Last January I made seven predictions about Microsoft's data and AI stack, promised to grade them honestly, and did exactly that in December — three clear rights, a proud one, two half-rights, one clean wrong, and a sustainability blind spot I never saw coming. The exercise was useful enough, and uncomfortable enough, to repeat. So here's year two: seven specific calls for 2025, written now, in the open, with a date on them, so that next December you can hold me to them.

Same standing bias, still worth naming: I architect on this stack for a living, so I'm closer to what breaks at 2am than to what sells from a keynote stage. Where the marketing says "revolution," I'll be the one asking what it costs, who governs it, and whether it survives contact with a real workload. Here's my bet on the year.

1. 2025 is the year of agents — and the year most agent projects quietly stall

The loudest story going into this year is agents: AI that doesn't just answer but acts — takes multi-step actions on your behalf across your data and tools. Microsoft is pushing hard here, and the capability is real. My prediction is two-sided: agents genuinely arrive as a usable pattern in 2025, and the majority of ambitious agent projects stall before production — not because the AI can't act, but because the underlying data, permissions, and governance aren't in a state where you'd dare let something act autonomously. The demos will be dazzling. The blocker will be the same unglamorous foundation it always is.

2. Governance stops being a project and becomes a precondition

Last year governance got a budget. This year it gets promoted — from a thing you do to a thing you must have before the interesting work is allowed to start. The reason is prediction one: you cannot responsibly let agents act on data you can't classify, whose lineage you can't trace, whose access you haven't controlled. So Purview and the whole governance layer stop being the reluctant back-office purchase and become the gate the AI ambitions have to pass through first. The organisations that governed early in 2024 will spend 2025 building; the ones that didn't will spend it retrofitting, and calling it a platform problem.

3. The AI Act moves from "in force" to "actually felt"

The EU AI Act became law in 2024, and this year its first real obligations start to bite — the prohibited-use bans, the early phases of the timeline. My call: 2025 is when the Act stops being a legal abstraction data teams could defer and becomes a live input to design decisions, especially for anyone touching high-risk domains. Not universal panic — I've argued the panic is misplaced — but a genuine shift from "we should look at that" to "we can't ship this until we've tiered it." Compliance quietly becomes an architecture constraint, not a slide.

4. Fabric's centre of gravity moves from analytics to application data

Fabric spent 2024 becoming the default analytics platform. My 2025 prediction is that it starts reaching beyond analytics into operational territory — the database and application-data workloads that used to live entirely elsewhere. The pitch becomes "keep even more of your data estate in one governed place," and the interesting architectural tension of the year is how far you actually want to pull operational workloads onto an analytics-born platform. I think the reach is real and the right answer is "further than sceptics expect, not as far as the marketing implies."

5. Copilot's second act is boring, and that's the point

The 2024 story was Copilot arriving. The 2025 story is Copilot maturing — and maturity looks boring on purpose. Fewer jaw-dropping demos, more quiet reliability on the narrow band of jobs it's actually good at: drafting measures, explaining models, summarising, accelerating the tedious middle of the work. My call is that the hype curve flattens while the actual usefulness keeps climbing, and the gap between what Copilot is sold as and what it's genuinely good for finally starts to close — not because the marketing gets humbler, but because users stop believing the demo and start using the tool.

6. Capacity cost becomes a genuine discipline, with tooling to match

Last year I predicted capacity cost would reach the boardroom and had to mark it half-right — it reached finance and the platform team, not quite the top table. This year I'll refine rather than repeat: 2025 is when Fabric cost management matures into a real discipline, with better native tooling, clearer patterns, and named owners — a proper FinOps practice around Capacity Units rather than the reactive bill-shock of the early adopters. The organisations that learned to read a capacity metrics report in 2024 spend 2025 looking prescient. The rest get one more memorable invoice before they learn.

7. "AI replaces the analyst" finally cools — and I'm calling it again anyway

Here's my one repeat offence. Last year I predicted the "AI replaces the analyst" narrative would peak and cool, and I got it wrong — it peaked and stayed loud. I'm calling it again for 2025, with a crucial change to my reasoning: I'm no longer predicting the timing of a mood, which I proved I can't do. I'm predicting a mechanism — that as more organisations actually deploy this stuff, enough of them get burned by confident-wrong answers that the conversation shifts from "replaces the analyst" to "needs the analyst more than ever, just for different work." It may still run late. But the direction is only becoming more certain, and I'd rather be wrong twice out loud than quietly drop the call I still believe.

What ties these together

Read the seven back and last year's thread has, if anything, tightened: the platform keeps getting more capable, and the human-and-governance layer keeps getting more decisive. Agents make action easy — so the scarce thing becomes the governed foundation that makes action safe. The AI Act makes rules explicit — so the scarce thing becomes people who can translate them into design. Copilot makes production easy — so the scarce thing, still, is the judgement to trust or reject what it produces. Every year the tools climb, and every year the value migrates further into the part the tools can't do.

That's 2025 in seven bets. Some will age well; at least one, going by my record, will embarrass me — which is the entire point of writing them down with a date instead of hedging into mush. I'll be back in December to grade myself honestly, wrong ones marked wrong, blind spots named. See you then.