Podcast — Episode 80: Rethinking AI in the Enterprise with Andrew Welch

The new episode of Leading IT — APAC Insights is live. This one's the deepest conversation we've had on the show this year. Joshua Rubens and I sat down with Andrew Welch — Microsoft MVP, founder of Cloud Lighthouse, executive director of the Centre for Trustworthy AI, and one of the most cited strategists in the Microsoft enterprise AI ecosystem — to work through what's actually happening when AI lands inside large enterprises in 2026, and what most leaders are getting wrong about it.

Listen here: pod.co/leading-it-apac-insights

Andrew has spent 25 years building large-scale enterprise technology in more than 12 countries — including, as he puts it, "famously on all 7 continents" (he once deployed Microsoft Dynamics CRM at the US Antarctic programme; even penguin relationship management is a thing). He now lives in Valencia and spends a lot of time with United Nations organisations working on trustworthy AI. He brings the kind of perspective you can only get from running the playbook at scale across very different regulatory and cultural contexts.

A few things from the conversation worth your time.

What we covered

1. The three categories of AI people in every organisation. Andrew breaks the workforce into optimists, head-in-the-sand ignorers, and the angry-LinkedIn doomers. The data behind the second group is the surprising one: in survey work he cites, 30–40% of people have admitted to actively sabotaging AI rollouts in their own organisations — most often because they're afraid it'll replace them. His counterpoint: the people getting promoted fastest right now are the ones who are figuring out how to use AI in their work. Whichever camp your team sits in, the way you roll the tech out has to be different for each one — and that needs executive ownership, not just sponsorship.

2. Why 80% of AI pilots are failing. Andrew normalised numbers from Harvard, Gartner and Boston Consulting Group and landed on roughly 80% as a defensible failure rate. The single biggest reason: organisations treat AI like ERP or CRM — a point solution you buy and implement. AI isn't a point solution. It's an ecosystem. "A point solution is about solving a known specific problem. What you're doing with AI is building an ecosystem, because the ecosystem is about data, security, and the AI thing itself, the tip of the iceberg. When you build an ecosystem, you're solving for an infinite number of unknown problems."

3. AI is a marble cake, not a layer cake. When I asked Andrew if AI is essentially a layer that sits across everything — operating system, network, AI on top — he pushed back firmly. "Don't think layer cake. Think marble cake. AI gets mixed through every part of the business; it isn't a tier you add on top." That single reframe changes how you scope, budget, and architect. Layer cake gets you bolted-on pilots that fail. Marble cake gets you an ecosystem where data, security, governance and capability all evolve together.

4. The strategy framework with five pillars and 25 dimensions. The Centre for Trustworthy AI publishes an open framework: strategy and vision · ecosystem architecture · AI workloads · responsible AI · scaling AI. Twenty-five sub-dimensions under those pillars. The point isn't to make you perfect at all of them before starting — "if you don't actually get on the bike, you're never going to learn how to ride the bike" — but to give leaders a structured way to assess where they are, and a way to target investment instead of spraying it. Worth using whether you're a 100-seat firm or the world's third-largest bank.

5. The "I want the car but don't touch my fuel system" problem. Andrew shared a quote from a senior decision-maker who told him: "I need you to figure out how to score the big AI wins, but I don't want you screwing with my data plumbing." His response: that's like asking for a car a hundred years ago but refusing to let the engineer touch fuel. You can't have AI outcomes without the data work underneath. Most "platform-first" AI investments stall because leadership refuses to fund the unglamorous foundations and then can't explain why the visible bits aren't paying off.

6. Mid-market is where the ecosystem-first argument gets sharpest. Mid-market firms (100–1000 staff) usually don't have a data lake, don't think they need one, and can't see why they should fund one when there's no immediate revenue lift. Andrew's argument: "In an organisation of that size you need a data lake more than the world's third-largest bank does, because you don't have the scale to be as wasteful. The bank has an essentially infinite pile of money to spend on 80 data lakes. You don't. You need to govern your data in one place, one time." If you're an AU mid-market CIO, this is the line you take into your CFO conversation.

7. Microsoft as glue, not model. Useful framing if you're choosing platforms in 2026. "OpenAI is going to continue building models. Anthropic is going to continue building models. Microsoft is going to build glue — that it sells to you to bring these models and capabilities together in an increasingly easier way." Andrew is open about being a Microsoft partisan, but his framing of Copilot as "the operating system for AI, the UI for AI — not a model itself, an orchestrator" is one most people get wrong about the Microsoft offering. It explains why so many "Copilot is bad at X" complaints miss the point.

8. Scientist AI as the under-discussed counterpart to agentic AI. Drawing on Yoshua Bengio's work, Andrew argues there's a road less travelled alongside the agentic-AI gold rush: AI as a brilliant research assistant working with a human on a hard problem, not autonomously. He's not against agents — he uses them — but he thinks the agentic-only narrative is missing a complementary pattern that, in his experience, generates substantial return without the same governance complexity.

9. Why the most regulated industries will win first. Counterintuitively, financial services, public sector, and other heavily-regulated organisations are emerging as the early AI winners. Not because they're more capable, but because their industries already make them take trust, security and compliance seriously. "The 'dog ate my homework' excuse — well, the AI wasn't regulated — doesn't fly. Your industry is still regulated. Those firms are organisationally predisposed to take this technology seriously."

The Paul Simon reference

Andrew's article "These Are the Days of Miracle and Wonder. AI is the long-distance call" (the title is a Paul Simon lyric from Graceland) was written after the UN's AI for Good Global Summit. The metaphor: just as the long-distance call shrank distance and changed what it felt like to be far from family, AI is the next moment where a previously expensive, scarce capability becomes ambient. The question for leaders isn't whether the technology will spread — it's what kind of organisation, culture and human experience you want to build around it.

The three things Andrew told us to do on Monday morning

1. Sit with the strategic question, not the prompt. "I'm worried AI is circumventing the human cognitive ability to sit alone with our thoughts. We've become more interested in the output of the prompt than the prompt itself." Go back to: what kind of company are we trying to build, what kind of world do we want to live in. Don't outsource that to a model.

2. Create quiet space for your own brain. Andrew and his wife each take a weekly day off where neither is parenting — explicitly to be alone with their thoughts. His company is rolling out a "thinking retreat" budget for staff, with the rule that you can't bring your kids and you can't go drinking with friends — just space to think. Whatever the form, the principle is to defend the quiet time AI is quietly eroding.

3. Leaders: build a deliberate strategy, architecture, and roadmap. You will not succeed by accident. Don't get caught in the pilot death spiral where the loudest internal need wins this quarter and the loudest different one wins next quarter. Have a forward-looking strategy, an ecosystem you're deliberately evolving, and a roadmap your team can act on.

The frame I'm taking away

The bit that landed hardest for me: most "AI strategy" documents I see in 2026 are still arguing about which tools, which use cases, which roll-out plan. Andrew's case is that those are the second-order questions. The first-order question is whether the technology is amplifying your people's judgment or quietly hollowing it out. Three years of "deferring to the model" and you've corroded the muscle the organisation actually runs on. Productivity in the abstract is the wrong goal. Human flourishing, exploration and culture are the right ones — and the technology choices follow from there.

If you're sizing up where AI fits in your organisation past the pilot phase, book a discovery call and we'll work through your specific situation. The patterns from Andrew's international enterprise lens map across to the Australian mid-market more cleanly than you'd expect.

---

Listen to Episode 80 on Leading IT — APAC Insights. The show covers IT strategy, AI, cloud, security, vendors and people leadership for CTOs, CIOs and IT leaders across the APAC region.

About the guest: Andrew Welch is founder, managing partner and chief strategy officer of Cloud Lighthouse, and executive director of the Centre for Trustworthy AI. He writes extensively on enterprise AI and cloud strategy; his best-known works include Crafting Your Future-Ready Enterprise AI Strategy, The Frontier Campus, Ecosystem-Oriented Architecture, and The Cloud Strategist. Find him on LinkedIn.