AI is finally actually just business

Ai4 doubled in size this year, and the crowd that filled The Venetian wasn't there to speculate – it was there to buy

Las Vegas is a strange place to go looking for proof that something is real, but that’s where I found it last week. Ai4, which started in 2018 as a 300-person gathering in Brooklyn and has since grown into North America’s largest AI industry conference, took over The Venetian this year with attendance that roughly doubled year-over-year – enough growth that it had to abandon last summer’s home at the MGM Grand just to fit everyone in. I was there moderating two panels, one on the agentic bank and one on the agentic brain, and I went in braced for the usual conference cocktail of frontier-model hype and hand-wavy futurism. What I got instead was three days of people talking about procurement.

The composition of the crowd was the first tell. The badges weren’t dominated by researchers or founders hunting their next round – they belonged overwhelmingly to IT and data leaders from banks, insurers, hospital systems, manufacturers, retailers, utilities and government agencies. These are the people who sign enterprise software contracts, and they showed up by the thousands, from organisations that would never describe themselves as tech companies. The programming followed the buyers: this year’s edition added 20 new industry and job-function tracks covering everything from financial services and healthcare to risk and compliance, energy and national security, and the enterprise speaking roster leaned on U.S. Bank, Mayo Clinic, Ford, PayPal and Uber rather than the frontier labs. Geoffrey Hinton, Fei-Fei Li and Andrew Ng shared a keynote stage – catnip for anyone craving the big existential debate – but the real business of the conference happened in the industry tracks and on an expo floor where roughly 400 vendors pitched to people carrying actual budgets and actual requirements documents.

My own sessions were a decent microcosm. The agentic bank panel could easily have drifted into science fiction – autonomous finance is exactly the kind of topic that invites it – but the conversation stayed stubbornly grounded in deployment mechanics: governance and auditability, how agents interface with core systems older than most of the people building on top of them, and where the returns actually showed up first. The agentic brain discussion followed the same arc, concerned less with grand claims about machine cognition than with the orchestration layer organisations are building to coordinate fleets of agents without losing control of them. And across both, the vocabulary that kept surfacing was the vocabulary of unit economics: ROI, token efficiency, right-sizing models to tasks, cost per resolved ticket or processed claim. Nobody in those rooms was buying ‘intelligence’ – they were buying outcomes at a price, and the audience questions made it clear they’ve learned to shop accordingly.

This isn’t just my read on the show floor, either. Futurum Group’s survey of 830 enterprise IT decision-makers earlier this year found that direct financial impact – revenue and profitability – nearly doubled as the primary way organizations measure AI success, displacing softer productivity metrics, while agentic AI surged more than 30% year-over-year as a top technology priority. Their conclusion was blunt: “the pilot phase of enterprise AI is over.” It’s a remarkable turnaround from where the narrative sat almost exactly a year ago, when MIT’s NANDA project released its now-infamous ‘GenAI Divide’ report – the one claiming 95% of enterprise AI pilots were delivering zero measurable return – in the same mid-August window as last year’s Ai4. That statistic did a full lap of every boardroom and earnings call in the western hemisphere, and it remains the load-bearing citation of the bubble thesis today. Twelve months later, the same conference doubled in size, packed with exactly the buyers that report said were getting nothing for their money.

None of this means every deployment works. Gartner still expects a big share of agentic AI projects to get cancelled over the next couple of years thanks to fuzzy ROI and weak controls, and they’re probably right – but that’s a statement about enterprise software, not about a speculative technology. ERP implementations have failed at spectacular rates for thirty years and nobody argues SAP is a bubble; failure inside a category is just what a category looks like. The bubble conversation, meanwhile, is genuinely live in the capital markets, and I don’t want to hand-wave it away. Fund managers have spent the past year ranking an AI bubble among their biggest tail risks, Goldman’s Jim Covello is still asking whether enterprises actually make or save money with this stuff, analysts are drawing open comparisons to 1999, hyperscaler capex is on track to clear half a trillion dollars this year, and the whole market is white-knuckling its way toward Nvidia’s earnings at the end of the month. Those fears aren’t crazy. Valuations at the model and infrastructure layer may well be ahead of themselves, and the circular financing arrangements between the biggest players deserve every ounce of scrutiny they’re getting.

But the bubble question and the utility question are two different questions, and conflating them is probably the most common analytical error in AI discourse right now. Financial bubbles form around real infrastructure all the time – railways, telegraph lines, fibre optic cable. The dot-com crash wiped out trillions in market value and did precisely nothing to slow the internet’s absorption of commerce, media and communication. If AI valuations correct, some shareholders will have a very bad year and a lot of overfunded startups will die – and not a single bank will un-deploy the agents working its loan operations, because those agents will still be cheaper and faster than whatever came before them. Enterprise budgets, once allocated to a category that demonstrably works, don’t get un-allocated; they get renewed, renegotiated and consolidated. Bubbles are about prices. Business is about budgets. Budgets are stickier.

What I watched at The Venetian – fittingly, a full-scale simulation of somewhere real – was AI completing the least glamorous and most important transition any technology makes: from spectacle to infrastructure. Cloud ran this exact route between roughly 2010 and 2016, from ‘is it even secure?’ to a default line item nobody debates, and the buyers wandering Ai4’s expo floor are now treating model selection the way they learned to treat cloud provider selection – as a vendor decision, with all the diligence and pleasant boredom that implies. The frontier labs will keep supplying the fireworks, and the bubble discourse will keep churning with every earnings report and capex announcement. But the quieter, more durable story is that AI has crossed over from ‘will this be real?’ to ‘how do we run it?’ – and no technology category, once across that line, has ever crossed back. The robots may or may not come for the S&P 500’s multiple. They’ve already made it onto the purchase order.