By Brad Ferris · 2 May 2026
For two years the standard objection to AI investment has been some version of "show me the money". This week the money showed up, at least on the seller's side of the table.
Google Cloud reported US$20.03 billion in quarterly revenue, up 63 per cent year on year, the first time it has cleared US$20 billion in a quarter, with a cloud backlog of roughly US$460 billion. AWS grew 28 per cent to US$37.6 billion, its fastest rate in fifteen quarters, and told analysts that its Bedrock AI service processed more tokens this quarter than in all prior years combined. Microsoft disclosed an AI business running at a US$37 billion annualised rate, up 123 per cent year on year.
Those are audited, reported, on-the-record numbers. Whatever you think of the capex race funding them, the claim that nobody is making revenue from AI no longer survives contact with the quarterly filings.
Here is the uncomfortable part. In the same week, the buyer-side research painted close to the opposite picture.
The Stanford AI Index 2026 found that 89 per cent of enterprise AI agent projects never reach production, with US$150,000 to US$800,000 wasted per failed implementation. Stanford's framing matters as much as the number: the dominant failure modes are operational and economic, not technical. Data plumbing, integration with existing systems, exception handling, governance. The model is rarely the thing that breaks.
Then there is Writer's 2026 enterprise AI survey of 2,400 leaders and knowledge workers, which found 79 per cent of organisations facing adoption challenges, a double-digit increase on last year, and 75 per cent of executives admitting their AI strategy is "more for show" than genuine internal guidance.
Both sets of numbers are true at the same time. The platforms are converting AI into revenue at extraordinary rates. Most of the businesses buying from them are not.
The gap is not a capability gap. The models available to a 40-person Brisbane distributor this week are the same models available to a Fortune 100 bank. The gap is a capture gap, and it lives in the unglamorous layer between the technology and the profit line.
A model subscription is the cheapest item on the invoice. The expensive items never appear on an invoice at all: defining the process you want improved, getting the underlying data into a usable state, redesigning the workflow around the tool, deciding what happens when the tool gets something wrong, and training the three people whose jobs change as a result. Stanford's data says that is where the 89 per cent die. Writer's data says most leadership teams have not honestly planned for any of it.
For a mid-market operator this is oddly good news. The failure rate is not a verdict on the technology. It is a verdict on how it gets deployed, and deployment discipline is something a smaller business can do better than a large one, because you can see your whole operation from one seat.
Four practical moves, in order.
Stop relitigating whether AI works. That debate ended with this earnings cycle. The live question for your business is narrower and more useful: which specific process, with which measurable baseline, is worth improving first.
Pick one process with a number attached. Quote turnaround time, invoice processing cost, first-response time on service tickets. If you cannot state the current baseline, fix that before buying anything.
Budget for the boring parts. If your plan allocates money for licences and nothing for data cleanup, process definition and training, you are budgeting to join Stanford's 89 per cent. A rough rule from the failure data: the change work costs a multiple of the software, not a fraction of it.
Kill zombie pilots. If you ran a trial in the last year that neither died nor shipped, make a decision on it this month. An undecided pilot consumes attention and teaches your team that AI projects are theatre, which is precisely the "for show" pattern Writer's survey describes.
The sellers have proven their side of the trade. The 2026 story for operators is proving the other side, one measured process at a time.
Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.