Good morning, C. AI is built into almost every app we open, but that doesn't automatically make it useful.

Microsoft's answer is to bring it all into one place with Copilot. Built around three new capabilities — Home, Code, and Autopilot — it's being pitched as an “OS for work” where you can chat, build apps, and work alongside agents.

In partnership with Microsoft, we sat down with Jacob Andreou, the company's EVP of Copilot, for an exclusive Q&A on what the new Copilot is trying to fix, where agents still need a human in the loop, and the lessons that changed how he thinks about AI products.

In today’s AI rundown:

LATEST DEVELOPMENTS

LESSONS LEARNED

🥇 What Microsoft had to relearn with Copilot

The Rundown: Microsoft removed Copilot from several Windows apps after finding they brought traffic but little value. Andreou says he hit the same wall as a user of the company's own consumer agent, and expects this fall's personal agents to hit it too.

Cheung: Microsoft pulled back on Copilot in some of its Windows apps. What did that teach you about where AI doesn't belong?

Andreou: We had to relearn that putting more people into the top of the funnel doesn't equal delivering more value to them. We picked a bunch of entry points that looked high traffic but were relatively low value. There were even a couple places where you could be using your Windows machine and then suddenly, Copilot. That doesn't help anyone.

Andreou added: I really, really instead believe in products earning their right to exist. Usage is fleeting unless you're delivering real value. We have less entry points that exist today, but have more usage per user and more people using it.

Cheung: What's something you've changed your mind about with AI agents, and what changed it?

Andreou: At the beginning of this year, we released Copilot Tasks in the consumer product. It increased my consumption a lot. I was ordering more food delivery, and it would call Ubers for me when my meetings ended. That felt magical for a few weeks, and then at some point it was like eating too much ice cream.

Andreou added: For the harder stuff in my life, like financial planning or planning a vacation around work commitments, it was kind of an inch deep. That was a big learning for me: something that was cool in the early days needs more meat to be something I feel amazing about using.

The consumer side of this right now is very hectic. You now have Muse, Instinct, and all these other folks. I think those products will learn similar lessons and evolve accordingly over time.

Why it matters: As companies rush to add AI to every product, Microsoft's experience points the other way. Cutting entry points raised usage per user, showing what keeps people coming back isn't convenience, but how much the AI can actually handle beneath the initial surface gains.

MODEL CHOICE

🎯 Does the model still matter?

The Rundown: Andreou says no single lab's models win at everything anymore, and that expecting workers to track which model suits which job has become a cost of its own. Copilot's answer is an Auto mode that picks for them while still offering the option to choose.

Cheung: Does the model still matter anymore, or is it all about the product on top?

Andreou: The model definitely matters. But having to behave as if the model really matters is a large tax to put on professionals and companies.

Andreou added: For the first half of this year, you could take the full line-up from a single provider and get the best experience across everything. That's changed. What became clear is that certain models became really good at certain tasks.

OpenAI's new class of models, starting with 5.6, were really smart, so in places like Excel they were really good. But Anthropic's models still had more of a native understanding of the visual consequences of the code they produce, so for drafting a PowerPoint, they still made sense.

It's become so nuanced that the majority of people are just drawn to solutions that lead to great output. With Copilot, we still want to let you pick a specific model, but we want Auto to give you the best output, whatever you're doing.

Why it matters: Performance now depends on the type of output rather than which model is currently on top. That's Andreou's case for a multi-model approach that routes each task to the model that fits best, while still letting people choose a specific model if they want.

PRODUCT FORM FACTORS

 🤖 Smarter models aren't enough

The Rundown: Andreou says AI is in an “overhang” moment, where the newest models lack the right products built around them to be useful to most people, and if this gap persists, AI’s value will centralize to a few companies.

Cheung: What problem were you trying to solve with the new Copilot that the old approach couldn't?

Andreou: Each generation of intelligence needs a product form factor for you to actually get value out of it. GPT-3 was generally under-appreciated until chat showed up and you could actually talk to the model for the first time. Opus 4.5 came alive via Claude Code and Cowork.

Today, we have Astra from OpenAI and Fable from Anthropic, and it feels like we're missing the product form factor to get the magic out of these models.

Andreou added: If we don't build products that translate raw intelligence into real value for real people, intelligence will centralize in the hands of a couple of companies, and value will centralize along those same lines. That’s what we’re trying to solve with Copilot.

Why it matters: Microsoft doesn't make either model Andreou names, so this is also a description of its own position. Its vision is that the product around a model will matter more than the model, and that the app you work in will shape your results more than the lab behind it.

AGENT AUTONOMY

🚪 How much freedom should an agent have?

The Rundown: The more freedom an AI agent has, the more it can get done, and the more things can go wrong. Andreou's rule of thumb: sort every action into one-way doors (irreversible) and two-way doors (reversible).

Cheung: Autonomy makes agents useful, and also concerns IT teams. How should companies decide how much freedom to give an agent?

Andreou: The tricky part about this is the same things that make those kinds of products amazing, their flexibility and their autonomy, are the same two things that make them scary to put to work in an enterprise.

My first experience with OpenClaw was setting it up inside my home network and it helped me fix a bunch of things on my home network, but it didn't have access to any of those things. It found a way in, which was a bit terrifying. Obviously, in an enterprise context, that's not acceptable.

Andreou added: If an agent is contacting another person, deleting a file, or making an irreversible decision, a human needs to be in the loop every time. For reversible work, it can operate independently inside clear boundaries and ask again if the task or risk changes.

With Autopilot, we pre-compute what the browser session should look like when the task first starts, and that becomes the boundary. If it tries to solve the problem differently, it has to ask for permission again. Agents are persistent. If they get stuck, they'll try something else, and that's where that second ask comes in handy.

Why it matters: Most talk about agent risk centers on what an agent can access. Andreou's home-network story is about what an agent does when its first plan fails. If you're testing agents at work, list the actions that can't be undone before you start, and ask how the tool behaves when it gets stuck.

AI & WORK

📬 What AI-native work looks like

The Rundown: Andreou handles his inbox through Autopilot, and argues a great AI-native employee doesn’t stop at making tasks a little faster — they reimagine their whole workflow.

Cheung: Based on what you've seen internally, how is Autopilot actually being used?

Andreou: The biggest surprise was that our top adopter was the sales team. They’re using it for everything from ongoing pipeline management to prepping for client calls to making sure they follow up at the right times.

Personally, I use it as my primary interface to email. I don't really open my Outlook inbox. When I wake up, it has a triage brief ready. When I want to dive into a thread, I just talk to Autopilot and it catches me up while knowing when to escalate. If it's an email from Satya that I haven't responded to, it'll escalate it to reach me.

Cheung: What does a great 'AI-native' employee do that others don't?

Andreou: They don't just use AI to make tasks a little faster. They rethink the workflow: what should I do, what should I delegate, and where does human judgment matter most.

Andreou added: One of our user-research leaders, who isn't a software engineer, used Code to build a shared app that continuously pulls together competitive comparisons and product research from across the organization. What started as an individual workflow is now something the broader team relies on every week.

Why it matters: Being AI-native still comes down to human judgment and decision-making. The research leader at Microsoft didn't need to know how to code, but did need the expertise to decide what was worth building.

GOING DEEPER

✈️ Explore the new Copilot

Discover how Copilot's new capabilities help you build, customize, and scale AI at work.

See new Copilot capabilities in action.

See you soon,

Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown