption and manage service-provider risks. That makes ownership and recovery arrangements part of the delivery problem. A platform cannot settle those responsibilities for its customer.

Teach the work people will inherit

The training question follows from the operating model. What must each group be able to do once the implementation becomes an everyday service? An executive may need to weigh a proposed use against its cost and risk. A data steward may need to approve and explain access. An engineer may need to diagnose a failed pipeline without waiting for the original author.

An illustrative exercise for a Singapore public body could follow an approved dataset from access request to published analysis, asking participants to explain each decision. For an Australian insurer, it could trace a claims assistant’s answer back to its source, then test how the team handles an unsupported response. Their value lies in making learning visible in the work, using the controls and escalation routes the organisation actually expects people to follow.

There is no reason to customise everything. Standard courses are useful for platform fundamentals and certification preparation. Banks and public agencies may already have substantial internal expertise. NAB’s staged rollout itself shows an institution organising access, training and governance around its own needs. Bespoke support earns its place when it closes a specific gap in the client’s systems, responsibilities or delivery experience.

The acceptance test should be equally specific. Can the team complete an agreed task, explain its decisions and recognise when to escalate? Attendance and examination results answer different questions. A client may choose to retain a managed service, but it still needs enough knowledge to oversee that relationship and make informed changes.

What a delivery partner should leave behind

The public Academy offer from Ultra Tendency separates certification training from bespoke learning adapted to roles, use cases and cloud environments. Ultra Tendency identifies Singapore public-sector work and Sydney finance and insurance as priorities, with customised training in both markets. That direction brings local delivery and client capability together, drawing on the company’s European engineering background and its technology- and cloud-provider-agnostic approach.

For Japanese enterprises investing across APAC, the useful distinction is between what can be shared and what needs local ownership. A common platform can support regional scale. The people responsible for each service still need to understand its data, users and consequences. Singapore and Sydney offer different examples of how that capability can be built.

The next project plan should name the people who will own the system after launch, the decisions they must make and the work they should be able to perform. Then it should give them time to practise before handover. Who owns that part of the plan, and is there a budget for it?