Can ChatGPT Replace Power BI? Someone's About to Ask You.
Every BI vendor has spent eighteen months rebuilding around conversational AI.
Somewhere between now and your next licence renewal, somebody senior is going to put a question to you in the tone people reserve for routine requests.
Could we replace Power BI with ChatGPT?
Not augment it. Replace it. And the reasoning will have nothing to do with technology. Somebody will have counted the dashboards, then counted the licences underneath them, and noticed that the two numbers have stopped reading like an asset register and started reading like a bill. Most of your estate, they’ll suspect, is floors nobody visits any more — dashboards built to answer a question somebody asked once, in a quarter nobody remembers, renewing annually ever since.
I know the tone because I’ve heard it. A few weeks ago an enterprise put exactly that question to me, and what I found on the way to an answer is worth having before your version of the meeting.
There are two respectable reactions, and you can probably hear both in your own building already.
Your BI leader’s reaction: absolutely not. A dashboard is a governed object. A certified report gives the same answer at nine o’clock that it gave at eight, carries its lineage with it, and stands up in an audit. And we’ve seen the chat-with-your-data film before — twice. Tableau retired Ask Data, its first natural-language interface, in 2024 for lack of adoption. Microsoft switches off Power BI’s original Q&A feature this December for the same reason. Natural-language BI has been the industry’s most reliably broken promise for a decade.
Your finance director’s reaction: have you seen the bill? Somewhere in the estate there’s a dashboard for every question anyone ever asked twice — each one built by someone, each one a seat, each seat a renewal. Meanwhile the conversational front ends grew up. Claude’s Artifacts became persistent, live dashboards in April — ask, and a working chart exists, connected to the data, refreshed when you reopen it. ChatGPT Enterprise connects straight to the warehouse. The thing the licences were paying for — a chart on demand — is now a sentence.
Both of them are right. That’s usually the sign the frame is wrong.
The vendors have already voted. In roughly eighteen months, every major BI vendor rebuilt its flagship around conversation. Amazon folded QuickSight into an agentic workspace last October — a first-tier cloud vendor’s BI product, now one tab in a chat suite. Microsoft retired the Power BI Premium SKUs and now meters every Copilot interaction against Fabric capacity: the seat is no longer the unit Microsoft grows on; the conversation is. Salesforce launched Tableau Next as “the world’s first agentic analytics platform”. ThoughtSpot declared its platform “100% agentic” — this from the company whose blog is literally titled “The day the dashboard died”. Databricks shipped a tool in June that imports Power BI and Tableau workbooks and rebuilds them on its own stack — removal vans, parked outside the estate.
And then the quiet one: in June, Microsoft released open-source skills that let Claude — another company’s assistant — query Power BI’s semantic models directly. Nearly every major BI product now ships an MCP server so that external chat clients can sit in front of it. When the incumbents start building the plumbing for their own disintermediation, the argument about where the front end is going is over.
So the replacement should be visible in the numbers by now. It isn’t. I went looking for a single named enterprise that has cut Power BI or Tableau seats in favour of ChatGPT or Claude, and there isn’t one in the public record — the only quantified claims are anonymous, and they live in vendor marketing. Power BI is still growing; Microsoft’s Fabric platform passed a $2 billion annual run rate in January. Dresner’s 2025 study found 90% of organisations holding or raising their BI budgets.
There’s exactly one tenant giving notice at the revenue line. Salesforce’s CFO has cited “increased softness in Tableau bookings and renewals” on earnings calls in December and May, while the company’s agent platform crossed $1 billion in annual recurring revenue. Tableau’s standalone chief executive role was eliminated in February. Qlik’s chief executive resigned in April. The dashboard franchises aren’t shrinking yet — but the people who ran them have started leaving the building.
The gap between the vendors’ conviction and the empty removal vans is explained by a set of numbers that never makes the sales deck. Anthropic — a company with every incentive to claim a chat window can do this job — published its own internal figures in June: Claude now handles about 95% of the company’s internal analytics questions, at roughly 95% accuracy. The same system’s accuracy before they built a governed semantic layer underneath it: 21%. AtScale had already measured the same cliff in 2024 — language models answered around 20% of business questions correctly against raw schemas, and 92.5% through a semantic layer. Gartner now predicts that 60% of agentic-analytics projects that wire a model to the data over MCP alone will fail by 2028, for want of exactly that layer. And determinism is a separate problem with the same shape: Microsoft’s own documentation says plainly that Copilot’s answers are nondeterministic and can be wrong. A certified report can’t carry that caveat.
There’s even a name now for what happens to teams that swing all the way to conversation: conversational debt — the insight that answered a question once, buried in a thread nobody can find. The chat-first teams that have written it up hit reproducibility problems within months, and the fix they reach for is a standing artefact. The output of a good conversation with your data, it turns out, is a dashboard. Just not a licensed one.
So the thing being replaced isn’t the dashboard. It’s the request queue — the long tail of screens built once per question and renewed forever, which is where the licences pile up. And the thing being built isn’t a chat rollout. It’s a semantic layer with an identity rail underneath it. The model orchestrates. The semantic layer computes. Identity decides what it’s allowed to compute with.
Dashboards won’t die of bad design. They’ll die at renewal — one unvisited floor at a time.
Here’s what that means for the estate with your name on it.
The deadline in this story isn’t a technology date; it’s your renewal date. Between now and then, the question will arrive whether you’ve prepared or not — and the difference between a panicked procurement conversation and a controlled one is which of you has the usage numbers first. The certified core of your estate — the numbers a board depends on, the reports a regulator can ask about, the screens somebody would miss under oath — is load-bearing, and it stays, because a certified number must be the same number twice. Everything else is the pile the question is really about.
Your security model moves too, and this part lands on your desk whichever way the front-end argument goes. When the chart is generated fresh for every question, access can’t live in which dashboard somebody was granted any more. It has to live in who’s asking. That’s an identity decision, not a BI feature — and if nobody in your building owns it yet, the replacement conversation will surface that gap at the worst possible moment.
And if your organisation swings all the way to conversation, you already know how that ends, because the teams ahead of you have written it up: conversational debt, reproducibility complaints within months, and a quiet rebuild of standing artefacts. Plan for that ending now and it’s cheap. Discover it later and you’ll have paid twice — once for the licences you cancelled, once for rebuilding what they did.
So, before the question arrives: pull ninety days of usage and split your estate into the two piles — the core somebody would miss under oath, and everything else. The second pile is your negotiation, whoever raises it first. Then put one question to anyone — vendor or colleague — who proposes conversation as the front end: when the answer comes back, who computed the number, the model or a governed semantic layer the business has signed off? If the answer is the model, you’re not looking at a replacement for your estate. You’re looking at a demo.
If this estate decision is heading for your desk: two minutes — find out what’s actually standing in its way. Run As You is free at the end.
The company that asked me had, I think, named the wrong thing and found the right one. They asked about a front end. What they’d actually found was a bill that had outlived the questions it was paying for. Yours will be the same decision when it comes — an estate decision, not a software decision. And estates don’t fall. They just stop being renewed — floor by floor, at dates already printed in your contracts. Choose the floors before the question chooses them for you.
Until next time,
Chris
P.S. — The whole resolution above hangs on one question: when the chart is built fresh for every question, who is the AI acting as? Run As You is my free book on exactly that — the badge, not the master key, and the pattern that lets a conversational front end touch governed data and survive the security review. An evening’s read: get it free.