OpenClaw era demasiado peligroso para instalarlo. 7 meses después, Microsoft lo lleva a tu empresa.

[0:00] Back in February, I covered the week OpenCL broke everything. A security panic erupted across the entire industry and every enterprise IT team I talked to. And Satya Nadella, Microsoft CEO, compared it to a virus. Seven months later though, Microsoft is selling it to your employer. That's what Autopilot is. Microsoft took the agent that was too dangerous to install. They gave it an identity your IT department can audit and they turned it into enterprise software with a support contract. That's
[0:31] a big deal because it's going to reach way, way more working people than anything OpenAI or Anthropic ships this year. Not because it's better, but because it shows up everywhere. Inside Excel, inside Outlook, inside Teams, the places you're already working. So, the AI you have access to is about to start doing a different kind of work. Not answering your questions, holding an assignment while you're off doing something else. What you do with it is whether it matters. And that part that part doesn't come automatically, okay?
[1:02] Let me show you what this tool can do, what Autopilot can do, and the five things that are going to decide which side of the AI usage map you're going to land on. One quick thing before I go on. When I started this channel a few years ago, I didn't start it to give you more stuff to watch. I started it so you could take what I cover here and actually put it to work. And that's why I'm launching Nate's Library MCP. Every prompt kit, every guide, every tool that I have ever published, all the stuff I talk about right here on this channel is
[1:32] pulled into the AI you already use via MCP. So, your agent is the librarian effectively and you can point it at any problem you have in front of you and have it apply my thinking to your actual work. It's the next best thing to having me in the room with you. If you pay for Substack, you got it. And setup takes just a few clicks and links in the description below. So, if that's interesting to you, if you'd like to hear more of me around your work, see what I think about what you're doing, that's how you get it. And if you're not a Copilot user, as many of us also are not, or if you use other AI models and you're like, I'm snoozing, I'm skipping
[2:03] this one, Nate, don't do it because the takeaways I'm going to give you, it turns out they work for any way you work with AI models. They're just particularly applicable with Copilot, so you will learn something, too. I promise. Stick with me and let's have fun. Microsoft reported more than 450 million paid Microsoft 365 commercial seats in January, which is a fancy way of saying they reach like half a billion people. Employers pay for those accounts. IT departments administrate them, right? We know this. Our email, our meetings, our spreadsheets, our company documents, they all live in
[2:34] Microsoft. Compare that with OpenAI's February figure of more than 9 million paying business users. 9 million, not a billion, right? Microsoft's paid workplace footprint absolutely dwarfs that business user base. Those 450 million seats aren't all paying for AI yet, and Autopilot is still in preview, true, but Microsoft already has the workplace relationships through which it can sell all of this AI. And in company after company that I talk to, Microsoft is the AI you're allowed to
[3:05] use. Now, no one is saying how many of those roughly half a billion people only have access to Copilot or Autopilot now at work. I can tell you that for a lot of people watching this, for a lot of you, this is your reality cuz I hear it back from you. You tell me. When I sit down with you, when I hear it in the comments, I have Copilot, I can't access that nice stuff at work. The agent your employer approves matters way, way more to your working day than the model everyone's arguing about online. I've played with Autopilot, and it's frankly,
[3:36] it's not super intelligent. I'm not going to pretend it is. The value is not from the intelligence. That's the point. The value comes from the data it can use and how well you know how to use AI. That's where I think the future gets case shaped. People with access to the same tools get very different results. So, there's two slopes to the K, right? A K goes up and a K goes down. I want you to be on the up slope that's getting more capable. Suppose you're responsible for a customer renewal next month,
[4:06] right? The customer emails to ask whether last week's problem has been resolved. Then the Teams conversation, then the meeting notes where somebody made a commitment. You're digging through all of this. Who knows how long it takes? You message the owner, you wait for an update. Assemble a briefing, all with data, right? Internal data. You haven't made a renewal decision yet. You've instead spent the entire morning getting everybody ready to have the conversation with proprietary company data. Copilot has been useful for pieces of that work, of course, because it can summarize the thread, or it can help
[4:36] write the reply. People use it for that all the time. But you still remember what needs checking. You still find the missing information. You proactively keep things moving. Now, imagine an agent with an ongoing assignment to help keep that renewal on the tracks. It checks developments against the customer commitment. It prepares the next step. It follows up where you've given it permission to do so. So, essentially, with Autopilot, you are delegating some of the responsibility you used to carry
[5:08] between prompts. That used to be on your shoulders. Here's the first thing I do with that possibility. Give the AI, give Autopilot, a particular job aimed at something you care about and tell it what good looks like. I pounded this as a drum so long. You got to tell the AI what good looks like. You can start this with the Copilot you already have. Maybe you don't have Autopilot yet. Maybe your IT department hasn't turned it on. Well, pick the next piece of work you're actually responsible for. For example, ask it to identify the unresolved
[5:38] customer commitments. Ask it to show you Ask it to tell you what needs to happen before the renewal conversation. Then you look at what it gives you. Does the data it collected and synthesized help you act? Did Copilot distinguish a promise someone made from something the customer requested, or did it come comfortable taking the briefing the Copilot put together into a meeting? I have used Copilot. I find it's it's a batting average, right? You know, 50 60% of the time, yes. 40 50% of the time, no. If you have a good previous
[6:09] briefing, that number goes up. Show it the example and explain what's useful about it. Tell the model what you like. You're giving it away to judge whether it has completed the assignment. In this case, it's really not an intelligence problem. It is a way we use the model problem, which is the whole point of this video. Asking Copilot to summarize everything leaves the entire standard, what you think is a good summary, largely unstated. Yet, I see time after time when I look over someone's shoulder and they're using Copilot, that's kind of the prompt that they're putting in.
[6:39] Now, Copilot needs information to do its job. You can't do it any other way. Every model needs that. So, if you go to support tickets, for example, right? A support ticket might be routine, or it might be the condition on which the customer agreed to renew. As a product person, I've seen both. An agent that sees the ticket, but misses the implicit commitment somewhere else in the data, can give you a perfectly readable briefing that misses the entire point. This is where Microsoft's data position really matters long-term.
[7:10] Its Work IQ announcement describes a renewal where the sales record looks healthy, but service information and customer correspondence reveal a bunch of flashing yellow and red lights. So, they're thinking about this, too. The demonstration now starts with a person asking a question. It's evidence of the business context Microsoft wants us to connect to, right? They want us to think this way. It's not really a demonstration of Autopilot discovering the problem on its own. The opportunity here that Microsoft is trying to sell is that they can take the
[7:40] understanding inside your business and make it reusable and useful. They can recover value from data that would sit there otherwise. Your company already has customer records, access permissions, assigned owners, and decisions recorded in meetings. If the agent can use those relationships, you spend less time reconstructing your business for it. That's a real serious advantage even when the model itself isn't the smartest one. Instead of worrying about how smart it is, worry about what data it has access to. That's
[8:10] the Microsoft story. That brings us to number two. The second takeaway here out of five is to put the right information in front of the model and check that it actually used it. So, for the renewal, give it the current account plan, give it the customer commitment, give it the relevant service. You get the idea. Give it all the information it needs. Tell it which source takes priority when documents disagree. Most people do not do this. This is a tip that will work for you regardless of which AI you use. And that works for any AI model by the way. If your setup includes Copilot
[8:41] notebooks as an example, you can actually collect project references and instructions right there in a way that's super Copilot accessible, which is super helpful. And you have to rebuild the context every time by feeding it in a prompt, which is a nice step that's definitely correlated to like project files and other AI tools. Then, ask Copilot to link the evidence for the claims that matter. Ask it to put together essentially an evidence tree. Open the links. If it says the problem is resolved, does the source say that? Is it looking at this week's update or
[9:11] last month's plan? Have you caught that? That's not just a Copilot issue. Other models have trouble with time, too. When the answer is weak, check what information the tool had before deciding that the model can't do it. I can't underline that enough. This works for ChatGPT, it works for for Claude, it works for Copilot. Your company may not have connected the relevant system yet, or your account may not have permissions to read it. There are legitimate business reasons why this might not happen yet, but that's often the issue, is information access. Knowing what's missing gives you something specific to
[9:41] fix, something specific to raise with the IT department. That's also why this launch deserves attention from people who have never opened a terminal. So, Cursor, Claude Code, Codex have made delegating work super familiar to those of us who have a development background. I've been working in Claude Code and Codex for months and months and months now. Microsoft is bringing that working arrangement into the applications that the rest of the company already uses. You shouldn't need to become an engineer to get an agent involved in something
[10:11] you need, like an account review. And the open Claude connection that I mentioned at the top, that is literal, by the way. Microsoft describes Scout, the earlier name for Autopilot, as literally powered by open Claude's open-source tech. Microsoft is monetizing trust in Microsoft. The buyer can bring an agent in through a vendor whose contracts, whose support, whose administrative tools the company already knows. That doesn't guarantee that the product is going to work perfectly every
[10:42] time, right? It just gives Microsoft a much, much easier starting point than a new vendor asking for access to all your company's information. If you doubt me, look at the graph of adoption for Slack versus adoption for Teams. It is the graph when you want to understand the power of distribution at scale. Microsoft reported more than 30 million paid Microsoft 365 Copilot seats in July. Stronger since, I'm sure. Microsoft already has a really substantial paid-AI business off of something that a lot of us have not
[11:12] considered state-of-the-art, and it has a whole lot more of the workplace base right in front of it to sell into than any other player. It can just pick up the phone and call existing customers all day long to monetize. And my expectation, frankly, is that AI is going to stay Microsoft-shaped in a lot of enterprises unless its challenger gives those companies a really compelling reason to change. They'll bring it in when they're satisfied with security, with access and control, and they'll trust Microsoft. Microsoft doesn't need to win every model comparison to keep that relationship. It
[11:44] just needs to make the product useful enough and manageable enough that the company keeps expanding what it allows us to do with it. Which brings us to the third takeaway. Takeaway again works for everybody, not just Copilot users, not just Autopilot users. Once an assignment works, make sure you can make it repeat. So, a useful renewal briefing every single morning is a different working arrangement with your AI from remembering to ask for one when you're stressed. Start by getting a result you would
[12:14] actually use. Then, decide what would cause the agent to run again. What's the trigger here? And what it's responsible for in between check-ins. Microsoft Autopilot announcement describes an agent that has identity and memory and computer and workspace, right? Like it's the open claw setup. And when it's talking about what it does, it does a lot of this repetitive motion, which is why I bring it up here. So, the supplier review example, it spans preparation, it spans meetings, it spans follow-ups. That's the direction the company is
[12:45] going, right? An assignment continues and you can kind of keep an eye on it while you are driving something else. Now, Autopilot is in expanded private preview, which means it's not going to roll out overnight to half a billion people, and we shouldn't expect that. So, if you don't yet have Autopilot, you can actually use scheduled Copilot prompts, which are an earlier way to repeat useful work where your license and company settings allow them. And why do I say that? Because you can obviously repeat tasks in Codex, you can repeat it in Claude. Why am I saying Copilot?
[13:16] Because again, data access matters. What if you only can get the data in Copilot? That's how you do it. So, now listen, a scheduled briefing is not the same as that true autopilot sense of an agent independently pursuing a responsibility. I don't know when Microsoft is going to get to the point where they're ready to roll this out to everybody. But, it's coming. You got to be ready for it. And if we go back to the renewal story, you can specify which accounts Copilot should check, how often it should check, what counts as a material change, and when it should bring you in. Be clear about whether it
[13:47] can contact anybody or only prepare a draft. Most people stick with the latter. It's usually smart to start with. And look at the next run of this recurring report. Did it actually notice what changed? Did it produce another summary of what you already knew? That's a common failure mode. Repeating a weak assignment just gives you more weak output to read. Microsoft has an enormous advantage in where the work already happens. But, Microsoft doesn't have the field to itself. The strongest argument against assuming Microsoft wins comes from taking model improvement
[14:18] itself very seriously. Think about that renewal process. Some of the steps only exist because people need to make decisions. Others exist because someone has to copy information into a report, wait for someone to read it, hold a meeting, explain what changed. A sufficiently capable model might find, probably will find, a better way, a more efficient way to get the renewal ready. It can reconcile the records continuously. It can prepare the decision when relevant conditions change. Perhaps you don't need the weekly report or the meeting that exists
[14:48] to explain it. They're irrelevant. They're for human coordination. Now, the intelligence is starting to change the process you were planning to automate. That's the strongest case for a lab like Anthropic scaling into the enterprise indefinitely. Which as models improve, they may work out for themselves more efficient ways to do corporate work than the processes companies have already encoded. So, Anthropic's account of managed agents actually gives an example of this. And they talk about this a lot. Supporting machinery or harnesses built
[15:18] around an older model's limitations become unnecessary with a better model. Now, that doesn't mean by itself that Claude can correctly redesign any given company. I'm not trying to say that, but it does give you a reason to ask how much of today's workflows are going to remain necessary long-term. And this is why Claude getting into Excel and PowerPoint starts to matter. Claude has add-ins for both of those applications, by the way, and also Word. In that world, a company can keep its Microsoft
[15:48] documents while its employees increasingly give Claude the value, the assignment to work on. Microsoft ends up retaining the data set applications, but another company becomes the place we start when we want work done. Now, Microsoft can compete by leveling up their models, and they have done so, and those apps and workflows, so they're competing, too. But, really, the competition here is not over who uses Excel, it's over who understands the job well enough that we keep coming back to delegate work to it. This is a useful
[16:18] test for your own assignments. Once the agent can produce the briefing you asked for in our example, ask whether there's a better way to get the job done with less process. Have it explain what a proposed change would remove, what information it would need, how you would check the result. You may discover you've taught AI to reproduce a report nobody needs, or to repeat steps that you thought were needed that the AI didn't. That is a much more interesting outcome than getting the report written faster. It has a lot more leverage long-term. So, and by the way,
[16:49] underneath this it's not just about working more efficiently. The cost for companies and AI runs through all of this. If agents keep checking, preparing, and following up, companies need to know what they're spending tokens on, right? So, Microsoft's auto feature is a router that chooses models based on accuracy, based on speed and cost, with longer-running agent products have effectively usage-based billing, so you kind of bill by how long they go. There are different arrangements here, but the economics of repeated work are getting wrapped into the Microsoft AI product. So, in this situation, that
[17:21] sets up some interesting tensions. Now, a cheaper model is useful when it can reliably handle the assignment, but the hard part, the part that frontier models need to ask is knowing when it can. That's what a router is supposed to do. It's supposed to choose which model should handle the work and do it correctly. In his launch interview with Deirdre Bosa, Nadella described learned routers and selection based on how demanding the task is. And we need to ask ourselves, what information the system needs to correctly make that choice. I'll give you an example. Suppose you type, "What is this?" You
[17:52] may be pointing at an obvious syntax error that doesn't need a smart model or some obvious issue with an email or something. You might be 3 hours into a fiendishly difficult code problem and type, "What is this?" You mean something entirely different. The words are exactly the same to the router. If the router only sees those words, it doesn't have the information to distinguish the jobs. Now, Microsoft has not disclosed all the inputs that router system uses. So, what I'm describing is really the shape of the problem any router has to
[18:23] solve to be correct. And so, in this situation, maybe you don't necessarily need the entire conversation every single time because that would be very expensive, but you at least need the context that affects the decision to be piped correctly. You need the underlying assignment. You need the earlier failures, the consequences of a mistake. Perhaps your explicit choice or asked to spend more reasoning. Companies can do some of this ahead of time by evaluating familiar workflows, by assigning models to their steps, etc. And a system can learn from your behavior and corrections, potentially, right? But the
[18:53] irony is, when you're using an auto router, you may be burn more tokens figuring all of that out than you would just sending it to a smart model in the first place. Difficult new problems may in fact require substantial reasoning just to figure out what they are. And this is where other tools have to come into play. I covered Jev, recently. Jev makes structured judgments very efficiently from supplied context. That may be a play here. It makes sense to me, at least for bounded decisions, you can start to test. And people have absolutely used Jev for model routing
[19:23] problems. It's an emerging space, but it's one possible solution. But regardless of whether you use Jev, you use the router solution Microsoft is offering, you must define the classes of work properly to avoid using dumber models. I'll give you one more example. I get passionate about this. Tidying a scheduling email and deciding how to answer a customer contract, both produce an email. Calling both email writing does not establish that a cheap model can handle both of those problems. So, the takeaway here is you have to
[19:54] figure out how to match the reasoning to the difficulty of the job. That's takeaway number four. Separate routine preparation from consequential interpretation. If your setup lets you choose a stronger model or reasoning mode, you got to use that deliberately for the tough part. Tell it what you've already tried and why the obvious answer failed. A short follow-up should not erase the fact that you're working on a really hard problem. This is a new skill for us humans. I'm emphasizing this cuz this is not intuitive. We didn't train for this. If you can't select the model, you can
[20:25] still make the task at least a little bit clearer, right? You can break it into steps, you can identify where the system keeps failing. Don't keep submitting the same request and hoping the fifth answer works better. I promise you it won't. I see people doing this and then giving up in frustration. You got to focus on giving the tool information that changes its chances of getting the job done right. And count the time you spend repairing the result. That matters, right? You should know how much time you're putting into this, at least so you can tell IT what the problem is and get them to fix. If a cheap model misses the customer
[20:56] commitment and you redo the briefing, the low model bill that IT gets does not capture the cost of what just happened. And I hear a lot of those anecdotes. A more expensive model may save you money by removing work. A simpler model may be entirely adequate for a tested routine, but you're going to have to get better at routing to figure that out consistently. So, in sum, a routing is solvable in a bunch of different settings, but that doesn't mean that the edge cases in routing don't matter a ton and aren't very hard
[21:27] to solve. Both sides, Microsoft and the labs, benefit when useful AI work expands, but they're absolutely competing for the customer relationship, even if they do deals together. We have a new prize fighter in the ring. As I was thinking through this story, Meta announced Meta Enterprise Platform, bringing news, Meta Business Agent, and its developer tools into a broader enterprise push. This follows Muse, of course, the personal agent that can keep working on assignments that I've talked about on this channel. It's a huge deal.
[21:57] And Zuckerberg is explicitly building on the momentum of having a number one downloaded app to explicitly go after business customers. I don't know yet how that's going to go, but there's a concrete small business connection here. In August, Meta announced tools that connect Facebook and Instagram business information, ad campaigns, Google Workspace, all together, right? The goal here is that they can use AI, Meta can, to help repair business docs, to help run recurring tasks. That gives Meta a way to start with work its business
[22:28] [clears throat] customers, often SMBs, already care about, understanding customers, evaluating campaigns, deciding what to do next, and so So, my read is that Meta sees an opportunity to turn its existing business relationships into a much broader role. Microsoft has the relationship around running your workplace already. Meta has relationships around reaching your customers. Those jobs kind of meet and collide inside a business. Meta is a challenger here, but Meta is a challenger with a vast trove of data, recurring work distribution, and a great AI team to build from. So, the
[22:59] enterprise contest is getting bigger everywhere you look. It's getting bigger than Open AI versus Anthropic. And for you, the question is what you learn to do with the tools you use. And this is the issue I see in enterprise after enterprise. A small group, call it 5%, figure out how to get useful work out of AI. And that's an actual number I've heard over and over again. I'm not just saying that. They try things, they learn what the tool needs, they get a self-improvement loop going, and they're off to the races. They just keep going and grow and grow and grow. Plenty of other people have the same access and get a fraction of the value. I've
[23:29] already told you it's not superintelligent, but it doesn't need to be. But for so many folks who are in the business context, what I want to say is this. If you are waiting for a perfect product, you are postponing the learning that makes the next product more useful to you. And so, the fifth takeaway I have for you, maybe the most important, is think obsessively about how you improve the next time you use AI. When the renewal briefing misses a commitment, going back to our story, find out why. Was the source missing?
[24:00] Was your instruction unclear? Get curious. Did the model see the evidence and fail to understand it? Those all call for different fixes. It's up to you to get curious and figure out what those are, change the instructions, add data, change how you handle reasoning, save the correction so that future runs will use it and get better. Check that it helped. When something works, share that with a colleague. Now you have more than one person discovering what makes AI reliable on your team. You've kicked off a loop, right? That's the goal. And
[24:32] you've also learned something you can use on the next assignment, whether your employer stays with Microsoft or approves another tool. Your ability to specify work supply context and evaluate the result is a skill we're all going to need. Every job, I don't care. And it travels with you. That's the case-shaped future I'm talking about. I'm going to give you the five takeaways right in a row in a list here. Number one, you've got to define what good looks like. Pick a real assignment, explain what would make the output usable. Number two, you have to
[25:02] supply relevant context and data to your model. This is why Microsoft is in a good position. Point it at the right sources, check the evidence. Number three, make a good assignment something that can happen again. Recur, set schedule, set responsibility, set limits, inspect the next run. Number four, match the reason to the difficulty. This is where we get into the auto routing challenge, right? But really, you want to give hard work the context and capability that it needs. Number five, improve the next run. Fix
[25:33] the cause of the mess, save what works, and share it. Make sure you're in a learning loop. Microsoft is bringing open cloth into the working environment, and a huge number of us are going to start experiencing this for the first time. If Microsoft is the AI your company allows, this is your chance to start handing off more useful work. And I get passionate about this because we all deserve to learn this skill. Pick an assignment this week and push on it with the AI you have in front of you. Be on the upper slope of the cage. Don't be
[26:03] stuck thinking AI is bad because I have Copilot. No, think of it as a skill issue. You can use it to go places that you probably didn't think were possible if you can put data and context together in a way that makes sense. I'll see you next time, and tell me your most interesting use cases for Copilot, your experiences with Autopilot in the comments.