Power BI is Marmite. ChatGPT won't help.

I was still miffed at his utter smugness when a CX Director we work with (and admire greatly!) confirmed something we've long suspected: "Most people in our business aren't comfortable with Power BI. To them it's heavy data wrangling—it's Marmite."

Half love it. Half—the half she needs to win over, the operations leads and service managers whose hands are on the actual levers—would rather prise off their toenails with a rusty spoon than open it.

As you may know, reader, the Marmite of BI is absolutely one of our favourite things. Proper analysts, the ones who can pivot, filter, segment, and cross-tabulate their way to a real answer in forty minutes flat, they deserve a platform that lets them play. The trouble is there aren't that many of them. Sometimes just one or two a team, and they're on holiday.

The ninety-five percent who are left—the people who actually need to act on customer feedback—don't want to get good at Power BI. They want answers. That gap is what this one is about.

The 25% problem

If you've watched a shiny new insight programme stall after six months, you're having the statistically normal experience.

BARC and the Eckerson Group have been tracking BI adoption for around two decades. Wayne Eckerson himself contends: "The average adoption rate of tools designed to help business users query, visualise, and analyse data and share insights has been stuck around 20% for many years."

Closer analysis suggests the actual number of people who use the BI tools their organisation has bought varies between 20 and 29 percent of a typical workforce. And the "many years" for which it's been stuck is seven. Seven Years! 0.4375 Eddies.

Voice of Customer is, if anything, worse. Forrester's 2025 State of Feedback Management survey found that 79% of businesses have no process to act on insights at all.

Only 27% of CX teams can effectively communicate insights in a timely way. Qualtrics XM Institute's own figures suggest that 71% of CX practitioners are sitting in the two lowest stages of maturity—and only 2% have reached "Embed".

This is not a "buy better software" problem. Organisations have bought the software. The software is not being used. The second user—the one who decides whether this investment lives or dies—turns off at the third click.

The obvious wrong answer

When a CX leader tells their vendor that the tool is too complicated for the wider team, the industry's reflex answer in 2026 is one word. Chat. Add an AI layer. Let people ask questions in plain English. Problem solved.

Back up, miladdo! Problem very much not solved.

The first problem is reliability. Peer-reviewed research in Frontiers in AI last year found that identical prompts sent to GPT-4o produced sentiment scores that swung by 0.3 to 0.6 across 100 runs.

Even at temperature zero—the setting that's meant to eliminate randomness—29% of outputs still varied. The paper has a name for this: the Model Variability Problem. You ask the same question twice. You get two different answers.

The second is what Microsoft themselves publish about their own Copilot. Deep in the Power BI Copilot documentation, there's a line that we think every CX leader should tattoo onto their monitor: "If you fail to prepare your data, Copilot produces mainly low-quality and inaccurate outputs that might be incorrect or even misleading."

That's Microsoft warning you about Microsoft's own product. They're right to. Pointing an LLM at a pile of raw customer feedback and asking it to tell you what's in there is one of the most confidently wrong things a piece of software can do.

The thing the data world figured out a year ago

While the VoC industry has been busy bolting chat interfaces onto raw verbatim and hoping for the best, the data-engineering world was having a completely different conversation, and they quietly settled it.

The conversation was: does AI need structure? And the answer the data people have converged on—dbt Labs, Cube, Hex, Benn Stancil, Microsoft's own Fabric team—is a unanimous yes. You build a semantic layer. A governed, pre-defined set of metrics, dimensions, and definitions. The AI writes queries against that. Not against the raw data.

Why? Not because it gets fewer answers wrong. The accuracy numbers are fine either way. It's because of what happens when it's wrong. Here's dbt's phrasing, which frankly is the sentence this whole article is organised around:

"With text-to-SQL, failure looks like a plausible but incorrect answer. With the Semantic Layer, failure looks like an error message."

Read that twice. It's a hill we'll bloody die on!

An AI working against raw data will cheerfully invent a number that's fluent, confident, defensible-looking and wrong. You have no way of knowing unless you already know the answer—in which case why did you ask?

An AI working against a semantic layer either gives you the right answer or tells you it can't. The failure is legible. You can see it. Which means you can fix it, or at least not ship it.

For a board report, a regulator conversation, or a C-level deck, "plausible but incorrect" is the one kind of failure you cannot survive.

The VoC equivalent

We think—and yes, we're monumentally biased—that the VoC industry hasn't yet had the BI world's conversation.

Most of the vendor landscape in 2026 is selling the chat layer as the headline feature. Ask Athena, XM Assist, Copilot, Answers. The chat is the hero. The structure underneath is an afterthought. Sometimes it isn't there at all—the LLM is reading raw verbatim and summarising.

One of our customers at a large housing association told us recently that "If I suddenly dropped 6.9 [on the platform] when the survey said 7.1, there're gonna be questions and guns coming towards me." That's not a theoretical risk. That's an analyst whose job depends on numbers reconciling. A chat interface that invents a different answer each time you ask is worse than no interface at all.

We've been kicking the tyres on this ourselves, in both directions.

Our Product Manager Izzie has been stress-testing Power BI Copilot (note: different from Microsoft Copilot! Confusing, I know) on our own BI semantic model for the past few weeks.

The good bits are impressive. Copilot autonomously explores the data structure, handles custom DAX, spins up new visualisations on the fly, and—crucially—keeps an audit log you can interrogate. It's significantly less black-box than you'd expect.

The caveats are real, though. It can only answer one question at a time. Its forecasting logic is still partially opaque. And if you ask the same vague question twice, you can get two subtly different justifications—all plausible, probably all valid, but dangerous if different people ask the same question in different ways.

I find myself returning to a line Iz read on Reddit: "Copilot doesn't fix weak BI foundations. It amplifies whatever you've built."

That, we think, is the whole game.

Head of Growth Steph has been running the other side of the experiment—Claude Code, in a secure European environment, working not against raw verbatim or the BI semantic model, but against the structured, themed data that's already come out of our platform.

With a specific methodology wrapped in a repeatable Claude skill, she's getting standardised reports that faithfully represent the data. A chat interface on top of a trustworthy structure. The output is genuinely useful. It's also reviewable, defensible, and—the bit that matters most—reproducible.

We're not yet ready to bring this to customers. There's calibration to do on both tiers, and a pile of questions we want to answer to our own satisfaction before anyone sees it. But the direction is clear—a two-tier architecture:

1. A governed taxonomy underneath.
2. Analysts who can play in Power BI on top of it.

Chat for the ninety-five percent who just want the answer—provided somebody they trust is looking at the data underneath.

That last phrase is doing a lot of work. The whole premise of the two-tier is that trust has been earned at the layer the chat is sitting on. The themes are defined. The classification is reviewable. The accuracy is measured. The governance is owned.

The structured taxonomy is the semantic layer of VoC—the thing that lets AI give you the same answer twice. As LLMs fold into every tool we touch, more people will just want the answer, not the analyst's working.

Which brings us back to Ian

Anyone who's been reading CX Corner for a while will remember Ian from Issue 49, who famously interrupted a demo to ask "when you're leaving it up to an AI model to make the call, who is liable?"

That question has not gone away. If anything, it has intensified. Over the last month we've heard versions of it on call after call. "Who trained this?" "What were the parameters?" "What were the prompts?" These are not people trying to catch us out. They're people trying to do their jobs without getting fired.

Two-tier is the architectural answer to Ian's question. When the AI is writing queries against a taxonomy you defined, reviewed, and signed off, liability is straightforward. You own the model of your business. The AI is a faster way to read it.

The EU AI Act calls for traceability and explainability; ISO 42001 for auditable outputs. A governed taxonomy gives you both. A chat interface sitting on raw verbatim gives you neither.

So, what does this mean for Wordnerds?

Wordnerds is the voice-of-customer semantic layer. Not a dashboard. Not a chat interface. Those are outputs. The product is the governed, organisation-specific, defensible layer of meaning underneath—the thing that makes every query repeatable and every answer defensible.

The analysts on our customers' teams love the platform's outputs. They can build on it—pull it into Power BI, layer their own metadata, cross-reference with operational data, explore.

Izzie's work is helping us make sure that experience gets sharper with Copilot sitting on top.

The other ninety-five percent of their organisation—the ones who just want to know what customers are saying about repairs this month—increasingly won't touch a dashboard. They'll ask. And they'll get a trustworthy answer because someone they trust—with a methodology we've tested ourselves—has done the work at the layer underneath.

Steph's work is where that part's being proven.

We would say that, wouldn't we. But—hear us out—it's also what the data engineering world has already concluded is the right architecture for every other category of analytics. VoC is where BI was twelve months ago. We'd like to spare you the intervening detour.

Until next time, keep learning!

Pete

---

P.S. At this point you will probably benefit from some context about my relationship with my son. Things you need to know:

1. He's a smartarse (Exhibit A: his Marmite/take it or leave it comment)
2. A couple of years ago he thought it would be hilarious to replace our names in his phone with stupid alternatives. I became "Fatty", Mum got "Spawn Point" (it was during his Minecraft obsession) and poor Amelie—his sister—is "Spare Organs"
3. As retribution he went into my phone as "Least Favourite Child", something that often gets me disapproving looks when my phone goes off in public. I basically can't win.

P.P.P.S. I'm so sorry we sent this out with an old title. In our defence Steph spent ages correcting the minutiae of all of the stupid little things I did wrong and none of us thought to check the MASSIVE text at the top. We are idiots and we are contrite.