You Don’t Need Audiophile AI

Microsoft dropped another model announcement this week. Honestly, I’ve stopped trying to keep count of how many of these cross my feed in a given month.

This one is called Microsoft-Decision-1, and it’s a different kind of animal. It isn’t a chatbot or a writing assistant. It’s a small, fast “decision-scoring” model in Microsoft Foundry, built by post-training Qwen3.5-9B, and it’s designed to do one thing very well: look at a fixed set of options and pick one. Routing, classification, prioritization, verification. Microsoft says it runs roughly 35 times faster than GPT-6 Sol, and the pricing is almost comical, at $0.042 per million input tokens with output tokens free (https://commandline.microsoft.com/microsoft-decision-1-model-foundry/).

I read through the announcement twice, and one line buried in the list of use cases stuck with me. I’ll come back to it in a minute.

The Models Really Are Amazing

The pace of what’s being built right now is remarkable. I don’t want anyone reading this as an old guy shaking his fist at the cloud. (I’m saving that for a different post.) Every few weeks there’s a new frontier model, a new reasoning mode, or a new small model that does in a fraction of a second what the big models did slowly and expensively not that long ago. I’m constantly experimenting with new tools, building solutions and testing the waters. Some of these new tools are real breakthroughs, and the people building them deserve the credit.

But there’s one thing I keep coming back to, whether I’m talking with customers or collaborating with peers in the community.

For Most of Us, the Model Barely Matters

If you’re an information worker using AI to draft an email, summarize a meeting, rewrite a paragraph so it sounds a little less grumpy, or rough out a slide deck, which model answers your request barely matters. Any of the mainstream options will do a perfectly good job. The differences between them on that kind of work are real, but they’re small, and noticing them mostly requires a side-by-side comparison and a lot of free time.

For everyday work, picking a model has become an audiophile hobby.

I say that with affection. I have been collecting vinyl records, cassettes, CD’s, and music-related t-shirts, posters, buttons, pins, and more since I was 11 years old, starting in the midst of the New Romantic and New Wave years until today. I also have a component stereo system, speakers, and turntable that cost much more than I’ve even let on to my wife. I love talking about music, but rarely run into anyone who is more than a casual listener these days. When I was a teenager and into my 20’s, the quality of your car stereo and speakers was still important…but these days, people accept whatever the vehicle comes with and never upgrade.

Yes, the quality of the medium matters. There is a vinyl versus CD versus streaming debate. There are people who can genuinely hear the difference, and there’s real engineering behind each format. Vinyl sounds warmer, richer, and best captures the full range of the artist recording, while CD and digital streaming cut off both the lows and the highs. But if you’re playing all three through a pair of cheap speakers, the format isn’t your problem. The speakers are.

Or think about video. You can shoot in 8K, with all the storage and processing headaches that come with it, and then export the final cut at 1080p for people who will watch it on their phones. All that extra resolution gets thrown away on the way out the door.

With everyday AI, the cheap speakers aren’t at the end of the chain. They’re at the beginning. They’re what you feed it.

The Bottleneck Is What You Feed It

When someone tells me Copilot (or ChatGPT, or whatever tool they happen to use) gave them a mediocre result, my first question is never “which model were you using?” It’s “what did you ask it, and what did you give it to work with?”

Most of the time, the problem isn’t the engine. It’s a vague prompt, like “write a status update” without saying who’s reading it, what changed, or what decision you need from them. Or it’s missing context, where the AI is expected to know things that only live in your head. Or it’s messy source material: three versions of the same policy sitting in one folder, a meeting transcript where half the decisions were made in a side chat, a deck pulled together from files nobody has cleaned up in years.

No model fixes that. A better model will just give you a more articulate version of the wrong answer.

I made a similar point in my recent post on building AI solutions in SharePoint (https://buckleyplanet.com/2026/10/ask-copilot-for-direction-build-in-sharepoint-for-results/). The value of an agent depends almost entirely on the content behind it, and the same is true of a single prompt.

The Real Upgrade Is You

If you want better results from AI tomorrow morning, the most effective upgrade available to you isn’t a new model. It’s you.

Write clearer prompts. Say who the audience is, what you want back, and what good looks like. Give it the context it can’t guess, like the goal, the constraints, and the backstory. And spend a little time getting your own house in order, because organized, current, well-labeled content makes every AI tool look smarter.

None of that is glamorous, and it won’t trend on social media. But it moves the needle far more than switching from one model to another.

Like many of you, I often start with the simple, generic prompt…but I don’t expect anything back but simple, generic responses. Ask a basic question, get a basic answer. When I am looking for something specific, or want a detailed, granular output, I am much more thoughtful about my prompt, the content it has access to, and the model with which I am requesting the action.

The Industry Quietly Agrees

Here’s that line from the Decision-1 announcement I mentioned. In the list of suggested use cases, Microsoft includes this:

Model routing: Evaluate an incoming request and select the best model for the task based on quality, cost, and latency requirements.

And in the closing section, Microsoft notes that “it’s increasingly important to choose the right model for the right job.”

Both of those things are true. But notice who they expect to be doing the choosing. Not you. Software.

Platforms are being built so that a fast, cheap model looks at your request and hands it off to whichever model makes the most sense, without you ever opening a dropdown menu. For most people, the best experience is to let the platform auto-select, and that’s exactly where the industry is heading.

So yes, the model matters. Choosing it just isn’t your job anymore.

Where the Model Does Matter

To be fair, there are real exceptions, and I don’t want to shoo them away.

• Reasoning-heavy work is the obvious one. For multi-step analysis, financial modeling, or contract review, the gap between models is less like vinyl versus streaming and more like a camera that can focus versus one that can’t. No amount of good lighting saves an out-of-focus shot.
• Agents taking actions is another. When AI is doing things on your behalf across many steps, small differences in accuracy compound. Microsoft’s own announcement makes a similar point about speed, noting that adding just 100 milliseconds to each of 20 sequential decisions adds two seconds to the workflow. Accuracy works the same way, except a mistake at step three usually costs a lot more than two seconds.
• Cost and speed at scale matter too. For an organization with thousands of seats, a few cents here and a few milliseconds there add up quickly. That’s a real decision, but it’s an IT and finance decision, not something every employee should be weighing before they ask for a meeting recap.
• And then there’s privacy and data boundaries. Where a model is hosted, who can access your data, and what happens to it afterward matter far more than any benchmark score. I dug into some of these questions in my post on what happens to your data inside Copilot Cowork (https://buckleyplanet.com/2026/09/what-actually-happens-to-your-data-inside-copilot-cowork-runtime/).

Notice a pattern? Each of these is mostly a decision for architects, admins, builders, and the people signing the invoices. Which is exactly where it belongs.

Celebrate the Models, Then Get Back to Work

I love that we live in a time when a company can ship a model that makes decisions in a blink of an eye for a fraction of a penny. I’ll keep reading the announcements, and I’ll keep geeking out over the benchmarks with my MVP friends.

But for the vast majority of the work people do in Microsoft 365 every day, let the platforms and your admins worry about which model answers. Spend your energy on the part only you control.

The best AI upgrade you can make this week is to ask a better question.