How Telecom and Mining Companies Make AI Work in Production

Recap of "AI Pilots Are Easy. These Two Leaders Figured Out the Hard Part." — Gartner Data & Analytics Summit, Sydney, June 2026

Walk the floor at any major data and analytics conference today and you will hear the same word from every vendor: trust. Trust in your data. Trusted AI. Trusted pipelines. It has become so ubiquitous that the word has almost stopped meaning anything.

Which is exactly why the opening line of Zach McIntyre's panel session at Gartner's Sydney summit landed the way it did.

"Every vendor is talking about trust," he told the room. "It's almost like trust dilution."

What followed wasn't a product pitch. It was something considerably more useful: a candid, detailed account of what two enterprise data leaders actually did to make AI work in production. Not in a pilot. Not in a proof of concept. In production, at scale, with real regulatory exposure and real business consequences if they got it wrong.

The problem most organisations are solving is wrong

The industry narrative right now runs something like this: if you want AI to work, upgrade your infrastructure, pick the right model, and plug it in. Vendors promise that the complexity can be abstracted away, that a sufficiently clever context layer or intelligent agent will compensate for whatever's missing in your underlying data estate.

The data says otherwise.

According to MIT's The GenAI Divide: State of AI in Business 2025 report — drawing on more than 300 enterprise deployments and 153 leadership surveys — 95% of enterprise generative AI pilots deliver no measurable profit-and-loss impact.

Gartner predicts that through 2026, 60% of AI projects will be abandoned due to inadequate data foundations, and that 63% of organisations either lack or are unsure they have the right data management practices for AI in the first place. 

The share of companies scrapping most of their AI initiatives jumped from 17% in 2024 to 42% in 2025, with the average company abandoning 46% of proofs of concept, according to S&P Global Market Intelligence.

This is the real story behind the AI confidence crisis… and it has almost nothing to do with models.

"If your data foundations don't have accuracy, quality, and context," Chris Michael told the Sydney audience, "then anything you do at the AI layer is just a hyperscaler on top of that problem. You accelerate garbage in, garbage out tenfold."

Michael is the Head of Analytics Products and Data Governance at TPG Telecom, one of Australia's largest telecommunications providers. The company was built through multiple rounds of M&A, carrying the full weight of that history in its data architecture.

His co-panelist, Dale Brimblecombe, Head of Data & Analytics at OceanaGold, arrived at the same conclusion from a different direction. A global gold mining operation spanning New Zealand, Australia, the Philippines, and North America, OceanaGold runs in a decentralised structure where data governance can't be mandated from the top; it has to be earned from the middle.

Two industries. Two very different constraint environments. One shared conviction: the foundations aren't optional.

What the ground truth actually looks like

The popular counter-narrative to heavy data governance is speed. Teams argue they can't afford to wait two years for a perfectly curated data environment; they need to move now, ship pilots, show value, and govern later. It's a reasonable concern dressed up as a strategy.

The problem is that "govern later" tends to become "govern never." And when AI gets deployed on top of ungoverned data at scale, the failure modes compound faster than any team can respond to.

Gartner's own research is stark on this point: 80% of governance initiatives will fail by 2027, primarily because they aren't connected to business outcomes. Not because governance is the wrong idea, but because it's being run as an IT project rather than a business capability. The organisations that tie governance to specific business outcomes and AI readiness will pull ahead. Those running it as a compliance exercise will fall behind and have little to show for it.

This is the distinction Michael and Brimblecombe kept returning to: governance not as a constraint, but as the thing that makes everything else possible.

Michael described the moment this clicked for his team as a shift in a single question asked in meetings. Early on, the question was: "Where did you get that data from?" Over time, as TPG's governance programme matured, that question changed to: "Are you using the certified source or the semantic model for that dataset?" Small difference in phrasing. Massive difference in what it implies about the organisation's data culture.

"It was a slow journey," Michael acknowledged. "But it's progressed really well."

Brimblecombe’s framing was even more direct:

"Start with governance. Even if it feels like too much, it's what's going to make AI go fast."

Building the foundation AI can actually use

For TPG Telecom, the path to governance ran through community. Faced with a federated analytics structure (with distributed teams across sales, marketing, and operations, each effectively siloed), Michael launched a community of practice that started with 50 people and has since grown to over 300 members, with 150 actively participating in every session.

The purpose was deliberately practical. Not to run governance for its own sake, but to identify the data that was actually being used to make decisions, nominate stewards who could curate it, trace critical data elements back to source systems, and build a certified semantic layer that the whole organisation could trust.

"We focused early on making sure the reporting layer and the semantic layer feeding that reporting layer were the most invested areas," Michael said. "We implemented this through certified and trusted data products in that semantic layer surfaced through Alation, which drove adoption across the organisation's reporting and analytics functions."

This is Alation's Semantic Model Mastering capability in practice. When an enterprise has multiple semantic models living across dozens of systems, it doesn't have governed definitions; it has six different answers to the question "what is a customer?" The community of practice became the mechanism for mastering that in one place, creating the single certified layer that downstream AI can actually reason over.

At OceanaGold, the foundation-building took a different form but required the same discipline. Dale's team spent years documenting the organisation's metrics with a level of rigour that initially met significant internal resistance: names and values, as well as calculation methods, plain-language business definitions, ownership, approval chains, and accountability. Today, OceanaGold has around 2,000 metrics documented in Alation.

"It seemed like a lot," Brimblecombe recalled. "There was pushback. But I could see, even before AI was the topic it is now, that if we were ever going to automate or continuously improve the platform, we'd need this level of rigour."

The payoff was unambiguous. "The first time we threw Claude at it, first try, the discipline paid off completely. It just worked."

That's not luck. That's the difference between an AI model working with rich institutional knowledge and one hallucinating atop ambiguity. The 2,000 documented KPIs are precisely what makes AI output trustworthy; the model isn't guessing at definitions, it's reading them.

The Alation perspective: Invisible infrastructure

There's a concept Brimblecombe introduced during the session that deserves to sit with anyone building or leading a data function right now: invisible infrastructure.

The data catalog was historically a place analysts went to find and annotate datasets. Today, it is becoming something else entirely. As agentic AI tools accelerate curation and as APIs make the catalog callable from anywhere in the enterprise stack, the catalog stops being a destination and becomes a knowledge layer. This is the foundation that makes accurate AI possible in the enterprise, whether or not the people using the downstream tools ever think about it.

"People might not engage with it directly," Brimblecombe said, "but when they connect their tools and ask a question, the catalog is the omnipresent knowledge laye that holds all of that institutional information."

He illustrated this with a recent example. The week before Gartner Sydney, his analyst, someone who wasn't coding six months ago, used Claude to generate a comprehensive operational report for one of OceanaGold's assets. The analyst extracted measures from the semantic model, cross-referenced them with what was documented in Alation, and pushed the results through the API. No traditional development process. No specialist SQL work.

The barrier to entry dropped because the knowledge layer was already there. The discipline of the previous years (the 2,000 documented metrics, the curation sessions, the rigour around naming conventions) is what made an analyst who wasn't coding six months ago productive with AI tools today.

This is the business case for what Alation's Agent Studio and CDE Manager are built to enable: a governed knowledge layer that AI agents can call with confidence, where critical data elements are tracked, trust flags are attached, and the institutional knowledge of the organisation is machine-readable. Not for the data team's benefit alone, but for every downstream consumer who connects a tool and asks a question.

"It's about making data access genuinely more accessible to the broader business," Brimblecombe said. "If you don't have the right foundations, when that scales, it gets worse way faster than you expect."

The Gartner D&A Summit 2026 keynote put a name to what underpins all of this: return on integrity. The organisations that invest in the trustworthiness of their data (in quality, lineage, governance, and documented institutional knowledge) are building compounding returns as AI capability scales. The organisations that skip those investments are building compounding risk.

The two-speed reality

One of the most practically useful frameworks from the session was Michael’s description of a two-speed operating model for governance: running the value-creating use cases the business needs today, while simultaneously building the foundations the business will need tomorrow.

"You can't just walk away for two years and build foundations," he said. "You still need to deliver while you build."

This matters because the false binary — "govern first" versus "ship now" — is precisely what leads organisations into pilot purgatory. They ship AI without foundations, hit the quality and trust problems, lose confidence, and start over. The alternative isn't to wait for perfect foundations before deploying. It's to build the foundations in parallel, with enough rigour that each new AI use case benefits from what came before rather than inheriting its problems.

The Data Products Marketplace model is designed for exactly this: packaging governed, certified data into reusable products that teams can consume with confidence, reducing the cost of doing the right thing for each new initiative. When the cost of accessing trusted data drops to near-zero for business consumers, the governance investment starts paying dividends across every use case, not just the ones the data team directly supports.

What to do today

Michael and Brimblecombe closed the session with one piece of advice each, not for next quarter, but for this week.

Brimblecombe: "Start with governance. Everything you know in your business: document it, because if you can't put it into a form that earns trust, nothing that sits on top of it will either. Governance isn't the thing that slows you down. It's the thing that makes it all go."

Michael: "Start with the business problem and the value. Everything you do from a solutions perspective has to tie back to what you're trying to achieve and where the value actually comes from. Don't start with a technology. Start with the problem."

Both pieces of advice point to the same root cause of AI failure… and the same correction. AI doesn't fail because the models are bad. It fails because the problem isn't defined clearly enough, the data isn't trusted enough, and the institutional knowledge the model needs to reason accurately is locked in people's heads rather than documented in a place that AI can access.

The organisations in this room that are still looking for the technology shortcut to production AI aren't going to find it. The organisations that accept this reality and start building the knowledge layer, this week, not next quarter, are the ones that will have something real to show for their AI investment twelve months from now.

FAQs

1. Why do most enterprise AI pilots fail to reach production? Most AI pilots fail because of weak data foundations, not weak models — poor data quality, ungoverned semantic definitions, and missing institutional context mean AI amplifies existing data problems rather than solving them.

2. What is a "two-speed" governance model? It's an operating approach where organizations run current, value-generating AI use cases while simultaneously building long-term data governance foundations in parallel, rather than choosing between shipping now or waiting years to govern first.

3. How did TPG Telecom build a data foundation for AI? TPG Telecom launched a cross-functional community of practice — growing from 50 to over 300 members — to identify critical datasets, assign data stewards, and build a certified semantic layer that unified reporting definitions across the business.

4. How many metrics did OceanaGold document before using AI successfully? OceanaGold documented roughly 2,000 business metrics, including definitions, calculation methods, and ownership, which enabled accurate AI outputs on the first attempt when applied to real reporting tasks.

5. What does "invisible infrastructure" mean in data governance? It describes how a data catalog evolves from a manual reference tool into an embedded, API-accessible knowledge layer that AI agents and business tools call automatically, without users directly interacting with the catalog itself.

6. What role does a data catalog play in enabling generative AI accuracy? A governed data catalog supplies AI systems with certified definitions, lineage, and documented business context, reducing hallucination risk and ensuring AI-generated outputs align with how the business actually defines its data.

The session "AI Pilots Are Easy. These Two Leaders Figured Out the Hard Part." was presented at the Gartner Data & Analytics Summit, Sydney, on 16 June 2026. Panellists were Chris Michael, Head of Analytics Products & Data Governance at TPG Telecom, and Dale Brimblecombe, Head of Data & Analytics at OceanaGold, moderated by Zach McIntyre, VP APAC at Alation.

To explore how Alation supports the foundations described in this session — including Semantic Model Mastering, CDE Manager, Data Products Marketplace, and Agent Studio — book a demo here or speak to your Alation account team.