Halfway through 2024, I can see the shape of the year clearly enough to name it: this is the year generative AI stopped being a demo on the data stack and became a thing organisations actually switched on — Copilot in Power BI, Copilot across Fabric, natural language pointed at the enterprise's own data at last. And it's the year a great many of those organisations ran headlong into a lesson that, in hindsight, was always waiting for them: an AI that can answer any question is only as good as the data model it's answering from, and most data models are not nearly as good as their owners believed. 2024 is the year Copilot met the data model — and the meeting has been, for a lot of people, a humbling one. Back in January I predicted the semantic model would become the real battleground of the year; I underestimated how fast, and how painfully, that would come true.

The promise, and the thing the promise hides

The promise of Copilot on your data is intoxicating, and I don't say that dismissively — it's genuinely a remarkable capability. Ask a question in plain English — "what were our top-performing regions last quarter, and how does that compare to the year before?" — and get an answer, a chart, a narrative, in seconds, without knowing a query language or waiting for an analyst. For a business that has spent years feeling locked out of its own data by technical gatekeeping, this is close to magic, and the demos are rightly dazzling.

But the demo hides the thing that actually determines whether it works, and the thing it hides is everything. When Copilot answers "what was our revenue," it isn't divining truth from the ether. It's reading your semantic model — the layer where someone, at some point, defined what "revenue" means: which transactions count, how returns are handled, when it's recognised, which currency, which filters. Copilot's answer is only as correct as that definition. If your model defines revenue well, Copilot gives a business user a trustworthy number in seconds, and that's transformative. If your model defines it badly, or three ways, or not at all — Copilot gives them a confident, well-formatted, completely wrong number in exactly the same seconds, with exactly the same authority. And a confident wrong answer, delivered beautifully and trusted instantly, is far more dangerous than no answer at all.

That's the meeting. The AI doesn't fix your data model. It broadcasts it — amplifies whatever was already there, good or bad, to every person who can type a question. And 2024 is the year a lot of organisations found out, in front of an audience, exactly what their data model actually said.

Why the AI makes the model matter more, not less

There's a seductive misreading of generative AI that runs the other way — the hope that AI would make the tedious data-modelling work unnecessary, that a clever enough model could just make sense of whatever mess you pointed it at, and free us from the discipline of definitions and governance. I understand the hope. It is exactly backwards, and 2024 is proving it backwards in the field.

Before Copilot, a bad data model was a contained problem. The damage was limited by the fact that only a handful of technical people ever touched the raw data directly, and they knew where the bodies were buried — they'd mentally correct for the quirk in the revenue table, remember that this measure double-counts, quietly route around the definitions that were wrong. The messy model was survivable because expertise stood between it and everyone else, absorbing its errors.

Copilot removes that buffer. It puts the data model in direct conversation with people who don't know where the bodies are buried, who have no way to sense that an answer is wrong, and who — reasonably — trust the confident machine. So every flaw in the model that used to be quietly absorbed by an expert now flows, unfiltered and authoritative, straight to a decision-maker. The AI didn't make the model matter less by making querying easy. It made the model matter enormously more, by removing every human who used to compensate for its flaws. The scarce, load-bearing thing is no longer the ability to query — the AI democratised that. It's the correctness of the thing being queried, and that was never democratic and never easy.

A worked example, because this is too easy to keep abstract

Let me put a real shape on it. Imagine a company where "active customer" means three subtly different things: to marketing it's anyone who opened an email in ninety days, to finance it's anyone who paid in the last quarter, and to the product team it's anyone who logged in this month. None of these is wrong. They're three legitimate answers to three slightly different questions, and for years the organisation lived with the ambiguity perfectly well — because the only people who ever queried the data directly were analysts who knew the ambiguity existed and would ask "active by which definition?" before answering. The expertise absorbed the collision.

Now switch on Copilot. A regional manager types "how many active customers do we have in the north?" and gets a number — instantly, confidently, formatted beautifully. She has no idea there are three possible answers. She doesn't know to ask which one she got. She takes the number into a meeting and makes a decision on it. Across the hall, someone in finance asks Copilot the same question, gets a different number from a different measure, and takes that into a different meeting. Both are "right." Both are trusted. And the organisation is now making decisions on silently contradictory figures, at machine speed, with total confidence — a situation strictly worse than the old one, where at least the ambiguity was visible to the experts who handled it.

Nothing about Copilot malfunctioned here. It did exactly what it was asked, against exactly the model it was given. The failure was entirely upstream, in a definitional disagreement the organisation had comfortably tolerated for years because expertise kept it contained — and which the AI, by removing the expertise from the loop, turned loose on everyone at once. That's the meeting, in miniature. The AI didn't create the problem. It industrialised one the organisation already had and had chosen not to resolve.

What the good 2024 stories have in common

I've watched organisations have genuinely good experiences with Copilot this year, and genuinely bad ones, and the dividing line is boringly consistent. It has almost nothing to do with how they configured the AI and almost everything to do with the state of what sat beneath it.

The organisations getting real value from Copilot in 2024 are the ones who had already done the unglamorous work: a well-governed semantic model with agreed definitions, clear ownership, sensible measures, and lineage you can trust. For them, Copilot is a force multiplier — it takes a solid foundation and makes it accessible to everyone, and the accessibility is pure gain because the foundation is sound. They spent years, often grudgingly, on data modelling and governance, and 2024 is the year that investment paid off in a form they never anticipated: it turned out to be the thing that made AI safe to switch on.

The organisations having a bad time are the ones who hoped the AI would let them skip that work — who pointed Copilot at an ungoverned sprawl of duplicated, undefined, unowned data and expected magic. They got the confident wrong answers, the "why does Copilot say our revenue is X when finance says Y," the erosion of trust that follows the first few times the magic machine is caught being wrong. And the cruel part is that it's harder to walk back. Once business users have tasted asking-anything, telling them "actually, don't trust it yet, we need to fix the model first" is a much worse conversation than if you'd never switched it on.

Generative AI on your data is a mirror, not a magic wand. It shows the whole organisation, at speed and with total confidence, exactly what your data model actually says — and if you didn't like what the model said quietly, you'll like it far less shouted.

The work 2024 is really forcing

So the real story of Copilot's arrival isn't an AI story at all. It's that the AI has made the old work — data modelling, definitions, governance, quality, ownership — suddenly, visibly, urgently valuable, after years of it being the thing that never got funded because it wasn't exciting. 2024 is, underneath the GenAI headlines, a year of organisations rediscovering that the semantic model is the crown jewel, and that everything trustworthy resolves back to it.

Concretely, the year is pushing serious teams toward a few things:

  • Treating the semantic model as a governed, first-class asset — not a byproduct of building reports, but the deliberate, owned, agreed definition of what the business's numbers mean, precisely because an AI is now going to broadcast those meanings to everyone.
  • Resolving the definition disagreements they'd been living with — because the two-departments-two-meanings-of-"churn" problem, survivable when experts absorbed it, becomes intolerable the moment Copilot confidently reports both and a leader has to pick.
  • Governing what Copilot can see and say — sensitivity, access, and certification, so the AI can't confidently surface something it shouldn't, or quote an uncertified model as gospel.
  • Being honest about readiness before switching it on — the discipline to say "our model isn't good enough to point an AI at yet," which is a much cheaper sentence to say before the demo than after.

None of that is new work. It's the work good data people have advocated for years and rarely got the budget for. What's new is the forcing function: generative AI has made the cost of a bad data model impossible to hide, and in doing so has finally made the case for governance that all the risk-and-compliance arguments never quite could. It turns out the killer app for data governance was AI — because AI is the thing that punishes its absence loudly enough for everyone to hear.

So what do you actually do — before you switch it on

If the meeting between Copilot and your data model is going to happen anyway — and in 2024 it is — the practical question is how to make it a good meeting rather than a humbling one. The honest answer is a sequence, and the order matters more than any single step.

Start by not switching it on everywhere at once. The instinct, under pressure to look modern, is to enable Copilot broadly and fast. Resist it. The blast radius of a confident wrong answer scales with how many people can ask, so the responsible move is to enable it first where your model is strongest — the one or two well-governed, certified areas you already trust — and prove the value there, rather than exposing your weakest definitions to your least technical users on day one.

Then, before widening it, do the unglamorous audit the AI is forcing on you: which of the measures people will actually ask about are well-defined and certified, and which are ambiguous, duplicated, or contested? That audit is the single highest-value thing a data team can do in 2024, because it converts a vague fear ("is our model good enough for AI?") into a concrete, finite work list ("these eight measures are ready; these five need definitions agreed before we let Copilot near them"). And resolving those five is not a technical task — it's the human negotiation of getting the departments who disagree about "active customer" into a room to settle it, precisely because an AI is about to broadcast whichever answer it's given. The tooling to enforce the agreement is trivial; the agreement is the work, as it always was.

The organisations that will look back on 2024 well are the ones that treated Copilot's arrival as the deadline that finally justified this sequencing — enable where you're strong, audit honestly, resolve the definitions, then widen. The ones that will look back on it badly are the ones that switched everything on because it demoed well, and let the AI discover their model's flaws in public, for them.

What the meeting leaves behind

When I look back on 2024 from some future vantage, I don't think I'll remember it as the year the AI got good, though it did. I'll remember it as the year the data model's importance became undeniable — the year the industry's long, boring, underfunded insistence that definitions and governance and quality actually matter got proven right by the flashiest technology of the decade, in the most public way imaginable.

The lasting lesson of Copilot meeting the data model is almost the opposite of what the hype implied. The hype said the AI would make the foundations matter less. The reality is that the AI made the foundations matter more than they ever have, by removing every buffer that used to hide their state. The organisations that will win with AI on their data aren't the ones with the cleverest AI configuration — that part is nearly commodity. They're the ones who did, or are now urgently doing, the unglamorous foundational work the AI made suddenly visible. Copilot met the data model this year. And the data model, it turns out, was the whole story all along — the AI just finally made everyone look at it.