In January I published seven predictions about where Microsoft's data and AI stack would go this year, and I promised I'd come back and mark my own homework rather than quietly let them expire. Build 2024 has just wrapped, which makes this the natural first checkpoint — five months in, one major conference of announcements to test my calls against. So here's an honest interim scorecard. Not a victory lap; a status check, including the bits that are aging badly.

The rule I'm holding myself to: I only get to grade against what's actually happened, not what I now suspect. A prediction isn't validated because it feels right in May. It's validated when something ships.

The ones that are clearly landing

Two of the seven are tracking almost embarrassingly well, and I'll take the easy points before the hard ones.

  • "Fabric becomes the default starting point." Everything about Build reinforced this. The messaging, the investment, the sheer gravitational pull of the announcements — Microsoft is treating Fabric as the centre of the data universe, and the field is following. I said greenfield builds would default to Fabric-first by year end; five months in, that's not looking bold, it's looking obvious. Which is a slightly uncomfortable thing to admit about your own "bold" prediction.
  • "Copilot goes from demo to daily driver for a narrow band of tasks." Copilot in Power BI reached general availability, and Build pushed Copilot deeper across the stack. Crucially, the way it's landing matches the prediction: it's genuinely useful for drafting and explaining, and still nobody sane is telling you to trust its plain-English answers unsupervised. The narrowness I predicted is holding. The hype has not collapsed the way it will need to, but the practical shape is right.

The ones that are too early to call

Two more are moving in the predicted direction but haven't earned a grade yet.

  • "Direct Lake turns out to be the most consequential feature." Still my sleeper pick, and Build gave it more polish and more surrounding capability. But "most consequential" is a claim about real-world adoption under real workloads, and it's simply too soon — the teams betting on it are still early, and the edge cases haven't finished revealing themselves. Ask me at year end. I'm still holding the bet; I just can't cash it yet.
  • "Capacity cost management becomes a boardroom conversation." The conditions are all assembling — more workloads on capacity, more ways to spend CUs — but the invoices that force the boardroom conversation take a couple of quarters to arrive and sting. This is a second-half-of-the-year prediction by nature. Watch this space; the bills are coming.

The one I'm quietly worried about

Here's where I'm honest against my own interest. My prediction that "the AI-replaces-the-analyst narrative peaks and cools" is the one I'm least sure of, and Build didn't help my case. If anything the narrative got louder, not quieter — every AI announcement carried a whiff of "you won't need the specialist for this anymore." I predicted a peak-and-deflate; five months in I can see the peak and no sign of the deflate.

Now, I still believe the deflation comes — I think it comes the expensive way, when enough organisations get burned by confident wrong answers. But "it'll cool eventually" is a weaker claim than the one I made, and if the cooling doesn't start showing by year end, I'll have to mark this one wrong. I'm flagging it now rather than hoping you forget I said it.

The point of grading your own predictions in public isn't to look clever in the ones you got right. It's to catch the one you're rationalising before it quietly becomes a lie you tell yourself.

What Build added that I didn't predict at all

Fairness cuts both ways, so let me note where the year is outrunning my imagination. The depth of the real-time and streaming investment at Build was more than I'd priced in back in January — Microsoft is clearly making "act on data as it arrives" a first-class part of Fabric, not a specialist corner. I under-weighted that, and I suspect it'll matter more by year end than my January self assumed. A good scorecard notes its blind spots, not just its hits.

And what the room made of it

A scorecard shouldn't only read the announcements; it should read the audience, and the community reaction to Build is itself evidence for grading my calls. Two things stood out in the practitioner threads. First, a note of Copilot fatigue is creeping in — not rejection, but a weariness at how every announcement now arrives wrapped in AI, and a hardening "show me it working on my real, messy data" posture. That supports my narrow-daily-driver call and, honestly, gives me a little hope for my nervous analyst-replacement prediction: the people who'd actually be replaced are pushing back on the framing, loudly, which is usually where a hype narrative starts to crack. Second, the cost anxiety is unmistakable — thread after thread asking what all these new workloads will do to capacity consumption. That's my boardroom-cost prediction assembling in real time, in the words of the people who'll be in that boardroom. The room, as ever, is a better forecaster than the keynote.

The two I haven't mentioned — and why they're the quiet engine

You'll notice I've said nothing yet about two of the seven: that governance would get a real budget, and that the semantic model would become the battleground. That's deliberate, because at the five-month mark they're the least visible and the most structural — the ones doing their work underneath the announcements rather than on stage.

  • Governance getting a budget doesn't announce itself at a conference; it shows up as line items and hires, slowly. What Build did was raise the pressure — more workloads, more sprawl, more surface to govern — without yet forcing the spend. It's a slow-burn call, and five months in it's smouldering exactly as I expected, not yet ablaze.
  • The semantic model as battleground is the one I'm quietly most confident in, because Microsoft keeps shipping features that only make sense if it's true. The direction of travel at Build — everything resolving back to a governed model, Copilot leaning on it for anything trustworthy — is my January thesis being built into the product itself. No grade yet, but the momentum is entirely my way.

Neither of these wins a scorecard in May. They're the two I expect to look most prescient in December, precisely because they're structural rather than flashy — and structural things move too slowly to demo and too surely to ignore.

Where this leaves the scorecard

Five months in: two landing, two too early but on track, two governance-and-cost calls waiting on the second half of the year to prove out, and one — the analyst-replacement deflation — that I'm watching nervously and may have to concede. Call it cautiously on track, with one flashing amber light I'm not going to pretend isn't flashing.

I'll do this again later in the year, and again at the end, because a prediction you never revisit is just marketing with a date on it. The value was never in being right in January. It's in being honest in May about which January calls the year is actually validating — and which one I might have talked myself into. See you at the next checkpoint.