Cloudwell Conversations: Sébastien Levert on Copilot, Cowork and Scout and Why the Meter Changes the Math

There’s a Copilot, a Cowork, a custom agent, an autopilot. What’s the simplest way to think about which is which?

Sébastien Levert (SL): Think about it as a spectrum of control. How much do you want to configure yourself, and how much of the platform do you want to plug into?

Declarative agents sit at the easy end. They’re configuration. You describe what you want the agent to do, in plain language. It’s not an IT project, there’s no deep integration work, and you won’t get to one hundred percent of some very specific vision. But it’s the fastest way to put AI in a person’s hands and let them shape their own way of working. That’s where most people should start, because you feel what AI can do for you almost for free.

Move along the spectrum and you get custom-engine agents. Now you want full control. Maybe you’re in an industry that needs a very specific or specially trained model, and you want the thing to map to your business process one to one. You can get all of that. The trade-off is integration. A custom-engine agent runs as part of Copilot but outside its context, so you don’t get the same richness of data flowing through it.

Then there are the autopilots, like Scout. These have their own schedule. They work alongside you, keep you in the loop, and message you when they hit something they need you for.

So: start declarative. If you need more control, build a custom-engine agent. And if you want a buddy that sits beside you and gets on with things, that’s an autopilot.

When do you actually move someone from Copilot to an autopilot like Scout?

SL: When the work is repetitive enough to teach, and you trust it enough to hand over.

The mental model I use is onboarding. When someone joins your team, you spend the first stretch teaching them. ‘Here’s how we do this. Here’s what to check every day, every week, every month.’ As a human, you scribble it in a notepad and build up the knowledge of how to work with you. You’re writing your own guardrails.

An autopilot is the same, except it learns far faster. You give it the doc, and it’s done. So the question becomes: which of your tasks look like that? The status report. The checking of every work item assigned to you to figure out what still matters. That’s the busy work, and it’s exactly what something like Scout will take off your plate and keep you posted on.

That’s the real shift. The busy work goes to the agent that’s waiting for a command. Your attention goes to the high-value job. And the test for whether to invest is simple. Doing more is not better. Doing the right thing is.

IT leaders are nervous about handing autonomy to something that doesn’t get tired. How do you keep an autopilot governed?

SL: On two levels. The hard floor, and the soft controls you set yourself.

The hard floor is in the platform. Scout is built for work, and it assumes the person using it might not be a security expert. It runs in execution containers, sandboxes on the machine. So even if the underlying agent decides the smart move is to delete every file on your computer, Windows stops it. That’s protection you get within Microsoft 365 without having to engineer it.

The soft controls are yours. Scout only acts on the signals you give it. It will not go and read your email unless you tell it to. You say: every fifteen minutes, pull my latest messages and draft replies. If you don’t ask, it doesn’t go off and decide for itself. It’s unleashed only in the directions you point it.

Past that, treat it like a relationship. You need to become the gardener of that experience. I keep a thirty-minute Scout meeting in my calendar every week. I go over what it created, I clean up, and I add memory. “By the way, when this happens, don’t do that.” It’s rough at the start and gets tighter the more you feed it. The more you do with it, the better it gets and the more you learn how to work with it.

Where do you draw the line on what it’s allowed to do on its own?

SL: By risk, and you stay elastic about it.

The risk of each action is different, so the guardrail should be different. A marketing autopilot sending cold outreach? Go for it, and loop me in when someone replies.

Anything that touches my relationships, my role, my reputation? I haven’t unleashed my Scout on that. It does the heavy lifting, helping with the thinking, the drafting, the cleaning, but a human stays in the loop, even if it’s only for a thumbs up.

Think about how you’d treat a new hire. At first you want a daily check-in on everything. Three years in, if they still came to you for every decision, that would get annoying for both of you. You’d want them to use their judgment. We’re early in that journey with agents, and I expect the comfort level to shift. For now, any decision with real impact stays with the person.

There’s a second thing leaders underestimate, and it’s about security, not autonomy. Inside the Microsoft boundary, your documents are already protected: data loss prevention, Purview, auditing, logs. Staying inside the platform is where a lot of the value is, because the safety nets are already there. But agents expose what I’d call security by obscurity. The document buried so deep in SharePoint that nobody would ever find it? An agent finds it, instantly. So, before you switch any of this on, your access strategy has to be real, not assumed.

Now the part keeping finance awake. How do you justify the cost of the agents that bill on usage?

SL: Start by accepting that there is a cost, and that’s normal. AI productivity costs money, the same way hiring an intern or a new employee does. We’ve been working a long time on flat licenses, and that’s shifting. GitHub Copilot already moved from a flat experience to a credit experience, and that’s a very different way of pricing. So, the discipline is to know what each task is worth before you run it.

And understand how the bill is built, because there are two layers. First there’s the seat. You want the full kit: Word, Excel, PowerPoint, and Work IQ connecting your data, so the everyday Copilot work is a flat, predictable cost per person.

The agent work sits on top of that. The heavy, autonomous tasks are the part that flexes with usage. Knowing which layer a cost belongs to is half the battle, because the predictable part and the variable part are two different budgeting conversations.

The practical move is to tier. There’s a question I get a lot: when should you reach for a frontier model, and when does a cheaper one do the job?

A huge amount of work does not require deep reasoning. Match the model to the task, and you control the cost.

You can tier the people, too. An IT admin can decide who gets what. Maybe leadership has access to the most powerful models, management to another tier, and a lot of routine roles to simpler ones, based on the scenarios that actually come up, not on hierarchy for its own sake.

And you need visibility. Like the smart meter on the wall showing daily electricity use. That’s exactly the model. I get a notification the morning after I charge my electric car: “heads up, your consumption went up.” Agents should do the same.

When Scout hits a budget limit, it shouldn’t just stop. It should come back and say: I can keep going, but it’ll cost two, three, four times the budget, and here’s the outcome you’ll get for it. You decide. You can even set automated approval thresholds. The job is to find the balance between autonomy and control, without making everything need a sign-off, because the moment it does, people route around you.

AI productivity has a cost. The same way hiring an intern does. The discipline is knowing what each task is worth before you run it.

Is the answer always the cloud, then?

SL: No, and this is the part most AI strategies miss. For a lot of scenarios, a one-time hardware purchase beats a monthly cloud bill.

We’ve spent years giving white-collar workers terrible laptops, just enough to survive the day. Then we ask those same people to use AI to solve problems, and the cloud bill runs into real money. Meanwhile the hardware can now run genuinely capable models locally. A modern Mac, a Copilot+ PC, the things they can do on-device are remarkable, and that capability is a fixed cost you’ve already paid.

It isn’t always slower, either. Watch a cloud model stream its answer back line by line and you’re basically looking at a 1980s terminal. Plenty of tasks don’t need the frontier at all.

So build your AI strategy across a mix: on-device for the everyday, the cloud and the frontier models for the work that truly needs them. Invest a little more in the kit, and you’ll save.

Microsoft made that bet visible at its Build conference in June, unveiling a line of agent-first devices, including the Surface Laptop Ultra and the Nvidia-powered Surface RTX Spark Dev Box, built to run AI agents on the machine rather than lean on the cloud for everything. See how Microsoft is pitching the new lineup.

Models are changing weekly, and one frontier model was pulled at short notice recently. How do you build on something that moves that fast?

SL: Don’t marry a single model.

There’s a line I keep coming back to, from a former OpenAI product lead: the model you’re using today is the worst one you’ll ever use. I’ve lived it. We’ve had scenarios where we changed not one line of code and not one line of prompt, and the thing went from mediocre to amazing inside two months, purely because the model underneath improved. So if you’re close to your goal and falling short, sometimes the right move is to wait a month. Copilot is shipping new models at speed.

The flip side is exposure. If you build everything on one frontier model and it becomes unavailable, you’re stuck. So test your scenario across several models. That way you always have somewhere to go. This is also why Copilot has an automatic mode that reads your problem and routes it to the right model for the task. It protects your quality, and it quietly protects your cost.

If you were briefing a CIO who’s nervous about all of this, what’s the one thing you’d tell them?

SL: Be curious before you’re cautious. This is the moment to run a curiosity program. Pick people across the organization and let them use it. You’ll find creativity at every level, including places you’d never have looked. People who aren’t software builders are now pitching real products and getting them funded. That’s the upside you can’t see from a spreadsheet.

And don’t sit at the back. If you’re in one of these roles, your professional life should be driven by being at the frontier. Be the first one to touch it, to feel the value yourself. Then, once you’ve got the confirmation, move to the back and push everyone else forward. That’s the order. Lead from the front to learn, then lead from behind to scale.

Sébastien Levert is Principal PM Architect for Copilot, Agents and Platform at Microsoft, based in Montreal, and a six-time Microsoft MVP. He spends his days, in his own words, in one app, Copilot, connected to Work IQ, and his weekends hiking with his young son. You can follow his work on LinkedIn.