Jason Langley breaks down the three technical pillars, the cost of experimentation, and the governance CIOs need to scale AI safely.
At the NZ CIO Innovation Summit, headline sponsor Oracle tackled one of the most pressing questions facing New Zealand businesses today: how do you move AI from experimentation to real enterprise scale? Fresh off a joint stage session with Beca CFO Mark Fleming, Oracle NZ Managing Director Jason Langley sat down with Brightstar to unpack the practical framework behind that talk — and offer a candid look at what’s actually holding organisations back.
The three pillars of Enterprise AI
Langley’s talk centred on three foundational requirements for scaling AI responsibly:
1. Trusted data. AI without access to enterprise data, Langley argued, stays generic — it can’t deliver real business value until it’s grounded in an organisation’s own context.
2. Security and governance. That data access has to be secured and well governed, addressing the exposure risks many organisations are only beginning to grapple with.
3. Choice and flexibility. Perhaps the most timely point: recent weeks have shown what happens when a vendor “turns something off” that a business has built itself around. Langley pointed to Oracle’s shift from a historically siloed model toward openness — partnering with Microsoft, Azure, AWS and GCP — so customers can choose where they host workloads and which AI models they run on top of them.
The human layer underneath it all
Beneath the technical pillars, Langley placed two “human” ones: judgement and domain expertise.
On judgement, his message was blunt — critically assess AI output rather than blindly accepting it. He shared a pointed example from the summit’s round table: receiving 30- and 40-page AI-generated summaries and asking teams a simple question — would you have produced this much content in a pre-AI world? If not, why now? The goal of AI, he argued, isn’t more volume; it’s the same quality, faster.
On expertise, Langley was equally direct about his own limits: generating content is one thing, but he’s not a business analyst or a finance professional — so building the tools those domains need should stay with the domain experts themselves.
Why Beca stands out
Having worked with Beca in various capacities over 30 years, Langley pointed to the company’s century-long history and its track record as an early, safe adopter — moving from predictive AI, through generative AI, into agentic AI across Oracle’s infrastructure, database and application layers. His advice to organisations still on the sidelines was simple: start now, or risk being left behind on commercial advantage.
“I look at AI as a muscle… if you’re not developing that muscle, you don’t get better at it.”
Experimentation has a price tag – and that’s okay
One of the most relatable moments came when Langley shared his own AI habits. He regularly uses ChatGPT (via Oracle’s integrated chat), Claude, and Gemini — noting Gemini’s particular strength with image-related tasks. He’s also experimented with training AI on his own voice and his team’s, so generic content can be quickly personalised and pushed out at scale, cutting through the inertia of a blank page.
He didn’t shy away from the cost of that experimentation either: around US$1,000 in tokens spent during a single month of trial and error, for output he estimated was worth roughly $50. His take: that’s the price of learning, and multiplied across an enterprise it adds up — but the freedom to experiment, within a capped budget, is what builds real AI fluency over time.
Advice for CIOs: Empower then watch them closely
Asked for parting advice, Langley’s answer centred on enablement paired with visibility. Give employees the freedom, support and safe environment to experiment — but maintain enough oversight to catch things early and course-correct. The real risk, he noted, isn’t malicious misuse; it’s employees unintentionally exposing sensitive data to unsanctioned models simply because they don’t understand the risk. Educating people on that risk, he said, is what allows them to self-correct.
His closing line captured the session’s throughline: scaling AI isn’t primarily a technology problem. It starts and ends with the human layer.
