At the ARC Retail AI Summit this week, there were lots of great numbers rattled off but this one really stood out for me: 93% of AI investment in retail goes to technology. 7% goes to process, people, culture and reskilling.
Irving Lee from Microsoft opened with the scale of it. Data centre capital expenditure hit US$930 billion in 2025, heading toward US$1.3 trillion by year's end, all in the space of six years. For context, Australia's entire gas and gold infrastructure build cost $77 billion over seven years. The NBN cost $54 billion over eleven. What's being poured into AI infrastructure right now dwarfs both, in a fraction of the time.
And it's landing. AI usage across retail is up 400% in eighteen months. Around 40% of use cases are making it into production, well above the commonly quoted figure of under 10%. Content generation running 50% cheaper and 200% faster. Call centre volume down 30%, saving Microsoft itself $800 million a year. Conversational commerce conversion up as much as 300%. The technology is working, in the narrow sense that it does what it's asked to do.
Then Leslie Lorenz and Travis Murphy from Snowflake showed what an agentic data platform looks like in practice, agents that surface insights, draft tasks and build pipelines without a person opening a query editor. David Phillips from Deloitte Digital went further still, walking through a demo of an 8,000-agent enterprise context layer, teams of agents debating a churn-recovery campaign and producing the finished thing, audience selection, product picks, visuals, styled models, in two and a half hours. Olivia Hitchens, who built the LeaseR platform for retail lease review, showed AI now doing lease analysis at roughly a tenth of the cost of a lawyer, and doing it well enough that she thinks ten lawyers with great AI will soon do the work of a hundred.
It's live, it's measured, and the return on investment is real. But it accrues heavily to the top 15% of companies executing well. The gap between the companies getting a return and the companies that aren't isn't a technology gap. Everyone in that room has access to roughly the same tools.
That's where I came in, on a panel about enabling people, leadership, workforce and productivity, alongside Ben Fish, Deloitte's Head of Organisation and Workforce Transformation, talking about the other 7%.
Companies roll out the technology, run some training, and then wonder why the productivity gains haven't shown up. The reason is almost never the model. It's that nobody redesigned the actual workflow the model was dropped into.
I talk about this as two separate gates, and most organisations have only opened one of them. Gate Two is capability, can people brief AI with intention, actually think strategically about what they're asking for, and evaluate what comes back with proper judgement. That's what most AI training teaches, and it's necessary. But Gate One is the one almost nobody names: legibility. Can your team describe how a piece of work actually gets done, consistently, the same way, by everyone doing it? If they can't, and in my experience most can't, then AI integration doesn't fail loudly. It fails confidently. Nobody catches the mistake because the thing being checked against was never solid to begin with.
The fix isn't complicated, which is part of why it gets skipped. A simple team AI charter, five lines written together rather than handed down: what we use AI for, what we don't, what we do when we're unsure, what data stays off it, who to ask. And one finding that surprised even me when I first came across it: the single biggest predictor of whether a team actually adopts AI well is whether their leader is visibly using it themselves. Not approving a budget. Using it, in front of the team, badly at first like everyone else. Reverse mentoring programs, pairing a CEO with a distinguished engineer, work for exactly this reason.
The fear conversation matters just as much, and it's where I think a lot of leaders get the tone wrong. You cannot argue someone out of being afraid for their job, and vague reassurance, telling a room AI will augment rather than replace, makes it worse, not better, because people can tell when a line is doing PR work instead of answering the actual question. What works is bringing people into the decision before it's made rather than managing the fallout after, and being specific about what an agent will and won't do rather than speaking in comfortable generalities. The useful reframe isn't "which roles do we cut," it's "what are the workflows, where does AI actually come in, and where does a human stay in the loop." That's a conversation people can engage with. The vague version isn't.
The clearest sign of this came from the technical talks themselves, without anyone intending it. Irving Lee named jagged intelligence as one of the three real blockers to scaling AI, the fact that it produces brilliant output and nonsensical output with the same confidence. That's Gate One, described from the infrastructure side. If your team can't tell the difference because they never agreed on what correct looks like in the first place, jagged intelligence isn't a model problem, it's a legibility problem the model is exposing. David Phillips flagged the same risk from the enterprise agent side: pockets of an organisation building their own agents without shared context governance leads straight to fragmentation. Again, Gate One, just with a different name on it.
Irving closed his own session with the numbers that back this up at the macro level. Australia is near the top of the world for consumer AI usage and retail experimentation, 70% of retailers here are experimenting with agents, nine points above the global average. But only 21% have actually scaled an agent into even one function, below the global average. We're enthusiastic starters and cautious finishers. That's not a criticism, it's a sensible instinct in a risk-averse sector. But it does mean the organisations that figure out the people side first are going to be the ones who convert experimentation into the top 15%, rather than staying stuck running pilots that never scale.
The infrastructure moment is real. The tools are extraordinary, and getting more capable by the month. But a 93 to 7 investment split doesn't fix itself just because the technology gets better. If anything, the gap gets more expensive to ignore the more capable the tools become, because the cost of a team that can't tell a brilliant output from a nonsensical one only goes up as the outputs get more convincing.
AI works when people do. That's not a soft add-on to the infrastructure story. Increasingly, it looks like the actual story.
Every engagement starts with a conversation
Start with a conversation