This is my second check-in on the seven predictions I made in January, and I want to do something different from the last one. In May, just after Build, I graded them like a scorecard — landing, too early, worried. Useful, but a scorecard misses the more interesting thing that happens when you live with your own predictions for nine months: they evolve. You start to see what they were really about, underneath the specifics. So this is less "am I winning" and more "what have these calls taught me about the year they were trying to describe."

And the honest answer is that all seven are quietly turning out to be the same prediction, wearing seven costumes.

The single thread I didn't fully see in January

Here's what I can see now that I couldn't quite articulate at the start of the year. Every one of my predictions was, underneath, a claim about a widening gap — the gap between how easy it's becoming to produce something with data, and how hard it remains to trust it.

Look at them through that lens and they line up:

  • Fabric makes it trivially easy to produce data platforms and workspaces — and I predicted governance would have to catch up, because trust didn't get easier at the same rate.
  • Copilot makes it trivially easy to produce answers and measures — and I predicted the value would concentrate in the narrow band where a human can still verify, because trust is the bottleneck.
  • The semantic-model battleground is entirely about this gap: producing a number is easy, agreeing it means the right thing is hard.

I predicted seven consequences. What I was actually predicting, without quite naming it, was that in 2024 the production of data work would race ahead and the trustworthiness of it would lag, and that every interesting problem this year would live in the widening space between the two. Nine months in, that's the through-line, and it's sharpened rather than softened.

The prediction that's validating hardest

If I had to point to the call aging best, it's the semantic-model one — and the evidence is that Microsoft keeps shipping products that only make sense if you believe it. The Metrics Layer that just landed is the clearest example: a whole feature built on the premise that the definition of a number is a first-class governed asset worth standardising. You don't build that unless the definitions are where the value and the pain concentrate. Microsoft's own roadmap is arguing my January case for me. I'll take it.

The prediction I'm still nervous about

And the one still keeping me honest is the same one that worried me in May: that the "AI replaces the analyst" narrative would peak and cool. Three-quarters through the year, I can report the peak has been magnificent and the cooling has been… theoretical. The narrative is still loud.

But I've refined why I still believe the deflation comes, even later than I hoped. It comes precisely because of the gap I described above. The louder the "just ask the AI" story gets, the more organisations wire production-of-answers directly to non-experts — and the wider the trust gap yawns underneath them. That's not stable. At some point a confident wrong answer drives a visibly wrong decision, and the correction begins. I mispredicted the timing of the cooling. I'm more convinced than in January about the mechanism. Whether that counts as right or wrong I'll settle honestly at year end.

Nine months of watching your own predictions teaches you something uncomfortable: the specifics you were proud of matter less than the pattern you didn't notice you were describing.

What's changed in how I'd advise

The practical upshot — because I don't write these to keep score, I write them to sharpen how I actually advise people — is that my counsel this year has narrowed to one sentence I keep repeating in rooms: invest in the trust side, because everyone's already over-invested in the production side.

The platforms will keep making it easier to produce. That race is won and it isn't yours to win. The scarce, defensible capability — the thing worth your budget and your best people — is everything that closes the trust gap: the governed definitions, the semantic layer, the classification, the human judgement about when an answer is safe to act on. That's where I'd have told you to spend in January if I'd seen the thread clearly, and it's what nine months of my own predictions evolving has made unmissable.

If I rewrote the seven today

Here's a test I find clarifying: if I could go back and rewrite January's list knowing what I know now, what would change? Less than I expected, and in an instructive direction.

I wouldn't remove any of the seven — every one still stands. But I'd collapse them. Three of my predictions — governance getting a budget, the semantic-model battleground, the capacity-cost reckoning — turn out to be one prediction: the trust-and-control side of the stack becomes where the money and the difficulty concentrate. And three others — Fabric as the default, Copilot as daily driver, Direct Lake as the quiet game-changer — are also one prediction: the production side gets radically easier and more consolidated. The seventh, the analyst-replacement deflation, is simply the collision between those two forces working itself out.

So if I rewrote it today, I'd make two big, opposed bets instead of seven medium ones, and I'd state the tension between them openly as the actual forecast. That's not hedging my way out of the original list. It's nine months teaching me that I'd chopped one clear idea into seven pieces and slightly hidden it from myself in the process — which is, I suppose, exactly what a check-in is for.

I'll do the final grading at year end, properly, with the honesty I promised — including on the one I might have to concede. But the real value of these check-ins has turned out not to be the marks. It's that they forced me to notice what I was actually claiming, and to say it more plainly each time. In January I made seven predictions. By September I think I was making one, and I'm finally able to state it: the platform got great at producing, the hard problem moved entirely to trusting, and the whole game is now played in the gap between.