The plan was simple: leave in March, travel with the family, come home in the summer, pick the blog back up where I left it. The first two happened on schedule. The last one didn't, quite — the break stretched past its intended end and the blog stayed dark a good while longer than the itinerary promised. I won't dress it up as strategy. Sometimes life reorganises your calendar without asking, and the honest thing is to say so, close the gap gently, and get back to the work. So: I'm back, a little later than billed, and glad to be here.
What's strange about stepping away from this field for a few months is how little you miss and how much you don't. The announcements keep coming — there's always a new capability, a renamed service, a keynote promising the end of some old drudgery. But step back far enough and the noise separates from the signal on its own. You return and find that most of what felt urgent in the spring has already been forgotten, and the two or three things that genuinely mattered are still sitting there, patient, waiting for someone to do the unglamorous work. That filtering is the real dividend of distance, and it shapes the short list below. This isn't everything happening in data and AI for the back half of the year. It's the handful of things I think are worth your attention — and, just as usefully, a couple I've decided to stop giving mine.
1. The Fabric conversation has finally grown up
For two years the question people asked me about Microsoft Fabric was should we move? It was the wrong question, and I said so at the time. The interesting question was never whether to adopt the platform — that was mostly a matter of when — but whether you could run it well once you had. We've crossed that line. The organisations I talk to have stopped asking whether Fabric is real and started discovering, sometimes painfully, that a unified platform concentrates your mistakes just as efficiently as it concentrates your data.
That's the shift I'm watching: from adoption to operation. Capacity management, the dreaded "why did our CU consumption double overnight," workspace sprawl, the governance model you should have designed before anyone built a single lakehouse — these are the questions of a maturing platform, and they're far less photogenic than a launch demo. Good. The value was always going to live here, in the operational discipline, not in the migration itself. If you moved to Fabric in the last eighteen months and haven't yet built a real cost and capacity model around it, that's your H2. Not another proof of concept — a FinOps habit.
2. Copilot has to earn its keep now
The AI-in-the-data-stack story is entering its most useful phase, which is the phase where the honeymoon ends and someone finally asks what it actually delivered. I've been openly sceptical about the gap between what Copilot demos promise and what it does on a Tuesday with your real, messy semantic model — and I stand by that scepticism. But scepticism isn't dismissal. The tools have quietly got better, the obvious failure modes are better understood, and there are now genuine, narrow places where an assistant in the query editor or the report canvas saves a real person real time.
What I'm watching is whether organisations get disciplined about where. The mistake I expect to see repeated all through the second half of the year is the blanket rollout — Copilot switched on everywhere, ROI assumed rather than measured, and a licence bill nobody can quite justify when the finance team comes asking. The counter-move is boring and correct: pick two or three concrete workflows, measure the before and after, and expand only where the numbers hold. Treat it like any other tool that costs money and makes claims. The hype cycle wants you to adopt on faith. Don't. Adopt on evidence, and you'll end up using it more, not less — because you'll actually know it works.
3. The EU AI Act stops being a slide and starts being a deadline
This is the one I'd move to the top of any board agenda. For a while the AI Act was something everyone acknowledged and nobody had operationalised — a compliance slide, a problem for later. Later is arriving. The high-risk obligations are the ones with teeth, and they don't care how good your model is; they care whether you can document what it does, govern the data underneath it, and show your work when someone asks. That is a data-governance problem wearing an AI costume, and the organisations that treated governance as overhead for the last decade are about to discover the bill.
Here's the quiet part I keep saying to the people who own this: the AI Act is not a reason to slow down your AI ambitions. It's a reason to build them on foundations that don't collapse under scrutiny. Lineage, cataloguing, access control, a real record of how a decision got made — the same governance plumbing that makes your data trustworthy is the plumbing that makes your AI defensible. If you've been putting that work off because it never had a deadline, you now have one. I'd rather you did it because it's right, but I'll take "because the regulator is coming" if that's what finally moves the budget.
What I've stopped watching
An honest watch-list needs a discard pile. I've stopped tracking the endless renaming and repackaging of services that do roughly what last year's version did — the churn is real, the substance rarely is. I've stopped taking "AI-powered" as information; it now describes almost everything and therefore distinguishes nothing. And I've mostly stopped engaging with the argument about whether any single vendor's platform is the answer, because after enough of these cycles you learn that the answer is always the same: it depends on the problem, the team, and the constraints, and anyone who tells you otherwise is selling something. I remain, as ever, Microsoft-deep and vendor-honest — happy to recommend the stack I know best, and just as happy to tell you when it's the wrong fit.
The thread
If there's a single line running through all of this, it's that the unglamorous fundamentals — cost discipline, measured adoption, real governance — are exactly where the value has quietly relocated while everyone was watching the demos. That's not a coincidence, and it's not new. It's the same lesson every hype cycle eventually teaches the people still standing at the end of it. The difference this time is that the regulation, the licence bills, and the operational reality are all arriving at once to enforce it.
It's good to be back at the desk. The break clarified more than I expected, and most of what it clarified was this: the work worth doing was never the loud part. Let's get on with the quiet part.