Anthropic’s own experiment suggests the answer is no, and shows what to do with them instead.
During the last two years I’ve been asked the same two questions by almost every CEO and transformation leader I meet. Can AI simply read all our data and answer our questions? And if it can, do we still need data analysts? People ask them as if they were separate topics, but they are the same question, and I have been giving the same answer every time: AI is not limited by its intelligence nearly as much as it is limited by the quality and meaning of the data we give it.
Until recently, that answer was based on experience. Today it is also backed by one of the world’s leading AI companies.
Start with the number
Anthropic recently published something remarkable. Inside their own company, the maker of Claude tested exactly the problem I’ve been seeing in organizations across Europe: they let their AI read their own data without a governed foundation underneath it, and the answers it produced were correct only 21% of the time, which means it was wrong roughly four times out of five. Once they built the foundation, meaning one agreed definition for every important number, accuracy climbed above 95% and reached around 99% in some areas.
The AI did not change and the data did not change either, but the only thing that changed was the foundation beneath them.
The right way to read this is as a business fact rather than a technical one. If you point AI at data that has no single agreed meaning, it will not pause to tell you it is unsure, it will simply hand you a confident and wrong number faster than any human could. An unsure person hesitates and double-checks before speaking, while an unsure machine answers anyway, and it answers with complete conviction.
That alone settles the first question. No, you should not simply point AI at your source data, and the reason is not that the AI is weak, the reason is that your data does not yet carry a single meaning for the AI to find. And that is exactly where the answer to the second question begins.
The reality check for most companies
Here is where the picture changes for most companies, because the setup Anthropic used is one that many cannot legally copy. Anthropic connected its AI directly to its own data, whereas for most companies data privacy usually keeps that data out of the cloud, and without nearby data centres capable of running the strongest models in-house, the choice is between technical workarounds that reach the cloud safely and smaller models that will not catch up with the frontier for a long time yet. That constraint sharpens the lesson from Anthropic, because if a company that builds the best AI in the world still needed a governed foundation to climb from twenty-one percent to ninety-five, then a company starting with weaker tools cannot skip that foundation and quietly hope the model makes up the difference. You cannot out-model Anthropic, but you can out-govern almost anyone. The foundation is the one lever that sits entirely in your own hands, because a weaker model working on well-governed data will beat the strongest model in the world working on a mess.
In my work with these companies, this is exactly where the word governance gets misunderstood, so it is worth saying plainly what it is and what it is not. Data governance is not the thing the regulator asks of you or a spreadsheet in which someone once wrote down a list of definitions that nobody has opened since. Real governance is a culture and a structure held together at the same time, which means a shared and enforced agreement across the whole company about what each important number means, about who owns that meaning when someone disputes it, and about how any figure on any screen can be traced back to the source it came from. It is the slow and unglamorous work of deciding that “active customer” will carry one definition rather than three, and then building the discipline to keep it that way as people leave, systems change, and new reports get built on top of the old ones.
This is exactly where many emerging markets are exposed, and the numbers make the gap difficult to argue with. In the OECD’s 2026 survey of more than eight hundred firms in emerging European markets, roughly half already report using AI, yet almost all of that use is generative tools of the ChatGPT kind, while the deeper foundations that would make AI trustworthy, cloud platforms and integrated business systems, reach only about one company in ten, and proper data analytics remains something that only a small minority of firms do at all. Those numbers do not surprise me, because they describe what I walk into almost every week: an enthusiastic owner, a handful of ChatGPT licences already in use, and underneath all of it the same unreconciled definitions of the same numbers that nobody has had time to fix. The survey confirms that this is the default state in many markets, enthusiasm resting on a thin foundation.
So the foundation is the work itself, and it is work that only people can do. You cannot quietly hand it to the machine as an optional first step to deal with later. Which brings us to the question every owner asks the moment I say the foundation has to come first. If I build all of this, if I finally give my data one governed meaning that a person or a machine can rely on, does that mean I no longer need the analysts?
The developer blueprint: what coding teaches us about analytics
Before you decide the fate of your analysts, it is worth looking at what just happened to software developers, because their industry spent the last half a year having this exact debate in public, and I followed it closely since every argument in it maps directly onto analytics.
Two camps formed almost immediately. The first argued that AI was quietly taking the jobs, and a Stanford study handed them their evidence: employment for developers aged 22 to 25 has fallen by almost 20% since late 2022, while their older colleagues in the same companies kept growing, so the youngest people doing the most routine work were precisely the ones who disappeared. The second camp, with Oxford Economics as its strongest voice, answered that AI was named in less than 5% of American layoffs and that the drivers were a weak economy and too many graduates chasing too few jobs, which makes AI a convenient story for investors rather than the cause. From what I see in the market every day, both are right about their own half. The routine entry-level work is the part that evaporates first, exactly as Stanford measured, and at the same time companies do use AI as a respectable cover for cuts they would have made anyway, exactly as Oxford suspected. The two studies describe the same shift from two different windows.
The lesson both camps eventually agreed on is the one that matters for your data team: AI absorbs the routine work, judgment becomes the scarce skill, and hollowing out your own team is the danger hiding behind the productivity gains. We have also seen where that road leads, because in 2008 companies froze entry-level hiring to save a visible line on the salary budget, and a few years later they could not find experienced people anywhere, for the simple reason that they had stopped growing them. That cost never appeared on any budget, which is what makes it the most expensive kind of cost, the one that never sends an invoice. The trigger was a recession then and it is AI now, but the mistake is identical: trading away the next decade to look efficient this quarter.
Notice who is playing it differently. In February 2026, while much of the industry was still cutting, IBM announced that it would triple its entry-level hiring, with its HR chief explaining plainly that AI can now do most of what the old junior job consisted of, so IBM redesigned the job instead of deleting it and moved its juniors toward customers and problem-solving. That is a calculated position taken on the next decade of experienced talent. And one honest note belongs here, because some of the developer panic was exaggerated, which is exactly why you should copy neither the hype nor the layoffs into your own business. Learn the pattern and leave the panic.
So, what about your analysts
Over the past decade I’ve asked executives the same simple question in banks, insurers and retailers, how many active customers do we have, and I’ve almost never received a single answer. Risk has one number, retention has another, growth has a third, and each of them is correct inside its own report, because each department uses a slightly different meaning of the word active. Nobody is wrong, and nobody can agree.
Now imagine putting AI on top of that. It will not stop to ask which meaning you had in mind, it will simply choose one and hand you a number with full confidence. I have watched that exact scene play out, a confident and wrong answer delivered to a room full of senior people, and everyone in the room nodded.
Every one of those companies sits in one of two situations. Either you do not yet have proper BI, which means you cannot reliably see your own numbers at all, or you have the reports but no agreed meaning behind them, so the same question produces three answers in three meetings. And above both situations sits an owner who loves AI and wants to use it now.
Here is the sentence that resolves both of the questions we started with. AI lacks context, and governance is context, which means the analyst is not the person the AI replaces, the analyst is the person who supplies the one thing the AI does not have.
The job itself does change, and it would be dishonest to pretend otherwise. The routine report-pulling and the standard weekly numbers do move to the machine, the same entry-level work that vanished for young developers, and pretending that nothing moves is how the “just describe it and ship it” crowd ended up with software that broke in production. But the harder work does not disappear anywhere. Deciding which definition is the right one, checking the number before it reaches the boardroom, connecting the data to a decision, all of that moves up, and it moves up to your analyst.
And this is bigger than your analysts. For a mid-market company in Europe the frontier models are walled off twice over: the cloud and data privacy keep your data out of their reach, and the price, heavily subsidized today, is unlikely to stay this cheap once the AI companies need to earn money rather than buy market share. That double wall is quietly reshaping who can win with data at all, and it may even open a door for companies that spent years a step behind, but that is a separate story and we will address it separately.
What to do instead of firing them
There are two moves, and IBM has already demonstrated both of them.
The first is to train them more. While others cut to save a visible salary line, you deepen the one capability that the cloud constraint says you must win on, and you take the time the machine frees up and point it at meaning, at context, and at the harder business questions that nobody had time to ask before.
The second is to embed them into your business processes, because they are not a separate team anymore, in the same way that AI is not IT anymore. AI stopped being a project that lives inside the IT department and moved into every process in the company, and your analysts need to make the same move, out of the reporting corner and next to the decision, as the people who own what the numbers mean.
So, will AI let you fire your data analysts?
No, it will not. What it fires is the part of their job that was never the point, and what it promotes is the part that always was.
