In January I made seven predictions about Microsoft's data and AI stack and promised to grade them honestly at year end — not on a curve, not quietly dropping the embarrassing ones. It's December. Time to pay up. I've checked in twice already, once after Build and once in September, and each time I got a little more honest about what I was actually claiming. This is the final reckoning, prediction by prediction, with a real grade on each and no wriggling.
The grades
1. Fabric becomes the default starting point. → Right. This one aged from "bold" to "obvious" inside six months. By year end, Fabric-first is the assumed posture for greenfield analytics on Azure, and the interesting conversation everywhere is what to deliberately keep outside it. Comfortably correct, and honestly not that brave in hindsight.
2. Copilot goes from demo to daily driver for a narrow band of tasks. → Right. Copilot in Power BI reached GA, spread across the stack, and landed exactly where I said it would: genuinely useful for drafting, explaining, and summarising, and still not trustworthy as an unsupervised oracle. The narrowness held. The people getting value are the ones treating it as a fast junior, not a replacement.
3. Direct Lake turns out to be the most consequential feature. → Partly. It matured, adoption grew, and for the teams using it well it did quietly dissolve the old import-versus-live trade-off. But "most consequential" was a bold superlative, and I can't honestly claim it beat Copilot or the governance push for sheer impact this year. Right direction, over-claimed magnitude. Half marks, honestly earned.
4. Governance gets a real budget. → Right. The Purview GA, the arrival of posture management, the general acceptance that Fabric's ease-of-creation demands governance to match — governance stopped being the slide nobody funds and started being a line item with an owner. Correct, and I suspect this one keeps compounding into next year.
5. The semantic model becomes the real battleground. → Right, and the one I'm proudest of. The Metrics Layer was Microsoft building my thesis into the product: proof that the definition of a number is now treated as a first-class, governed, contested asset. Everything trustworthy — including Copilot's answers — kept resolving back to the governed model. Called it.
6. Capacity cost management becomes a boardroom conversation. → Partly. The invoices started biting and capacity planning became a real discipline, but I over-stated how far it climbed. It reached the platform team and finance; it didn't quite reach the board in most places. Right that it became a serious conversation, wrong that it became a top-table one. Half marks.
7. The "AI replaces the analyst" narrative peaks and cools. → Wrong. I have to take the loss cleanly. I predicted the peak-and-deflate; we got the peak and no deflation. The narrative was as loud in December as in January. I still believe the cooling comes — I refined why across the year — but "it'll happen eventually, later than I said" is not the prediction I made. I said 2024. It didn't. Wrong, and I'm marking it wrong.
The thing I completely missed
Fairness demands I grade my blind spots as harshly as my hits, and here's the biggest one: I said essentially nothing in January about sustainability data. Then Microsoft leaned into sustainability-focused data solutions — the ability to bring environmental and emissions reporting into the same governed analytics estate as everything else — and it turned out to be a more substantive thread than my January self imagined. I didn't predict it because I under-weighted how fast regulatory and reporting pressure would make "can we actually measure and report our environmental data" a real board-level question with a real Microsoft answer attached. A genuine miss, and worth naming.
A prediction scorecard is only worth anything if you grade the miss as loudly as the hits. Five-and-two-halves right feels good; the honest headline is that I got the loud one wrong and didn't see the sustainability one at all.
Why the one I got wrong was always the hardest
It's worth dwelling on prediction seven, because which one I got wrong is instructive. The six I got right or half-right were all predictions about technology — what would ship, what would consolidate, where the difficulty would move. Technology predictions are relatively tractable, because roadmaps have momentum and platforms move in visible directions. If you watch the signals, you can see them coming.
The one I got wrong was the only prediction about human behaviour at scale — how a narrative would rise and fall in the collective imagination. And that's a genuinely harder thing to forecast, because narratives don't obey roadmaps. A hype cycle cools when enough people get burned badly enough to change the conversation, and "enough people" and "enough burning" run on a timeline nobody controls. I predicted the mechanism correctly and the timing not at all — which in forecasting terms just means wrong. The lesson I'm banking: be far more humble about calling the timing of a mood, however confident I am about its eventual direction. Technology has momentum you can read. Sentiment has its own clock.
The tally, and what it taught me
So: three clear rights, one proud right, two half-rights, one clean wrong, and one significant blind spot. On a generous count that's a good year of forecasting. On an honest one it's a reminder that the calls I was most confident about (Fabric, the semantic model) held, the one I most wanted to be true (the AI-hype deflation) didn't, and the future had a whole category of surprise (sustainability) I wasn't even looking at.
The real lesson isn't the score. It's what the check-ins taught me along the way: that my seven predictions were really two — the production side getting radically easier, the trust side staying hard — and that everything interesting happened in the widening gap between them. That's the frame I'll carry into next year's predictions, which I'll write in January and grade again next December, wrong ones and all. Because a forecast you're not willing to mark wrong in public was never a forecast. It was just a nicer way of saying what you hoped.