Power BI's Copilot has moved somewhere new and, on the surface, delightful: into the mobile app and into email subscriptions. Now the AI-written "here's what changed and why it matters" summary doesn't wait for you to open a report — it arrives on your phone, or lands in your inbox with the morning subscription. Insight without the dashboard. For the busy executive and the field team who never open Power BI Desktop, it sounds like exactly the delivery mechanism analytics always needed. And it might be. But I want to talk about why this specific move — Copilot summaries, pushed, to people who won't verify them — is the riskiest configuration of AI-in-BI yet, precisely because it's the most convenient.

Push is a different thing from pull

Everything I've written about Copilot this year has, until now, been about pull: someone actively asks a question or requests a measure, gets an AI-generated response, and — crucially — is right there, in the tool, able to look at the underlying report and sanity-check what came back. The human is present at the moment of generation. That presence is what makes it safe-ish.

Mobile and subscription Copilot flips that to push. The summary is generated and delivered, to someone who did not ask a precise question, is not looking at the underlying data, and — this is the key part — receives it in a context that signals "this is a finished, trustworthy conclusion." An executive reading a Copilot narrative in their inbox over coffee is not in verification mode. They're in consumption mode. They're going to read "revenue softened in the northern region, likely driven by the seasonal dip" as a fact about their business, not as an AI's interpretation that might be built on a mislabelled column.

That's the shift. Pull puts the AI's output in front of someone positioned to check it. Push puts it in front of someone positioned to act on it — often the most senior, busiest, least-likely-to-open-the-report people in the organisation. You've optimised the delivery of AI-generated interpretation to exactly the audience least able to catch it when it's wrong.

Why the convenience is the danger

I want to be fair: the convenience is real and the use case is legitimate. Decision-makers genuinely don't have time to open every dashboard, and a good summary delivered to where they already are is a real productivity gain. The problem isn't that this is useless. The problem is that its usefulness and its danger are the same property.

The whole value proposition is "you can trust this enough to not open the report." That's the pitch. And the moment people believe that pitch — the moment the summary becomes a substitute for looking rather than a prompt to look — every weakness of Copilot's interpretation gets a direct, unfiltered line to a decision. The narrative sounds authoritative because good prose sounds authoritative. It arrives through a trusted channel because it's your own BI platform. And it reaches someone who, reasonably, treats "the summary from our analytics system" as a reliable input.

  • The interpretation is Copilot's, not the data's. "Revenue softened, likely driven by the seasonal dip" — the number might be right and the "likely driven by" entirely invented. Causal-sounding narrative is where these summaries are most confident and least reliable.
  • The reader can't see what was left out. A summary is a set of editorial choices about what mattered. Copilot made those choices, and the recipient has no way to know what it decided wasn't worth mentioning — which might be the thing that actually mattered.
  • The channel launders confidence. The same words carry more authority in an executive's inbox than on a screen they know is a live, queryable report. Delivery format changes how much scrutiny people apply, and this format lowers it.
The safest place for an AI summary is next to the data it summarises, in front of someone who might check. This feature does the opposite: it delivers the summary away from the data, to the person least likely to look.

How to ship it responsibly

None of this means don't use it. It means use it with the guardrails the convenience quietly removed:

Reserve pushed Copilot summaries for the well-governed, well-understood models where you'd actually stake your reputation on the underlying numbers — not the messy ones where the interpretation is a coin toss. Keep the path back to the source one tap away, and culturally encourage people to take it for anything they'll act on — the summary as a prompt to look, not a replacement for looking. And be honest in how it's framed: a small "AI-generated summary — open the report to verify" does real work in keeping the reader in the right mental mode. The goal is to keep the convenience while refusing the illusion that convenience equals correctness.

The verdict

Copilot in mobile and subscriptions is a genuinely useful delivery mechanism, and I understand exactly why people will love it — it meets decision-makers where they are and spares them a tool they were never going to open. That's real.

But it's also the point where AI-generated interpretation gets the most direct, least-scrutinised path to the most consequential readers in your organisation, and that deserves respect rather than enthusiasm. The feature's great strength — "you don't need to open the report" — is precisely the belief that makes it dangerous when the report was the only thing that would have caught the error. Use it for the models you trust completely, keep the source one tap away, and never let "insights on the go" quietly become "conclusions without checking." Convenience this good is worth having. It's also worth being a little afraid of.