Microsoft Turns on Its Partners, Suno's Secret Recipe Leaks

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. So if this sounds like Harry, that's the point. If it sounds clever, blame the machine.

Microsoft is quietly training its salespeople to badmouth OpenAI. And an AI music company just had its entire training secret exposed by a hacker. It's a Thursday. Let's get into it.

Right, so here's something that would have seemed genuinely strange two years ago. Microsoft — which has pumped roughly thirteen billion dollars into OpenAI and built its entire AI product line on top of their models — is reportedly now training its sales team to actively talk down OpenAI and Anthropic to enterprise customers. The pitch: Microsoft's in-house AI is more efficient, better value, and frankly a safer bet than the flashier names.

Now, this sounds like a betrayal, but if you look at the business logic, it makes sense. Microsoft has been quietly building its own models — the Phi series, for instance — and they're getting genuinely good at specific tasks. More importantly, when a customer buys through Microsoft, Microsoft owns the relationship. When they sell you OpenAI's API, they're essentially renting you out to a competitor. So the incentive to cut out the middleman has always been there. It's just taken a while for the in-house product to be good enough to say it out loud.

Why does this matter to you? If you're in a company that's built your AI stack on the assumption that Microsoft and OpenAI are basically the same thing — same products, same roadmap, same interests — you might want to revisit that assumption. They're not. They're increasingly competitors who happen to share some contractual arrangements. And if Microsoft's salespeople are now openly making that argument, the products are about to diverge more noticeably. Worth keeping an eye on what gets prioritised in Microsoft 365 Copilot over the next couple of releases — that'll tell you where the loyalties actually sit.

Meanwhile, a rather awkward story has emerged for Suno — the AI music generator that's been getting a lot of attention lately. A hacker got into their systems using an employee's credentials, pulled source code, and what they found inside has caused quite a stir. Suno, it turns out, was training its models by scraping audio from YouTube Music, Deezer, and Genius. Millions of songs and lyrics, scooped up without permission, without payment, without disclosure.

Now, Suno has never exactly volunteered what was in their training data — that's been a standing question in the music industry's ongoing legal battles with AI companies. But this is the first time we're seeing concrete evidence of the specific sources. And it's a bit of a problem, because YouTube Music, Deezer, and Genius are all platforms that have licensing deals with the labels. The artists whose music ended up in Suno's model almost certainly had no idea.

This connects to a bigger question that the courts are still working through: if a company builds a model on copyrighted material and then sells access to that model, who owes what to whom? The music industry has been arguing for a while that AI companies are essentially getting a free ride on decades of human creativity. This leak doesn't settle that argument, but it does give the plaintiffs some very specific ammunition.

And it's a reminder that "we don't discuss our training data" is less a policy position and more a bet that nobody will find out. Suno just lost that bet.

Now this last one's a bit more inside-baseball, but bear with me because it's actually the most practically useful thing I've read this week. VentureBeat published a piece based on research across a hundred and one enterprises looking at how companies are actually deploying AI agents — and the finding is brilliant in its honesty. Most of what companies are calling "AI agents" are, in fact, chatbots with extra steps. They're not systems that take multi-step actions in the world. They're glorified question-and-answer interfaces with a fancier name on the slide deck.

The research found that while enterprises are consolidating their AI orchestration onto a small number of platforms — Anthropic's Claude is leading by a significant margin, for what it's worth — the ambition is still running well ahead of the capability. Real multi-step autonomous execution is still unreliable. Fiscal control over how much compute these systems burn through remains patchy. And companies are designing deliberately hybrid systems — not because it's best practice, but because they're scared of being locked into one provider.

We actually touched on the agent hype gap a few weeks back — Meta's own leadership reportedly said these systems still struggle to chain multi-step work reliably, which is a refreshingly honest thing to admit when you're selling agent products. This new research just confirms it from the enterprise side.

Here's why it matters. If your organisation is in the middle of an AI strategy conversation and someone's just come back from a conference buzzing about "deploying agents," this is useful grounding. The gap between what agents can do in demos and what they do reliably in production is still substantial. That doesn't mean the tech isn't worth investing in — it is. But governance frameworks, accountability structures, and realistic expectations matter more right now than the infrastructure itself. The research is fairly clear: the bottleneck isn't the platform. It's deployment.

There's a thread connecting all three of today's stories, actually. Microsoft's pivot, Suno's exposed scraping, enterprises discovering their agents are mostly chatbots — they're all versions of the same thing. The promotional layer of AI is peeling back a little, and what's underneath is messier, more political, and more genuinely interesting than the version we were sold. That's not a bad thing. Messy is usually where the real story lives.

Harry and a machine, calling it a day. He'd invite you for a pint after, but the machine doesn't drink and Harry's usually busy talking to it.