It's January, the season when everyone with a keyboard and an opinion publishes their predictions, and most of them are hedged so carefully they can never be wrong. So let me do the riskier version: seven specific calls about where Microsoft's data and AI stack is heading this year, written now, in the open, with a date attached — so that in December you can come back and mark my homework.
I'm making these as someone who architects on this stack for a living, not as an analyst watching from the stands. The bias that comes with that is worth naming: I'm closer to what breaks than to what sells. So where the marketing says "revolution", I'll usually be the one asking what it costs and who has to maintain it at 2am. Here's my bet on the year.
1. Fabric stops being an add-on and becomes the default starting point
Microsoft Fabric went generally available in November, and the tempting way to treat it is as one more product to evaluate — against Synapse, against Databricks, against the warehouse you already run. My prediction: by the end of 2024 that framing flips. For any greenfield analytics build on Azure, Fabric becomes the default — the option you have to argue your way out of, not into.
Not because it's finished. It isn't. But because Microsoft has clearly decided this is the front door, and the entire field-and-partner machine will push everyone through it. The real question for the year isn't "should we look at Fabric" — it's "what do we deliberately keep outside it, and can we say why out loud."
2. Copilot goes from demo to daily driver — for a narrow band of tasks
Every keynote right now shows Copilot writing a finished report from a single sentence. That's the demo. The reality this year will be smaller and far more useful than that: Copilot becomes a genuine daily tool for a narrow set of jobs — drafting a first-pass DAX measure (the formula language behind every Power BI metric), explaining a query someone else wrote, summarising what's actually in a model — and stays unreliable at the thing everyone's selling, which is "ask a question in plain English and trust the answer."
The teams that win with it in 2024 are the ones who treat it as a fast junior who needs checking, not an oracle who replaces the checking. Copilot for Power BI reaches general availability this year and lands exactly there: real productivity on the boring parts, and a hard human gate on anything a decision rests on.
3. Direct Lake quietly turns out to be the most consequential feature
While Copilot takes the headlines, my pick for the change that actually reshapes how we design is Direct Lake — Fabric's mode that lets a semantic model read straight from the Parquet files sitting in OneLake, with no import and no scheduled refresh, at something close to import-mode speed.
That sounds like plumbing. It isn't. For fifteen years the fundamental tension in Power BI has been import (fast, but you're copying the data and refreshing it on a schedule) versus DirectQuery (live, but often painfully slow). Direct Lake is a serious attempt to dissolve that trade-off entirely. If it holds up under real workloads — and that "if" is the whole story of the year — a lot of refresh windows, a lot of duplicated storage, and a lot of 6am "the refresh failed again" calls simply stop existing. I think it delivers on enough of that to change default designs, even as the edge cases keep us honest.
4. Governance stops being optional — this is the year Purview gets a budget
Here's a less fun one. Fabric makes it dramatically easier to create a workspace, spin up a lakehouse, and land data. Anything that makes creation easy makes sprawl easy. By mid-year, organisations that adopted Fabric with enthusiasm will look up and find they can no longer answer basic questions: what data do we actually have, who owns it, and which of these forty semantic models is the one finance is allowed to quote.
So 2024 is the year governance moves from a slide nobody funds to a line item somebody owns. Microsoft Purview — classification, lineage, a proper data catalogue, sensitivity labels — becomes the reluctant necessity, bought not because anyone fell in love with governance but because the platform's own ease-of-use forced the question.
Every capability that makes it easier to create data quietly raises the price of not governing it. Fabric is exceptionally good at making it easy to create data.
5. The semantic model becomes the real battleground
As data spreads across lakehouses, warehouses, and notebooks, the question "where is the single version of the truth" gets louder, not quieter. My call: the semantic model — the layer where you define what "revenue" and "active customer" actually mean — becomes the contested centre of gravity this year.
This is the fight I care about most, because underneath it isn't a technical fight at all. Two teams with two definitions of "churn" don't have a tooling problem; they have an unresolved disagreement wearing a tooling costume. The organisations that treat the semantic layer as a governance and definition problem — a negotiation between people about meaning — will get value from all the shiny compute beneath it. The ones that treat it as a modelling detail will industrialise their disagreements faster than ever, and call the result a platform.
6. Capacity cost management becomes a boardroom conversation
Fabric runs on capacity: you buy a pool of Capacity Units — CUs, the single currency everything in Fabric spends — and every workload draws from that pool. The model is genuinely clever. It smooths short spikes and lets you burst above your baseline for brief periods instead of forcing you to size for your worst minute. It is also genuinely easy to misjudge.
My prediction: by the second half of the year, "why is our capacity throttling" and "why did we buy an F64 when an F16 would have carried us" become questions that reach finance, not just the platform team. Cost stops being something you reconcile in arrears and becomes a design constraint you plan around from the very first workspace. The teams that learn to read a capacity metrics report early will look prescient by Q4 — and the ones that don't will get a memorable invoice.
7. The "AI replaces the analyst" narrative peaks — and cools
Finally, the one I'd be glad to be wrong about. The loudest story right now is that generative AI makes the analyst redundant: everyone just asks questions in plain language and the machine answers. I think that narrative peaks in 2024 and then quietly deflates — because organisations discover, the expensive way, that a confident wrong answer is worse than no answer, and that "plain language in, trustworthy number out" depends entirely on the unglamorous layer underneath (the agreed definitions, the governed model, the lineage) that AI doesn't provide and can't fake.
The job doesn't vanish. It moves. It becomes less about producing the number and more about guaranteeing the number means what the person asking thinks it means. As the counting gets automated, the meaning becomes the whole game — and reading meaning is a human's job for a good while yet.
What ties these together
Read the list back and one thread runs through all of it. Every prediction is a version of the same shift: the platform keeps getting easier, and the human judgement keeps getting more valuable. Fabric makes building easy — so the scarce skill becomes deciding what to build and what to govern. Copilot makes producing easy — so the scarce skill becomes knowing when to trust the output. Direct Lake makes the plumbing disappear — so the questions that remain are the ones that were always genuinely hard.
That's my year in seven bets. Some will age well; some will embarrass me — which is exactly why it's worth writing them down with a date on them instead of hedging into mush. I'll come back at the end of 2024 and grade myself honestly, because the entire point of a prediction is that it can be wrong, out loud, in public. See you in December.