Microsoft launches Copilot Cowork

value with Cisco. And without further ad ado, we have Charles Lammana from Microsoft in the research radio. Now he's in the TV room. Charles, how are you doing?

Good. Thanks for having me on, guys.

Thanks so much for coming on the show. Uh why don't you give us a little bit of an introduction on yourself and then we can get into the news.

Sure thing. I'm on the product and engineering side at Microsoft. Uh my team builds out things like agents, a bunch of the the platform work and our business applications and some low code stuff. So, a bunch of odds and ends

and what uh what's the latest and greatest in your world?

Yes, a lot of big news today. So, today we announced co-opilot co-work

which really takes co-pilot beyond chat. So, it can take actions and automate tasks and pick delegation. Basically, we had some new agent capabilities in the office apps and a bunch of agent management solutions.

Yeah. How are you thinking about uh the the open claw boom, people writing code locally versus Microsoft has Azure? Like it's so easy to throw everything in the cloud. And I feel like uh one of the barriers that a lot of people have is like, do I need to go get a Mac Mini? Can I just delegate to the cloud? Some of the apps don't just by default build an app in the cloud or they don't have functionality. How are you feeling about that trade-off these days? Yeah, I mean we're super excited by OpenClaw and all that means. I think it's like the resurgence of the PC. Like what's the best thing you can give an agent? A computer, you know, let it run the apps, let it log in as you, let it messages you. Uh I think like a Windows PC in the cloud is probably the main way people will ultimately use something like OpenClaw at work.

Yeah. How do you think about uh model selection? Obviously, Microsoft has a has a time-honored partnership with OpenAI at this point, but uh you know, Microsoft has their own AI lab and uh you can get pretty much any model on Azure. Do customers want to pick models? How do you think about uh what is what's the best tool for the job in the products that you're building? But then also, how much do you want to give to the user? Because there's sometimes when I go into a chat app and I'm like, I just want auto. You should be figuring this out. I I know the details of pro versus thinking, but I don't want to think at all.

Yeah. So, that's one of the things that we think is super important for co-pilot and a lot of our products is the fact that it's multimodel. And the reason is because the best model for a given kind of task is not always the same and it's not even the same from month to month or quarter to quarter. Like if I look at Microsoft 365 copilot, chat is like the main way people interact with it. Our auto router is going to put you in OpenAI models because that's what we think works best with our work grounding in our work context. But then if you go to co-work the new announcement today and you fire off a task and you want it to run for a long period of time in the background use a bunch of tools maybe generate some code use like the terminal and stuff that uses the clot family of models and I go to co-pilot one place and get both and then we're going to keep adding more models models like from MAI and other places where it makes sense but it's really about the experience the user wants not so much picking exactly the version or the type.

Yeah. Uh how have conversations been with business leaders in terms of deploying a product like co-pilot co-work because the the the the ramp up for software engineers was smooth and everyone was aware of it tracking everything and going from you know better autocomplete to agentic coding was was pretty intelligible at every step of the way for most software developers. But for someone who's lived their life in spreadsheets and email and maybe not really thought about building automations all that often, what does actual deployment of these tools look like?

Yeah, it's such a great question because I think there's like two pieces to it. The first is you have IT teams which are nervous with if you imagine every single employee spinning up a whole bunch of agents that can create documents, send emails, run workflows. Um, so they want to make sure they have the right governance and controls in place. That's why another big part of the announcement today was something we call agent 365, which makes it easy for IT teams to manage and deploy agents from Microsoft and other companies. So that's one part. And then you have the end users. How do you teach a typical Excel or PowerPoint or Outlook user this idea of agentic working? I mean, even in coders, it it's taken time and it's probably not fully diffuse. Um, so that's the thing that we think we're in a really unique position at Microsoft in terms of being able to do because we can bring things like co-work to users where they already spend their time

inside of the office apps, inside of Windows, and increasingly inside the co-pilot app.

Uh, and we can present it in a way where it's digestible to hundreds of millions of office users, not just tens of millions of coders.

Is the majority of your business model still seatbased at this point? because I imagine that there will be a transition at some point. How much of that transition is already happening?

Yeah, we're we're still very seat based on like the M365 side, but like the good news at Microsoft, we kind of know how to do consumption businesses because we got Azure, we got a bunch of other very large capacity businesses. So the fact that we have user licensing and then consumption and pay as you go licensing for capacity all integrated with one commercial kind of vehicle with customers, we think we're in a pretty good position to make that transition over the next couple of years.

There was some uh there was some chatter on X this weekend over uh how different labs are pricing tokens and plans potentially to take market share from each other. There's a little bit of a capital fight playing out. How are you communicating?

Yeah. More specifically competing with the application layer.

Yes. Yeah. And so and so uh my question is how are you talking to business leaders, CTOs about how they should think about forecasting token budgets and spend and compute budgets. If they if they're used to, okay, I have licenses and licenses maybe go up in price and my headcount goes up and I'm used to that. But all of a sudden it's like, well, everyone's paying $200 a month, but they might be $5,000 a month in a couple months, but also the models are getting cheaper and the inference is getting cheaper and the capabilities are going up and it's very hard to predict. What are you talking to? What are you telling to business leaders and uh CTO's?

Yeah. Yeah. So, I the the first thing is you have to embrace this technology otherwise you're going to be left behind right now because the productivity gains are just so significant. And if you go talk to a coder

and you told them, hey, I'm going to take away like GitHub copilot and the agent coding capabilities, they'd be like, I refuse to work in this environment. Like it's just it's just inhumane almost. So So I think the the same type of thing is going to happen for all information work, all office work. So

9 months from now, I think if you went to somebody and said, "We're going to take away your agentic tools like co-pilot co-work," they'd be like, "No way. I'm not going to go back to the old way of working." So, so I think the first is there's a degree of inevitability because the benefit is so large and there's such strong pull from the end users.

That said, you everybody has budgets. So,

what you're going to see and what you're already starting to see is a bunch of controls where IT teams can start to manage exactly how many tokens or how much dollars each user can spend. And they're going to allocate that in a way based on where they think they're getting return and diminishing return. and it'll probably be like a expense report approval hierarchy type thing. So you'll see all of that show up, but part of just doing work now is you need a token budget that that's the future of how we think about you how you budget for a team or a department.

Where do you have any exciting case studies on the nons software engineering agentic work side of the business? uh because when when agentic coding came to software development uh you know there was a shift in a lot of software developers uh workflows where they went from writing a lot of code to reading a lot of code reviewing PRs they they started operating at higher and higher levels of abstraction but the decision what to build how to prioritize things how to build features there's still conversations and software engineers are still doing work uh what does that look like in the context of of of a finance department or an operations department or a sales group like where when the rubber meets the road like where are you seeing like okay this this is ready this is the next uh this is the next sub area or category that will really really benefit from the productivity gains from agentic workflows

yeah I mean a general thing and then maybe a more specific thing uh what we're seeing is like you how do you measure productivity is itself a challenge I mean this is this predates AI but right lines of code, number of pull requests, number of bugs fixed. I don't think I've ever heard a good engineering leader say that's how I measure the success of my team, right? It shows up in other ways. Do you have a great product people love? Is your cost structure good? What's your install base look like? So, so I think we have to get better at capturing that final business outcome because that's what's going to get the CFO to open up the pocketbook and continue to invest uh in more tokens and more AI. So if I go then take that and apply to kind of other disciplines like you what you don't want to say is oh our sales people send a lot more emails or our finance professionals create a hundred more spreadsheets like

that's more like of a nightmare scenario for AI. So

one billion spreadsheets per company.

No no I mean yeah we saw someone on the timeline say that their their token budget went way up but their revenue didn't go up and so they're like I don't know if it's uh if it's worth it. like you got to show results.

Exactly. So that that is what the question is going to be everywhere. But the the good the good news is

um engineers historically have resisted clear measurement of impact because you know there's a little bit of art to it.

But like sales professionals or finance professionals or customer service and and support reps,

they are so metric driven. Like if I'm a sales rep, what's my quota? What's my attainment? What's my revenue per head? Every single sales department has all of those things instrumented. And we've seen with M365 Copilot, we kind of did a bunch of experiments and we published a white paper on it like an almost 10% increase in revenue per seller just by giving them this the family of M365 copilot tools and things like our sales extension. Like that's real dollars in the bank. Um same thing for like back office finance processes. We've seen people use like our agents to do like accounts reconciliation and we have like a great case study we did there where it was like 30% of all back office finance was automated out by this thing. So th those are examples of real topline and bottom line benefit. Uh and that that's what everybody's going to be talking about the next couple of years. It's not going to be you know emails, documents, files, etc.

Yeah. Yeah. Uh I mean you're in a you're in a high level business role yourself. I find myself doing a lot of like knowledge retrieval, deep research reports, but have you personally been using a gentic, I mean this product specifically in your day-to-day? Like even as something as simple as like a Sam Alman told us that he has a self like he has some sort of like to-do list that does itself or something. I still don't understand exactly what that product is, but how are you using these tools these days?

Yeah. So the for for like co-pilot coork for example I I've I've loved that like my favorite magic moment has been around calendar management.

I was telling the story to someone like

I go go look out 3 months in the future and try to see you know how crazy does my calendar and travel look like.

Interesting.

And I was able to decline 17 meetings for me.

Okay right there. Nothing gives me more joy than declining a meeting. You sold me.

Yes. And I say like I went to Outlook and just saw all the meetings disappear and it was like such a serene moment for me and and like it did a deep analysis. It went and said like who's my manager, who do I work with, am I optional or required, what's the topic and it recommended the 17 and I went through all 17 said you know what decline and it was right. I mean so like that's one thing that I use it for. Um like another thing that I like right now is I mean I'm an engineering background so I'm not going to go create the most beautiful content all the time. So I use it now to pull together information design docs, meeting notes and then go create either like architecture diagrams, presentations or spreadsheets or docs.

And I mean I know it's just more doc generation but it's very much more kind of like fire and forget. So it works really well.

Yeah. Yeah. I mean even if it's just like the bones of a of a PowerPoint presentation with the actual data prepopulated that would just previously be a lot of copy paste. Go run this analysis, copy it over, make sure it's the same font. like there's so much grunt work that can be automated with these tools. I'm super excited for that. Um I I where do you see the the design PowerPoint generation uh tool use? I mean, people have been talking about like it's hard to create an RL environment with a verifiable reward around a great PDF, but it's easy to do that with coding. Do you do you have like an internal data set of like the best PowerPoints ever created your that you're like rlinging against or where do you see that going?

Yeah, I mean I think that is something that there's a little bit of taste in all of this cuz like depending on the company, some companies like lots of infographics and lots of colors.

Yep.

Other companies like super fact-based, no images. So like, you know, there's a little bit of an eye of the beholder, but absolutely we have a a very robust harness and like gym environment where we're able to quickly evaluate models, prompts, tools, and our own like fine-tuning efforts to see if it's producing better content. So I think like this like uh PowerPoints, Word documents, spreadsheets, it's going to be like code

and you you're starting to see it um already, but it's going to be like code for sure.

Yeah. I mean it like it it a lot of these things are like XML under the hood anyway and and can be puppeteered like code anyway. Uh well congratulations on the launch. Uh Jord,

I hope you don't mind. I've been calling co-pilot coco.

Coco.

So I think hopefully you like it. Yeah, it's a very cute name.

But congrats congrats on the launch and great to meet you.

Thank you so much for taking the time to come on the show. Cheers. We'll talk to you soon. Let me tell you about MongoDB. What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI, own the data platform