Agents Aren't Taking Your Jobs. They're Creating More Work Instead.

Look, we were told agents would take our Look, we were told agents would take our stuff off our plate. Some of us were stuff off our plate. Some of us were stuff off our plate. Some of us were told agents would take our jobs. It told agents would take our jobs. It told agents would take our jobs. It turns out agents aren't going to take turns out agents aren't going to take turns out agents aren't going to take our jobs because agents are generating our jobs because agents are generating our jobs because agents are generating so much work for humans. They're giving so much work for humans. They're giving so much work for humans. They're giving us more work to do. Everybody buying an us more work to do. Everybody buying an us more work to do. Everybody buying an agent is doing the same math. They're agent is doing the same math. They're agent is doing the same math. They're thinking the agent does the work, I need thinking the agent does the work, I need thinking the agent does the work, I need fewer people doing the work. They're all fewer people doing the work. They're all fewer people doing the work. They're all wrong. The volume numbers on open router wrong. The volume numbers on open router wrong. The volume numbers on open router don't lie. Agent token usage is up don't lie. Agent token usage is up don't lie. Agent token usage is up 14-fold between February and August. 14-fold between February and August. 14-fold between February and August. Agents now burn more than five tokens Agents now burn more than five tokens Agents now burn more than five tokens for every single one a human burns. And for every single one a human burns. And for every single one a human burns. And OpenAI is on the same page. It says its OpenAI is on the same page. It says its OpenAI is on the same page. It says its heaviest Codex users generate more than heaviest Codex users generate more than heaviest Codex users generate more than 60 hours of agent activity a day. 60 hours of agent activity a day. 60 hours of agent activity a day. Nobody's watching 60 hours of work. This Nobody's watching 60 hours of work. This Nobody's watching 60 hours of work. This video is about what happens after that. video is about what happens after that. video is about what happens after that. What happens after we discover the What happens after we discover the What happens after we discover the agents give us so much work. How does agents give us so much work. How does agents give us so much work. How does that affect us? How does that affect us that affect us? How does that affect us that affect us? How does that affect us at various company scales? How do we go at various company scales? How do we go at various company scales? How do we go from individual to small medium business from individual to small medium business from individual to small medium business to enterprise? We're going to cover all to enterprise? We're going to cover all to enterprise? We're going to cover all of it in this video. Yes, with specific of it in this video. Yes, with specific of it in this video. Yes, with specific examples. We create agents to take work examples. We create agents to take work examples. We create agents to take work away from us, and yet the evidence so away from us, and yet the evidence so away from us, and yet the evidence so far shows that agents create more work far shows that agents create more work far shows that agents create more work for humans. This video is about how for humans. This video is about how for humans. This video is about how agents are actually showing up in the agents are actually showing up in the agents are actually showing up in the workplace. The surprises we found so far workplace. The surprises we found so far workplace. The surprises we found so far and what that suggests for workers, what and what that suggests for workers, what and what that suggests for workers, what that suggests for small businesses, and that suggests for small businesses, and that suggests for small businesses, and what it suggests for enterprises. Look, what it suggests for enterprises. Look, what it suggests for enterprises. Look, right now it seems like most of us are right now it seems like most of us are right now it seems like most of us are saying agents handle more of the saying agents handle more of the saying agents handle more of the execution. I've heard that a lot. I've execution. I've heard that a lot. I've execution. I've heard that a lot. I've said it a lot. I've heard it back from said it a lot. I've heard it back from said it a lot. I've heard it back from leaders as well. Uh and that means that

leaders as well. Uh and that means that leaders as well. Uh and that means that people are deciding what runs, people people are deciding what runs, people people are deciding what runs, people are giving the agent what it needs, are giving the agent what it needs, are giving the agent what it needs, people are checking the results, and people are checking the results, and people are checking the results, and people are taking over when it fails. people are taking over when it fails. people are taking over when it fails. The question is, does that add up to The question is, does that add up to The question is, does that add up to more work for all of us? Who does that more work for all of us? Who does that more work for all of us? Who does that work? How does that work involve as work? How does that work involve as work? How does that work involve as agents get better? Let's talk about agents get better? Let's talk about agents get better? Let's talk about that, too. It at the personal level this that, too. It at the personal level this that, too. It at the personal level this is simple, right? One person can manage is simple, right? One person can manage is simple, right? One person can manage one agent for themselves. That's the one agent for themselves. That's the one agent for themselves. That's the whole open claw phenomenon, and we'll whole open claw phenomenon, and we'll whole open claw phenomenon, and we'll we'll talk about that. A small business we'll talk about that. A small business we'll talk about that. A small business needs multiple agents and often pays a needs multiple agents and often pays a needs multiple agents and often pays a vendor to manage it, and we're going to vendor to manage it, and we're going to vendor to manage it, and we're going to talk about specific examples there, too. talk about specific examples there, too. talk about specific examples there, too. An enterprise can hire engineers and An enterprise can hire engineers and An enterprise can hire engineers and deployment teams to build all of that deployment teams to build all of that deployment teams to build all of that agent management into the company and agent management into the company and agent management into the company and customize workflows, and we'll talk customize workflows, and we'll talk customize workflows, and we'll talk about what that looks like and why about what that looks like and why about what that looks like and why that's a different case than small that's a different case than small that's a different case than small businesses and why agents show up so businesses and why agents show up so businesses and why agents show up so differently in larger companies. Let's differently in larger companies. Let's differently in larger companies. Let's talk about small businesses first. Small talk about small businesses first. Small talk about small businesses first. Small businesses are getting mixed results businesses are getting mixed results businesses are getting mixed results from agents. You see that in the from agents. You see that in the from agents. You see that in the conversations I have privately with conversations I have privately with conversations I have privately with small business leaders, but you also see small business leaders, but you also see small business leaders, but you also see it in publicly reported results, which it in publicly reported results, which it in publicly reported results, which we'll talk about here as well.

we'll talk about here as well. we'll talk about here as well. Enterprise leaders I talk to, meanwhile, Enterprise leaders I talk to, meanwhile, Enterprise leaders I talk to, meanwhile, are talking about better returns, and are talking about better returns, and are talking about better returns, and we'll get to that a little bit later in we'll get to that a little bit later in we'll get to that a little bit later in the video and why that is. Let's start the video and why that is. Let's start the video and why that is. Let's start with small businesses and legal. Legal with small businesses and legal. Legal with small businesses and legal. Legal is an interesting case. Most law firms is an interesting case. Most law firms is an interesting case. Most law firms are small businesses, but of course, law are small businesses, but of course, law are small businesses, but of course, law is a very small part of the SMB category is a very small part of the SMB category is a very small part of the SMB category overall. Why is legal use growing overall. Why is legal use growing overall. Why is legal use growing quickly? Honestly, it's because legal is quickly? Honestly, it's because legal is quickly? Honestly, it's because legal is a verifiable domain. A verifiable domain a verifiable domain. A verifiable domain a verifiable domain. A verifiable domain is a domain where you can prove is a domain where you can prove is a domain where you can prove something is right or wrong, incorrect something is right or wrong, incorrect something is right or wrong, incorrect or correct. It's why coding has done so or correct. It's why coding has done so or correct. It's why coding has done so well. Well, law is like almost a version well. Well, law is like almost a version well. Well, law is like almost a version of coding for human letters and of coding for human letters and of coding for human letters and language, right? We are either correct language, right? We are either correct language, right? We are either correct and in line with the law or not. And and in line with the law or not. And and in line with the law or not. And sure, there's interpretation and lawyers sure, there's interpretation and lawyers sure, there's interpretation and lawyers have arguments back and forth, but it's have arguments back and forth, but it's have arguments back and forth, but it's a much more verifiable domain than a lot a much more verifiable domain than a lot a much more verifiable domain than a lot of the other ones out there. Perhaps of the other ones out there. Perhaps of the other ones out there. Perhaps that's why legal agentic AI use, as that's why legal agentic AI use, as that's why legal agentic AI use, as measured by usage of Codex, is up measured by usage of Codex, is up measured by usage of Codex, is up roughly 108 X since January. What I find roughly 108 X since January. What I find roughly 108 X since January. What I find practically is kind of the opposite.

practically is kind of the opposite. practically is kind of the opposite. When I talk to SMB owners, they're When I talk to SMB owners, they're When I talk to SMB owners, they're typically cash-strapped and they're typically cash-strapped and they're typically cash-strapped and they're time-poor, and they're trying to use the time-poor, and they're trying to use the time-poor, and they're trying to use the agents to buy time back. And as they get agents to buy time back. And as they get agents to buy time back. And as they get their agents to do more, they're their agents to do more, they're their agents to do more, they're spending more time on agent management. spending more time on agent management. spending more time on agent management. And it's not just legal, on open router And it's not just legal, on open router And it's not just legal, on open router agent token usage has grown about agent token usage has grown about agent token usage has grown about 14-fold between February and August and 14-fold between February and August and 14-fold between February and August and now exceeds human token use by more than now exceeds human token use by more than now exceeds human token use by more than five to one. Agents are using so many five to one. Agents are using so many five to one. Agents are using so many tokens these days and they are going to tokens these days and they are going to tokens these days and they are going to need to be managed by somebody. OpenAI need to be managed by somebody. OpenAI need to be managed by somebody. OpenAI meanwhile says its heaviest Codex users meanwhile says its heaviest Codex users meanwhile says its heaviest Codex users regularly generate more than 60 hours of regularly generate more than 60 hours of regularly generate more than 60 hours of agent activity every single day by agent activity every single day by agent activity every single day by running several agents at once. Now, no running several agents at once. Now, no running several agents at once. Now, no one is watching 60 hours of work step by one is watching 60 hours of work step by one is watching 60 hours of work step by step, you can't. The person is leveling step, you can't. The person is leveling step, you can't. The person is leveling up above the loop and that's one of the up above the loop and that's one of the up above the loop and that's one of the trends I see in individuals, it's kind trends I see in individuals, it's kind trends I see in individuals, it's kind of leaking into small businesses as of leaking into small businesses as of leaking into small businesses as well. The person is choosing jobs, the well. The person is choosing jobs, the well. The person is choosing jobs, the person is starting runs, the person is person is starting runs, the person is person is starting runs, the person is checking what comes back and the person checking what comes back and the person checking what comes back and the person is deciding what needs attention. Now, is deciding what needs attention. Now, is deciding what needs attention. Now, as agents get better, any given agent as agents get better, any given agent as agents get better, any given agent may require less supervision even though may require less supervision even though may require less supervision even though the number of agents and the amount of the number of agents and the amount of the number of agents and the amount of work continues to grow. And this is over work continues to grow. And this is over work continues to grow. And this is over the last 6 months, I think a classic the last 6 months, I think a classic the last 6 months, I think a classic example of what we would call the Jevons example of what we would call the Jevons example of what we would call the Jevons effect. The idea that making something effect. The idea that making something effect. The idea that making something more efficient can increase total usage.

more efficient can increase total usage. more efficient can increase total usage. It's why the human role doesn't simply It's why the human role doesn't simply It's why the human role doesn't simply shrink as the agent improves. Anthropic shrink as the agent improves. Anthropic shrink as the agent improves. Anthropic saw a similar pattern when they studied saw a similar pattern when they studied saw a similar pattern when they studied about 400,000 Claude code sessions and about 400,000 Claude code sessions and about 400,000 Claude code sessions and the division of labor is visible there the division of labor is visible there the division of labor is visible there as well. In a typical session, the human as well. In a typical session, the human as well. In a typical session, the human made about 70% of the planning decisions made about 70% of the planning decisions made about 70% of the planning decisions while the agent made almost all of the while the agent made almost all of the while the agent made almost all of the execution decisions. So, the person execution decisions. So, the person execution decisions. So, the person chose what ought to be built and then chose what ought to be built and then chose what ought to be built and then supplied the context and the agent ended supplied the context and the agent ended supplied the context and the agent ended up doing the work of like finding the up doing the work of like finding the up doing the work of like finding the files and writing code and changing files and writing code and changing files and writing code and changing things and running tests. This things and running tests. This things and running tests. This absolutely matches how I use Claude, absolutely matches how I use Claude, absolutely matches how I use Claude, right? And it matches how most of the right? And it matches how most of the right? And it matches how most of the people I talk to use Claude. The more people I talk to use Claude. The more people I talk to use Claude. The more experienced users did something very experienced users did something very experienced users did something very interesting. They approved more actions interesting. They approved more actions interesting. They approved more actions automatically, but they also interrupted automatically, but they also interrupted automatically, but they also interrupted the agent more often when it went in in the agent more often when it went in in the agent more often when it went in in wrong direction. On about 9% of wrong direction. On about 9% of wrong direction. On about 9% of conversational turns, experienced users conversational turns, experienced users conversational turns, experienced users interrupted the agent compared with only interrupted the agent compared with only interrupted the agent compared with only 5% for newer users who seemed more 5% for newer users who seemed more 5% for newer users who seemed more deferential to Claude. In other words, deferential to Claude. In other words, deferential to Claude. In other words, more experienced users watched fewer more experienced users watched fewer more experienced users watched fewer small steps and got better at noticing small steps and got better at noticing small steps and got better at noticing when the whole run was going kind of off when the whole run was going kind of off when the whole run was going kind of off the rails. Domain knowledge mattered the rails. Domain knowledge mattered the rails. Domain knowledge mattered here, too, and that makes me think of here, too, and that makes me think of here, too, and that makes me think of law again. Sessions led by experts law again. Sessions led by experts law again. Sessions led by experts averaged about 12 agent actions for averaged about 12 agent actions for averaged about 12 agent actions for every single instruction compared with every single instruction compared with every single instruction compared with only five for novices. In other words, only five for novices. In other words, only five for novices. In other words, knowing the problem mattered more than knowing the problem mattered more than knowing the problem mattered more than knowing how to code because the expert knowing how to code because the expert knowing how to code because the expert could describe the job really could describe the job really could describe the job really specifically. There's that verifiable specifically. There's that verifiable specifically. There's that verifiable domain again. Recognize a plausible domain again. Recognize a plausible domain again. Recognize a plausible mistake and tell whether result was mistake and tell whether result was mistake and tell whether result was usable. And this is what is beginning to

usable. And this is what is beginning to usable. And this is what is beginning to define the human job, especially for define the human job, especially for define the human job, especially for individuals and for small businesses. individuals and for small businesses. individuals and for small businesses. People are picking work the agent can People are picking work the agent can People are picking work the agent can finish, verifiable domains. They're finish, verifiable domains. They're finish, verifiable domains. They're giving it the files and permissions it giving it the files and permissions it giving it the files and permissions it needs to do that work, and they're needs to do that work, and they're needs to do that work, and they're telling it what a good result looks like telling it what a good result looks like telling it what a good result looks like up front and in advance. Then, they know up front and in advance. Then, they know up front and in advance. Then, they know when to let it go because they have the when to let it go because they have the when to let it go because they have the domain knowledge versus when to stop it domain knowledge versus when to stop it domain knowledge versus when to stop it and steer it. They know how to check the and steer it. They know how to check the and steer it. They know how to check the answer against whatever they want to answer against whatever they want to answer against whatever they want to check it against, right? Maybe it's the check it against, right? Maybe it's the check it against, right? Maybe it's the law code for lawyers, maybe it's a test law code for lawyers, maybe it's a test law code for lawyers, maybe it's a test of some sort, and they know what to of some sort, and they know what to of some sort, and they know what to change when the same mistake keeps change when the same mistake keeps change when the same mistake keeps happening. One person can carry all of happening. One person can carry all of happening. One person can carry all of that in these tiny businesses because that in these tiny businesses because that in these tiny businesses because the same person holds the goal and the the same person holds the goal and the the same person holds the goal and the context, the permissions. They hold all context, the permissions. They hold all context, the permissions. They hold all of that domain knowledge, right? If the of that domain knowledge, right? If the of that domain knowledge, right? If the agent ends up writing something poor, agent ends up writing something poor, agent ends up writing something poor, then the user can just say, "You This is then the user can just say, "You This is then the user can just say, "You This is crap. I'm not going to use it." If it crap. I'm not going to use it." If it crap. I'm not going to use it." If it takes research in the wrong direction, takes research in the wrong direction, takes research in the wrong direction, the the law firm partner can say, "Hey, the the law firm partner can say, "Hey, the the law firm partner can say, "Hey, this is not where I want to go." And the this is not where I want to go." And the this is not where I want to go." And the failure then is also constrained to that failure then is also constrained to that failure then is also constrained to that person's work. And that can be not true person's work. And that can be not true person's work. And that can be not true for a business when the business needs for a business when the business needs for a business when the business needs to depend on the result. And this is to depend on the result. And this is to depend on the result. And this is where there begins to be a divergence.

where there begins to be a divergence. where there begins to be a divergence. Failure stays at one-person scale when Failure stays at one-person scale when Failure stays at one-person scale when we use AI individually. Failure can get we use AI individually. Failure can get we use AI individually. Failure can get much larger even at the SMB scale when much larger even at the SMB scale when much larger even at the SMB scale when we use this in a business context. Part we use this in a business context. Part we use this in a business context. Part of what accelerates the failure case is of what accelerates the failure case is of what accelerates the failure case is when agents are used in ways that are when agents are used in ways that are when agents are used in ways that are not for verifiable domains. There's not for verifiable domains. There's not for verifiable domains. There's many, many small business owners in many, many small business owners in many, many small business owners in plumbing, in electrical, in in tax. plumbing, in electrical, in in tax. plumbing, in electrical, in in tax. Regardless of where you're picking, Regardless of where you're picking, Regardless of where you're picking, there's lots of small business there's lots of small business there's lots of small business categories that are not legal that you categories that are not legal that you categories that are not legal that you can go after here. They don't want a new can go after here. They don't want a new can go after here. They don't want a new job managing agents. They already have a job managing agents. They already have a job managing agents. They already have a job. The The business owner wants job. The The business owner wants job. The The business owner wants work done, wants outcomes, right? They work done, wants outcomes, right? They work done, wants outcomes, right? They want calls answered after hours. They want calls answered after hours. They want calls answered after hours. They want quotes sent while the customer is want quotes sent while the customer is want quotes sent while the customer is still interested. They want invoices still interested. They want invoices still interested. They want invoices collected. They want appointments booked collected. They want appointments booked collected. They want appointments booked correctly. All of these things that correctly. All of these things that correctly. All of these things that would help their business. They don't would help their business. They don't would help their business. They don't want the job to be managing agents want the job to be managing agents want the job to be managing agents because that would keep them from because that would keep them from because that would keep them from actually managing the business and actually managing the business and actually managing the business and helping it grow. Those kinds of outcomes helping it grow. Those kinds of outcomes helping it grow. Those kinds of outcomes are not outcomes that small businesses are not outcomes that small businesses are not outcomes that small businesses can purchase for 20 or 40 bucks a month. can purchase for 20 or 40 bucks a month. can purchase for 20 or 40 bucks a month. And that's the amount they're paying. JP And that's the amount they're paying. JP And that's the amount they're paying. JP Morgan looked at explicit AI service Morgan looked at explicit AI service Morgan looked at explicit AI service payments across 4.6 million small payments across 4.6 million small payments across 4.6 million small businesses and found that almost 2/3 of businesses and found that almost 2/3 of businesses and found that almost 2/3 of them are paying for AI at a rate of them are paying for AI at a rate of them are paying for AI at a rate of about 40 bucks a month. So, maybe two about 40 bucks a month. So, maybe two about 40 bucks a month. So, maybe two seats, not even pro on Claude or Open seats, not even pro on Claude or Open seats, not even pro on Claude or Open AI. And when you pay that amount, you're AI. And when you pay that amount, you're AI. And when you pay that amount, you're not going to be expecting phenomenal not going to be expecting phenomenal not going to be expecting phenomenal agent results that take care of all of agent results that take care of all of agent results that take care of all of your bookings for you. You're just not your bookings for you. You're just not your bookings for you. You're just not going to get that. What you're going to going to get that. What you're going to going to get that. What you're going to get is a glorified chatbot assistant.

get is a glorified chatbot assistant. get is a glorified chatbot assistant. Goldman Sachs surveyed 1,256 owners in Goldman Sachs surveyed 1,256 owners in Goldman Sachs surveyed 1,256 owners in its 10,000 small businesses program and its 10,000 small businesses program and its 10,000 small businesses program and only 14% only 14% only 14% of those owners said AI was fully of those owners said AI was fully of those owners said AI was fully integrated into their core operations, integrated into their core operations, integrated into their core operations, whatever that means. While 73% whatever that means. While 73% whatever that means. While 73% said they needed more training and said they needed more training and said they needed more training and resources to implement and evaluate it. resources to implement and evaluate it. resources to implement and evaluate it. That absolutely matches anecdotally what That absolutely matches anecdotally what That absolutely matches anecdotally what I hear from small business owners. They I hear from small business owners. They I hear from small business owners. They don't have time, they don't have don't have time, they don't have don't have time, they don't have capital, they don't have resources. They capital, they don't have resources. They capital, they don't have resources. They need help to actually get to the need help to actually get to the need help to actually get to the outcomes I just described, to more outcomes I just described, to more outcomes I just described, to more pipeline, to more booked appointments, pipeline, to more booked appointments, pipeline, to more booked appointments, to the to the results they're looking to the to the results they're looking to the to the results they're looking for that don't require them to have the for that don't require them to have the for that don't require them to have the job of managing agents. A $40 job of managing agents. A $40 job of managing agents. A $40 subscription is just not going to handle subscription is just not going to handle subscription is just not going to handle appointment booking. It's just not going appointment booking. It's just not going appointment booking. It's just not going to handle figuring out that something to handle figuring out that something to handle figuring out that something broke 3 days ago and proactively fixing broke 3 days ago and proactively fixing broke 3 days ago and proactively fixing it. That's not what those chatbots are it. That's not what those chatbots are it. That's not what those chatbots are designed to do. And the problem is that designed to do. And the problem is that designed to do. And the problem is that precisely because SMB owners are precisely because SMB owners are precisely because SMB owners are strapped for time, the person who strapped for time, the person who strapped for time, the person who understands the domain is also the one understands the domain is also the one understands the domain is also the one who has to hire and who has to sell and who has to hire and who has to sell and who has to hire and who has to sell and check the finances and who has to deal check the finances and who has to deal check the finances and who has to deal with whatever else broke that day and with whatever else broke that day and with whatever else broke that day and book the next day's work. So, frequently book the next day's work. So, frequently book the next day's work. So, frequently in these cases, what I see and what in these cases, what I see and what in these cases, what I see and what others see is a pattern of vendors others see is a pattern of vendors others see is a pattern of vendors ending up picking up AI management for ending up picking up AI management for ending up picking up AI management for SMB owners. And so, if you are hiring SMB owners. And so, if you are hiring SMB owners. And so, if you are hiring some to do AI transformation as an SMB some to do AI transformation as an SMB some to do AI transformation as an SMB owner, the vendor takes over the AI owner, the vendor takes over the AI owner, the vendor takes over the AI management job. The vendor chooses the management job. The vendor chooses the management job. The vendor chooses the workflow. The vendor connects the workflow. The vendor connects the workflow. The vendor connects the software. The vendor checks results and software. The vendor checks results and software. The vendor checks results and changes rules and shows where the changes rules and shows where the changes rules and shows where the customer is making or saving money.

customer is making or saving money. customer is making or saving money. Which kind of makes sense if you're Which kind of makes sense if you're Which kind of makes sense if you're strapped for time, but also limits the strapped for time, but also limits the strapped for time, but also limits the impact of what AI can do for you. impact of what AI can do for you. impact of what AI can do for you. Because for for for a second, like in a Because for for for a second, like in a Because for for for a second, like in a vendor pitch, this sounds good, right? vendor pitch, this sounds good, right? vendor pitch, this sounds good, right? The owner gets the result. The owner The owner gets the result. The owner The owner gets the result. The owner gets the appointments booked. The vendor gets the appointments booked. The vendor gets the appointments booked. The vendor handles everything around the agent. handles everything around the agent. handles everything around the agent. Except that the vendor has now taken Except that the vendor has now taken Except that the vendor has now taken responsibility for a core part of the responsibility for a core part of the responsibility for a core part of the business. Pocket OS, this is an example, business. Pocket OS, this is an example, business. Pocket OS, this is an example, is a small software company whose is a small software company whose is a small software company whose customers run car rental businesses. Its customers run car rental businesses. Its customers run car rental businesses. Its founder asked Cursor to handle a routine founder asked Cursor to handle a routine founder asked Cursor to handle a routine task in a test environment. Now, you're task in a test environment. Now, you're task in a test environment. Now, you're probably going to guess and yes, you're probably going to guess and yes, you're probably going to guess and yes, you're right. The agent hit a credential right. The agent hit a credential right. The agent hit a credential problem, found an account-wide railway problem, found an account-wide railway problem, found an account-wide railway token in another file. Account-wide token in another file. Account-wide token in another file. Account-wide tokens are dangerous. And decided to tokens are dangerous. And decided to tokens are dangerous. And decided to delete an entire storage volume. Nine delete an entire storage volume. Nine delete an entire storage volume. Nine seconds later, the live database and all seconds later, the live database and all seconds later, the live database and all of the ordinary backups were gone. of the ordinary backups were gone. of the ordinary backups were gone. Rental operators could no longer find Rental operators could no longer find Rental operators could no longer find reservations, they could no longer reservations, they could no longer reservations, they could no longer assign cars to customers arriving at the assign cars to customers arriving at the assign cars to customers arriving at the desk. Everything was offline. desk. Everything was offline. desk. Everything was offline. Now, in this case, Railway eventually Now, in this case, Railway eventually Now, in this case, Railway eventually was able to recover the data from an was able to recover the data from an was able to recover the data from an off-site disaster backup, but the poor off-site disaster backup, but the poor off-site disaster backup, but the poor founder of Pocket OS spent the next 30 founder of Pocket OS spent the next 30 founder of Pocket OS spent the next 30 hours working with every single client hours working with every single client hours working with every single client pen to keep them operating. The agent pen to keep them operating. The agent pen to keep them operating. The agent did that much damage in 9 seconds. And did that much damage in 9 seconds. And did that much damage in 9 seconds. And the human recovery, of course, took much the human recovery, of course, took much the human recovery, of course, took much longer. It took 30 hours, and that longer. It took 30 hours, and that longer. It took 30 hours, and that doesn't include all the pain and doesn't include all the pain and doesn't include all the pain and suffering and lack of trust that that suffering and lack of trust that that suffering and lack of trust that that wrecked up. And so, wrecked up. And so, wrecked up. And so, any vendor who is running an agent for a any vendor who is running an agent for a any vendor who is running an agent for a small business has to price in that kind

small business has to price in that kind small business has to price in that kind of risk because of the fragility of the of risk because of the fragility of the of risk because of the fragility of the systems they are managing. systems they are managing. systems they are managing. Someone has to check the workflow. Someone has to check the workflow. Someone has to check the workflow. Someone has to limit what the agent is Someone has to limit what the agent is Someone has to limit what the agent is able to read and write to. Someone has able to read and write to. Someone has able to read and write to. Someone has to notice those kinds of failures before to notice those kinds of failures before to notice those kinds of failures before they happen. Someone has to have a they happen. Someone has to have a they happen. Someone has to have a disaster plan that helps with recovery. disaster plan that helps with recovery. disaster plan that helps with recovery. I think stories like Pocket OS is part I think stories like Pocket OS is part I think stories like Pocket OS is part of why we see such a dramatic mixture of of why we see such a dramatic mixture of of why we see such a dramatic mixture of results for small and medium businesses. results for small and medium businesses. results for small and medium businesses. Legal may be off the charts in token Legal may be off the charts in token Legal may be off the charts in token usage. Legal is the exception. Most usage. Legal is the exception. Most usage. Legal is the exception. Most small businesses have less verifiable small businesses have less verifiable small businesses have less verifiable domains. They have less resources. They domains. They have less resources. They domains. They have less resources. They have less time. And they struggle with have less time. And they struggle with have less time. And they struggle with getting agents to work well, especially getting agents to work well, especially getting agents to work well, especially when vendors outside are managing when vendors outside are managing when vendors outside are managing agents. agents. agents. The vendor managing agents can work. You The vendor managing agents can work. You The vendor managing agents can work. You have to have a good vendor, strong have to have a good vendor, strong have to have a good vendor, strong alignment with what the business is alignment with what the business is alignment with what the business is going after. And in the meantime, the going after. And in the meantime, the going after. And in the meantime, the customer, the the the the small business customer, the the the the small business customer, the the the the small business may end up buying a cheap agent, but may end up buying a cheap agent, but may end up buying a cheap agent, but getting a useful product out of that getting a useful product out of that getting a useful product out of that cheap agent, getting useful work done, cheap agent, getting useful work done, cheap agent, getting useful work done, still often depends on the human that is still often depends on the human that is still often depends on the human that is wrapping around that agent. And so in wrapping around that agent. And so in wrapping around that agent. And so in that situation, the value is still often that situation, the value is still often that situation, the value is still often coming from the small business itself, coming from the small business itself, coming from the small business itself, not from the agent, because the agent not from the agent, because the agent not from the agent, because the agent isn't sufficiently embedded in the small isn't sufficiently embedded in the small isn't sufficiently embedded in the small business cuz the vendor doesn't know how business cuz the vendor doesn't know how business cuz the vendor doesn't know how to do that. So in that case, if the SMB to do that. So in that case, if the SMB to do that. So in that case, if the SMB can't get that value out, the vendor can't get that value out, the vendor can't get that value out, the vendor often charges enough to cover the value often charges enough to cover the value often charges enough to cover the value of what they got. The vendor walks away

of what they got. The vendor walks away of what they got. The vendor walks away seeing a success story. The vendor said seeing a success story. The vendor said seeing a success story. The vendor said we did our part and they part ways, but we did our part and they part ways, but we did our part and they part ways, but there's not a lot of value transfer. there's not a lot of value transfer. there's not a lot of value transfer. There's not an agent left that actually There's not an agent left that actually There's not an agent left that actually adds to the SMB experience. And again, adds to the SMB experience. And again, adds to the SMB experience. And again, why is that? Because the SMB didn't have why is that? Because the SMB didn't have why is that? Because the SMB didn't have the capital to invest to get this right the capital to invest to get this right the capital to invest to get this right by deeply integrating it, because they by deeply integrating it, because they by deeply integrating it, because they don't have the time to invest, because don't have the time to invest, because don't have the time to invest, because when they're outsourcing to a vendor, when they're outsourcing to a vendor, when they're outsourcing to a vendor, the vendor's making choices that favor the vendor's making choices that favor the vendor's making choices that favor the vendor even if inadvertently, and the vendor even if inadvertently, and the vendor even if inadvertently, and because not having a verifiable domain because not having a verifiable domain because not having a verifiable domain means a ton of work. A verifiable means a ton of work. A verifiable means a ton of work. A verifiable domain, by the way, that is just can I domain, by the way, that is just can I domain, by the way, that is just can I prove that the agent did it well or not prove that the agent did it well or not prove that the agent did it well or not relatively easily. And that's relatively easily. And that's relatively easily. And that's surprisingly hard to do in a lot of surprisingly hard to do in a lot of surprisingly hard to do in a lot of business domains outside coding. It's business domains outside coding. It's business domains outside coding. It's why legal took off. It's why I think a why legal took off. It's why I think a why legal took off. It's why I think a lot of other small business domains lot of other small business domains lot of other small business domains struggle. And if you're struggling, struggle. And if you're struggling, struggle. And if you're struggling, you'll need capital to make that domain you'll need capital to make that domain you'll need capital to make that domain verifiable and SMBs are classically verifiable and SMBs are classically verifiable and SMBs are classically short of capital. It's not easy to be an short of capital. It's not easy to be an short of capital. It's not easy to be an SMB and to have to invest in something SMB and to have to invest in something SMB and to have to invest in something like making a net new note domain like making a net new note domain like making a net new note domain verifiable. Okay.

verifiable. Okay. verifiable. Okay. That's the rough news for SMBs. That's That's the rough news for SMBs. That's That's the rough news for SMBs. That's what I'm actually seeing in the field. what I'm actually seeing in the field. what I'm actually seeing in the field. That's what the research is showing. That's what the research is showing. That's what the research is showing. Let's move to enterprises and see how Let's move to enterprises and see how Let's move to enterprises and see how the story looks a little different. An the story looks a little different. An the story looks a little different. An enterprise, of course, faces at least enterprise, of course, faces at least enterprise, of course, faces at least the same management cost if not more. the same management cost if not more. the same management cost if not more. And the difference is an enterprise is And the difference is an enterprise is And the difference is an enterprise is better capitalized and it has more better capitalized and it has more better capitalized and it has more people to keep that work inside the people to keep that work inside the people to keep that work inside the company. So when I talk to enterprise company. So when I talk to enterprise company. So when I talk to enterprise leaders who are putting agents into leaders who are putting agents into leaders who are putting agents into production, I hear consistently much production, I hear consistently much production, I hear consistently much better returns than I hear from small better returns than I hear from small better returns than I hear from small businesses. And I want to be very clear businesses. And I want to be very clear businesses. And I want to be very clear about why that is. It's not that there's about why that is. It's not that there's about why that is. It's not that there's a public study comparing return on a public study comparing return on a public study comparing return on investment between the two groups. This investment between the two groups. This investment between the two groups. This is part of the advantage of talking with is part of the advantage of talking with is part of the advantage of talking with a lot of folks as you start to see a lot of folks as you start to see a lot of folks as you start to see patterns emerge. What we do see in patterns emerge. What we do see in patterns emerge. What we do see in public backs this up though. OpenAI has public backs this up though. OpenAI has public backs this up though. OpenAI has compared its heaviest enterprise users compared its heaviest enterprise users compared its heaviest enterprise users with typical users. And in January, the with typical users. And in January, the with typical users. And in January, the heaviest enterprise users generated heaviest enterprise users generated heaviest enterprise users generated about 2.6 times as many output tokens about 2.6 times as many output tokens about 2.6 times as many output tokens per person as other users. By June, that per person as other users. By June, that per person as other users. By June, that difference had actually grown up to 8.3 difference had actually grown up to 8.3 difference had actually grown up to 8.3 times. In other words, enterprises that times. In other words, enterprises that times. In other words, enterprises that were firmly adopting AI were were firmly adopting AI were were firmly adopting AI were accelerating away from the group. Those accelerating away from the group. Those accelerating away from the group. Those firms were twice as likely to use firms were twice as likely to use firms were twice as likely to use plugins, six times as likely to be using plugins, six times as likely to be using plugins, six times as likely to be using skills, meaning they'd connected the skills, meaning they'd connected the skills, meaning they'd connected the models to more company systems and built models to more company systems and built models to more company systems and built more repeatable ways to use them.

more repeatable ways to use them. more repeatable ways to use them. Classic patterns of enterprise adoption Classic patterns of enterprise adoption Classic patterns of enterprise adoption and interaction there. Now, when you and interaction there. Now, when you and interaction there. Now, when you look at that from inside the firm, which look at that from inside the firm, which look at that from inside the firm, which I've done, you can hide a lot in an I've done, you can hide a lot in an I've done, you can hide a lot in an enterprise demo, just like you can hide enterprise demo, just like you can hide enterprise demo, just like you can hide a lot in SMB demo. But when you are a lot in SMB demo. But when you are a lot in SMB demo. But when you are inside the enterprise itself, you have inside the enterprise itself, you have inside the enterprise itself, you have to do a lot of work just to light to do a lot of work just to light to do a lot of work just to light something up and do it well. So you have something up and do it well. So you have something up and do it well. So you have to be able to get files from a really to be able to get files from a really to be able to get files from a really wide variety of places. You have to wide variety of places. You have to wide variety of places. You have to grant access across a wide variety of grant access across a wide variety of grant access across a wide variety of tools and do so within infosec policies. tools and do so within infosec policies. tools and do so within infosec policies. You have to monitor and log all of your You have to monitor and log all of your You have to monitor and log all of your agent runs. You have to consistently agent runs. You have to consistently agent runs. You have to consistently have a bar for deciding whether the have a bar for deciding whether the have a bar for deciding whether the agent run was any good. In an SMB, all agent run was any good. In an SMB, all agent run was any good. In an SMB, all of that might be one person's job. But of that might be one person's job. But of that might be one person's job. But in an enterprise, that's a manager's in an enterprise, that's a manager's in an enterprise, that's a manager's job, it's a security team's job, it's a job, it's a security team's job, it's a job, it's a security team's job, it's a quality control person's job, it's a quality control person's job, it's a quality control person's job, it's a product manager's job. There's just a product manager's job. There's just a product manager's job. There's just a whole team that's involved in setting whole team that's involved in setting whole team that's involved in setting this up and getting it to work. And so I this up and getting it to work. And so I this up and getting it to work. And so I think that although it leads to slower think that although it leads to slower think that although it leads to slower progress initially, progress initially, progress initially, ultimately it leads to deeper ultimately it leads to deeper ultimately it leads to deeper integration. Now, OpenAI and Anthropic integration. Now, OpenAI and Anthropic integration. Now, OpenAI and Anthropic have both realized that deeper have both realized that deeper have both realized that deeper enterprise integration takes time and enterprise integration takes time and enterprise integration takes time and takes investment. That's why I've talked takes investment. That's why I've talked takes investment. That's why I've talked about them both creating FDE deployment about them both creating FDE deployment about them both creating FDE deployment companies, right? OpenAI has one that companies, right? OpenAI has one that companies, right? OpenAI has one that has hundreds of engineers, deployment has hundreds of engineers, deployment has hundreds of engineers, deployment specialists, they've committed billions specialists, they've committed billions specialists, they've committed billions of dollars against it. Anthropic has a of dollars against it. Anthropic has a of dollars against it. Anthropic has a similar initiative. The job of those similar initiative. The job of those similar initiative. The job of those engineers, and I've met them, is to sit engineers, and I've met them, is to sit engineers, and I've met them, is to sit with leaders, sit with operators, sit with leaders, sit with operators, sit with leaders, sit with operators, sit with frontline staff, choose the

with frontline staff, choose the with frontline staff, choose the workflows, connect up the data and workflows, connect up the data and workflows, connect up the data and tools, set up the controls, test the tools, set up the controls, test the tools, set up the controls, test the systems, and then really ultimately systems, and then really ultimately systems, and then really ultimately ensure that work gets done differently, ensure that work gets done differently, ensure that work gets done differently, ensure impact. That's a ton of human ensure impact. That's a ton of human ensure impact. That's a ton of human work. And it is necessary to get agents work. And it is necessary to get agents work. And it is necessary to get agents to be productive. And an enterprise can to be productive. And an enterprise can to be productive. And an enterprise can afford to pay the bill, frankly, because afford to pay the bill, frankly, because afford to pay the bill, frankly, because the cost is spread across thousands of the cost is spread across thousands of the cost is spread across thousands of employees, the cost is spread across a employees, the cost is spread across a employees, the cost is spread across a large amount of revenue, and the upside large amount of revenue, and the upside large amount of revenue, and the upside is correspondingly big if they get it is correspondingly big if they get it is correspondingly big if they get it right. Whereas a small business would right. Whereas a small business would right. Whereas a small business would either ask the vendor to absorb that either ask the vendor to absorb that either ask the vendor to absorb that work or live with a much thinner work or live with a much thinner work or live with a much thinner implementation that has correspondingly implementation that has correspondingly implementation that has correspondingly less upside. That may be why I see less upside. That may be why I see less upside. That may be why I see enterprise leaders reporting better enterprise leaders reporting better enterprise leaders reporting better returns to me consistently from agents returns to me consistently from agents returns to me consistently from agents than SMB leaders. The base model is not than SMB leaders. The base model is not than SMB leaders. The base model is not the difference. They both have access to the difference. They both have access to the difference. They both have access to frontier intelligence. What changes is frontier intelligence. What changes is frontier intelligence. What changes is how much work the company has done how much work the company has done how much work the company has done around the agent. So, why is legal around the agent. So, why is legal around the agent. So, why is legal different? And can legal point the way different? And can legal point the way different? And can legal point the way for small business owners? I've talked for small business owners? I've talked for small business owners? I've talked about verifiable domains. I think that's about verifiable domains. I think that's about verifiable domains. I think that's important. And And even though we know important. And And even though we know important. And And even though we know that Codex users in legal drew like a that Codex users in legal drew like a that Codex users in legal drew like a 108-fold, we also should be aware that 108-fold, we also should be aware that 108-fold, we also should be aware that Codex is not the only one reporting this Codex is not the only one reporting this Codex is not the only one reporting this data. This is a widespread pattern that data. This is a widespread pattern that data. This is a widespread pattern that shows us widespread AI adoption in legal shows us widespread AI adoption in legal shows us widespread AI adoption in legal firms found that 71% of solo lawyers and firms found that 71% of solo lawyers and firms found that 71% of solo lawyers and 75% of small firms are already using AI

75% of small firms are already using AI 75% of small firms are already using AI in the legal field. Codex has their in the legal field. Codex has their in the legal field. Codex has their results. And those are statistics, but results. And those are statistics, but results. And those are statistics, but we also have actual stories from legal we also have actual stories from legal we also have actual stories from legal firms that we can learn from as well. firms that we can learn from as well. firms that we can learn from as well. Traubly and Singer is a small personal Traubly and Singer is a small personal Traubly and Singer is a small personal injury firm in Washington, D.C. It uses injury firm in Washington, D.C. It uses injury firm in Washington, D.C. It uses Even Up to review medical records and Even Up to review medical records and Even Up to review medical records and prepare the first draft of demand prepare the first draft of demand prepare the first draft of demand packages. The firm says that one packages. The firm says that one packages. The firm says that one workflow saves about 40 staff hours a workflow saves about 40 staff hours a workflow saves about 40 staff hours a month and lets it avoid another hire. month and lets it avoid another hire. month and lets it avoid another hire. The lawyers are not having to invent a The lawyers are not having to invent a The lawyers are not having to invent a new way to judge whether what the agent new way to judge whether what the agent new way to judge whether what the agent does is right. All they're doing is does is right. All they're doing is does is right. All they're doing is comparing the package with the records comparing the package with the records comparing the package with the records in the case, which they would have to do in the case, which they would have to do in the case, which they would have to do anyway. It's not more work to that. It's anyway. It's not more work to that. It's anyway. It's not more work to that. It's not that the agents are always perfect. not that the agents are always perfect. not that the agents are always perfect. Uh a recent benchmark gave systems 1,300 Uh a recent benchmark gave systems 1,300 Uh a recent benchmark gave systems 1,300 excerpts from legal briefs and asked excerpts from legal briefs and asked excerpts from legal briefs and asked them to find citation errors, and even them to find citation errors, and even them to find citation errors, and even the best systems missed subtle problems the best systems missed subtle problems the best systems missed subtle problems and too many steps. So, lawyers still and too many steps. So, lawyers still and too many steps. So, lawyers still have to verify all of that work, but the have to verify all of that work, but the have to verify all of that work, but the difference is they have a way to verify difference is they have a way to verify difference is they have a way to verify it. They have an existing process. They it. They have an existing process. They it. They have an existing process. They were doing it anyway. Why not do it with were doing it anyway. Why not do it with were doing it anyway. Why not do it with an agent? So, if we look at where an agent? So, if we look at where an agent? So, if we look at where lawyers are using AI, I think we get lawyers are using AI, I think we get lawyers are using AI, I think we get lessons for how SMBs can start to use AI lessons for how SMBs can start to use AI lessons for how SMBs can start to use AI more productively. Start with what more productively. Start with what more productively. Start with what you're already doing. Lawyers are not you're already doing. Lawyers are not you're already doing. Lawyers are not changing what they're doing when they changing what they're doing when they changing what they're doing when they review a demand package. They're doing review a demand package. They're doing review a demand package. They're doing what they would do anyway. The agent what they would do anyway. The agent what they would do anyway. The agent just created the work. Where are there just created the work. Where are there just created the work. Where are there areas inside SMBs where small medium areas inside SMBs where small medium areas inside SMBs where small medium businesses are already doing the work?

businesses are already doing the work? businesses are already doing the work? Maybe they're preparing a quote for the Maybe they're preparing a quote for the Maybe they're preparing a quote for the client, and and the manager has to client, and and the manager has to client, and and the manager has to review the quote either way. Maybe the review the quote either way. Maybe the review the quote either way. Maybe the agent can step in there. When you are agent can step in there. When you are agent can step in there. When you are trying to find areas for agents to yield trying to find areas for agents to yield trying to find areas for agents to yield value, and you don't have a ton of value, and you don't have a ton of value, and you don't have a ton of capital, I think there's a great lesson capital, I think there's a great lesson capital, I think there's a great lesson there. Find places where you would do there. Find places where you would do there. Find places where you would do the work anyway, and let the agent do the work anyway, and let the agent do the work anyway, and let the agent do the preparatory step. This also has the preparatory step. This also has the preparatory step. This also has implications for vendors. Vendors still implications for vendors. Vendors still implications for vendors. Vendors still need like Steve Jobs level product need like Steve Jobs level product need like Steve Jobs level product stubbornness, especially if they're stubbornness, especially if they're stubbornness, especially if they're working with SMBs. They cannot expect to working with SMBs. They cannot expect to working with SMBs. They cannot expect to drop a general agent into a company and drop a general agent into a company and drop a general agent into a company and hope that the owner magically finds hope that the owner magically finds hope that the owner magically finds value. Instead, you should be looking at value. Instead, you should be looking at value. Instead, you should be looking at where the owner is spending their time where the owner is spending their time where the owner is spending their time anyway and finding places to make the anyway and finding places to make the anyway and finding places to make the inputs easier for the owner so they can inputs easier for the owner so they can inputs easier for the owner so they can do the checks they'd be doing, not waste do the checks they'd be doing, not waste do the checks they'd be doing, not waste additional time, and magically get time additional time, and magically get time additional time, and magically get time back on the inputs and get value back back on the inputs and get value back back on the inputs and get value back that way. If you're not thinking that that way. If you're not thinking that that way. If you're not thinking that way, you're probably accidentally way, you're probably accidentally way, you're probably accidentally creating human work for managing agents creating human work for managing agents creating human work for managing agents when you're selling saving time for when you're selling saving time for when you're selling saving time for humans. And that does not add up over humans. And that does not add up over humans. And that does not add up over time. That does not work. This gets more time. That does not work. This gets more time. That does not work. This gets more complicated when we talk about teams.

complicated when we talk about teams. complicated when we talk about teams. And teams and teams and AI usage is And teams and teams and AI usage is And teams and teams and AI usage is something that is a consistent pattern something that is a consistent pattern something that is a consistent pattern across SMB and enterprise, and I think across SMB and enterprise, and I think across SMB and enterprise, and I think it's useful to understand how teams work it's useful to understand how teams work it's useful to understand how teams work with AI. In a field experiment with 776 with AI. In a field experiment with 776 with AI. In a field experiment with 776 Procter & Gamble employees. So, this is Procter & Gamble employees. So, this is Procter & Gamble employees. So, this is enterprise. One person using AI produced enterprise. One person using AI produced enterprise. One person using AI produced work that was about as good as a work that was about as good as a work that was about as good as a two-person team without it. That sounds two-person team without it. That sounds two-person team without it. That sounds great. But, among the best 10% of great. But, among the best 10% of great. But, among the best 10% of answers, only teams using AI were able answers, only teams using AI were able answers, only teams using AI were able to improve. So, solo workers using AI to improve. So, solo workers using AI to improve. So, solo workers using AI did not. In other words, the experiment did not. In other words, the experiment did not. In other words, the experiment showed that you still need teams of showed that you still need teams of showed that you still need teams of humans working with AI to iterate and humans working with AI to iterate and humans working with AI to iterate and get the very best answers to a problem get the very best answers to a problem get the very best answers to a problem versus just having individuals use AI. versus just having individuals use AI. versus just having individuals use AI. But, that yields a bunch of management But, that yields a bunch of management But, that yields a bunch of management questions, doesn't it? Who manages the questions, doesn't it? Who manages the questions, doesn't it? Who manages the agent when it's a team? Who keeps track agent when it's a team? Who keeps track agent when it's a team? Who keeps track of what the agent is doing? Who of what the agent is doing? Who of what the agent is doing? Who prioritizes what the agent does? Who is prioritizes what the agent does? Who is prioritizes what the agent does? Who is managing the agent's runs overnight? managing the agent's runs overnight? managing the agent's runs overnight? These are real questions that startups These are real questions that startups These are real questions that startups are grappling with right now that I are grappling with right now that I are grappling with right now that I talked to because you want to take talked to because you want to take talked to because you want to take advantage of those overnight runs. Your advantage of those overnight runs. Your advantage of those overnight runs. Your agent can run all the time. You want to agent can run all the time. You want to agent can run all the time. You want to have that value running overnight. how have that value running overnight. how have that value running overnight. how do you make sure that you're using that do you make sure that you're using that do you make sure that you're using that time well? To be very honest with you, time well? To be very honest with you, time well? To be very honest with you, most of the way that's answered in small most of the way that's answered in small most of the way that's answered in small businesses is that the owner steps in businesses is that the owner steps in businesses is that the owner steps in and decides where the agent will be and decides where the agent will be and decides where the agent will be allocated. And most of the time where allocated. And most of the time where allocated. And most of the time where the agent is allocated is very the agent is allocated is very the agent is allocated is very specifically against stuff that is specifically against stuff that is specifically against stuff that is verifiable and that is against the verifiable and that is against the verifiable and that is against the revenue line. Which kind of makes sense revenue line. Which kind of makes sense revenue line. Which kind of makes sense if you're the owner. We are going to

if you're the owner. We are going to if you're the owner. We are going to have to come up with more sophisticated have to come up with more sophisticated have to come up with more sophisticated answers as agents continue to scale at answers as agents continue to scale at answers as agents continue to scale at companies. Because what we're going to companies. Because what we're going to companies. Because what we're going to find is that it's not just one or two find is that it's not just one or two find is that it's not just one or two agents, it's not just a few agents. agents, it's not just a few agents. agents, it's not just a few agents. We're going to find more and more and We're going to find more and more and We're going to find more and more and more agents and agents scaling faster more agents and agents scaling faster more agents and agents scaling faster and faster at a company until even small and faster at a company until even small and faster at a company until even small medium businesses have to deal with medium businesses have to deal with medium businesses have to deal with agents at a relatively significant agents at a relatively significant agents at a relatively significant scale. Maybe 10 or 20 agents for a team scale. Maybe 10 or 20 agents for a team scale. Maybe 10 or 20 agents for a team of 10. Well, now one owner can't define of 10. Well, now one owner can't define of 10. Well, now one owner can't define all of that. You need to think through all of that. You need to think through all of that. You need to think through the coordination problems. This is going the coordination problems. This is going the coordination problems. This is going to impose a new agent management tax and to impose a new agent management tax and to impose a new agent management tax and a new set of startups that will be a new set of startups that will be a new set of startups that will be responsible for effectively helping you responsible for effectively helping you responsible for effectively helping you manage your agents. If you want to find manage your agents. If you want to find manage your agents. If you want to find a trend for 2027, there it is. That's a trend for 2027, there it is. That's a trend for 2027, there it is. That's it. But if we step back, let's step back it. But if we step back, let's step back it. But if we step back, let's step back from legal for a minute, let's step back from legal for a minute, let's step back from legal for a minute, let's step back from small medium business, let's step from small medium business, let's step from small medium business, let's step back from enterprise. Is there a larger back from enterprise. Is there a larger back from enterprise. Is there a larger lesson we can all learn here? Agents are lesson we can all learn here? Agents are lesson we can all learn here? Agents are doing more and more work. Agents are doing more and more work. Agents are doing more and more work. Agents are generating more work as they do more generating more work as they do more generating more work as they do more work. The work of management of agents work. The work of management of agents work. The work of management of agents is real. And so people are having to is real. And so people are having to is real. And so people are having to change their jobs to manage agents.

change their jobs to manage agents. change their jobs to manage agents. And they are doing so by moving what we And they are doing so by moving what we And they are doing so by moving what we would call above the loop. You know how would call above the loop. You know how would call above the loop. You know how we talk about humans in the loop for AI? we talk about humans in the loop for AI? we talk about humans in the loop for AI? Increasingly in 2026, people are above Increasingly in 2026, people are above Increasingly in 2026, people are above the loop. They're deciding what the the loop. They're deciding what the the loop. They're deciding what the agent is going to do, the way I agent is going to do, the way I agent is going to do, the way I described it as small business owner described it as small business owner described it as small business owner deciding it. They're deciding who may deciding it. They're deciding who may deciding it. They're deciding who may act with approval, which failures act with approval, which failures act with approval, which failures matter, what happens next. And if you're matter, what happens next. And if you're matter, what happens next. And if you're wondering from an enterprise perspective wondering from an enterprise perspective wondering from an enterprise perspective how this looks, if if you, you know, how this looks, if if you, you know, how this looks, if if you, you know, your CEO is not defining how everybody your CEO is not defining how everybody your CEO is not defining how everybody at Procter & Gamble's using the agents. at Procter & Gamble's using the agents. at Procter & Gamble's using the agents. No, they're not. This is done at the No, they're not. This is done at the No, they're not. This is done at the managerial level. And the managerial managerial level. And the managerial managerial level. And the managerial level, like L7 roughly at Amazon, right? level, like L7 roughly at Amazon, right? level, like L7 roughly at Amazon, right? Like that is a role that is becoming Like that is a role that is becoming Like that is a role that is becoming more and more, as I talk to these folks, more and more, as I talk to these folks, more and more, as I talk to these folks, an agent management role, an agent an agent management role, an agent an agent management role, an agent allocation role, as much as a human allocation role, as much as a human allocation role, as much as a human management role. And these teams, these management role. And these teams, these management role. And these teams, these managers, were never trained for that. managers, were never trained for that. managers, were never trained for that. And they are running into, in microcosm And they are running into, in microcosm And they are running into, in microcosm form, the same issues that small medium form, the same issues that small medium form, the same issues that small medium businesses are running into. And I've businesses are running into. And I've businesses are running into. And I've heard it over and over again. They talk heard it over and over again. They talk heard it over and over again. They talk about the problem of figuring out how to about the problem of figuring out how to about the problem of figuring out how to allocate agent runtimes against cheap allocate agent runtimes against cheap allocate agent runtimes against cheap token hours on the clock. How do you token hours on the clock. How do you token hours on the clock. How do you make sure that you are taking advantage make sure that you are taking advantage make sure that you are taking advantage of the fact that cloud is less used of the fact that cloud is less used of the fact that cloud is less used overnight in the US? How do you make overnight in the US? How do you make overnight in the US? How do you make sure that you are using your agents sure that you are using your agents sure that you are using your agents against spots that have the heart against spots that have the heart against spots that have the heart highest ROI, so you can show that you're highest ROI, so you can show that you're highest ROI, so you can show that you're using them that way when your VP comes using them that way when your VP comes using them that way when your VP comes calling? They have a lot of the same calling? They have a lot of the same calling? They have a lot of the same concerns, but they're doing it within a concerns, but they're doing it within a concerns, but they're doing it within a matrix organization. And because they matrix organization. And because they matrix organization. And because they have a lot of the same concerns, the have a lot of the same concerns, the have a lot of the same concerns, the human above the loop issues remain for

human above the loop issues remain for human above the loop issues remain for both small medium businesses and both small medium businesses and both small medium businesses and managers in enterprise context. They managers in enterprise context. They managers in enterprise context. They both need enough knowledge of the work both need enough knowledge of the work both need enough knowledge of the work to tell when the result is wrong. They to tell when the result is wrong. They to tell when the result is wrong. They need enough authority to change the need enough authority to change the need enough authority to change the system if it fails. They need to be system if it fails. They need to be system if it fails. They need to be confident that what they are doing is confident that what they are doing is confident that what they are doing is going to yield a better result for their going to yield a better result for their going to yield a better result for their business, whether it's a business unit business, whether it's a business unit business, whether it's a business unit they're managing in a larger company or they're managing in a larger company or they're managing in a larger company or the overall business as a small business the overall business as a small business the overall business as a small business owner. I'd like to close by laddering owner. I'd like to close by laddering owner. I'd like to close by laddering out a few of my takeaways. I talked to a out a few of my takeaways. I talked to a out a few of my takeaways. I talked to a bunch of small business owners for this bunch of small business owners for this bunch of small business owners for this video. I've talked enterprise managers. video. I've talked enterprise managers. video. I've talked enterprise managers. I've talked to enterprise leaders. What I've talked to enterprise leaders. What I've talked to enterprise leaders. What are we taking away here that we can all are we taking away here that we can all are we taking away here that we can all learn from as we head into the rest of learn from as we head into the rest of learn from as we head into the rest of 2026 and think about agents differently? 2026 and think about agents differently? 2026 and think about agents differently? Where is the puck going here? Agents are Where is the puck going here? Agents are Where is the puck going here? Agents are creating more work for humans, not less. creating more work for humans, not less. creating more work for humans, not less. And I think that's a misperception that And I think that's a misperception that And I think that's a misperception that we've generated when we've said agents we've generated when we've said agents we've generated when we've said agents sort of take work away from humans. No, sort of take work away from humans. No, sort of take work away from humans. No, the examples we see empirically at all the examples we see empirically at all the examples we see empirically at all scales are that agents generate more scales are that agents generate more scales are that agents generate more work for humans. Two, we don't have good work for humans. Two, we don't have good work for humans. Two, we don't have good established norms around what managing established norms around what managing established norms around what managing agents look like and the places where agents look like and the places where agents look like and the places where the norms are forming the fastest are in the norms are forming the fastest are in the norms are forming the fastest are in verifiable domains. I mentioned legal, I verifiable domains. I mentioned legal, I verifiable domains. I mentioned legal, I mentioned coding. You can look at mentioned coding. You can look at mentioned coding. You can look at healthcare as a verifiable domain as healthcare as a verifiable domain as healthcare as a verifiable domain as well. Where you have verifiable domains, well. Where you have verifiable domains, well. Where you have verifiable domains, where it's easy to see what's what's where it's easy to see what's what's where it's easy to see what's what's good and right and what's bad and wrong.

good and right and what's bad and wrong. good and right and what's bad and wrong. You have faster progress on agents and You have faster progress on agents and You have faster progress on agents and agent management norms because it's easy agent management norms because it's easy agent management norms because it's easy to see when the agent did something to see when the agent did something to see when the agent did something right. Where you don't is where things right. Where you don't is where things right. Where you don't is where things get sticky and where you start to see get sticky and where you start to see get sticky and where you start to see divergence between small medium divergence between small medium divergence between small medium businesses and enterprises. Enterprises businesses and enterprises. Enterprises businesses and enterprises. Enterprises have the capital and the resources to have the capital and the resources to have the capital and the resources to put agents into sticky, hard to get put agents into sticky, hard to get put agents into sticky, hard to get value situations and if there's strong value situations and if there's strong value situations and if there's strong management, strong will to win, they're management, strong will to win, they're management, strong will to win, they're able to deeply integrate agents and get able to deeply integrate agents and get able to deeply integrate agents and get value in those situations. Good examples value in those situations. Good examples value in those situations. Good examples are figuring out pricing are figuring out pricing are figuring out pricing recommendations, figuring out recommendations, figuring out recommendations, figuring out item recommendations in carts, figuring item recommendations in carts, figuring item recommendations in carts, figuring out how to write better product out how to write better product out how to write better product requirement documents that allow agents requirement documents that allow agents requirement documents that allow agents to develop against that work more to develop against that work more to develop against that work more effectively but still demonstrably meet effectively but still demonstrably meet effectively but still demonstrably meet customer requests and customer customer requests and customer customer requests and customer requirements. Another one that's sticky requirements. Another one that's sticky requirements. Another one that's sticky and difficult to validate that and difficult to validate that and difficult to validate that enterprises are good at is getting enterprises are good at is getting enterprises are good at is getting really good at deck creation. That's really good at deck creation. That's really good at deck creation. That's notoriously non-verifiable domain, notoriously non-verifiable domain, notoriously non-verifiable domain, right? What how good are PowerPoints? right? What how good are PowerPoints? right? What how good are PowerPoints? How do you know? Enterprises have the How do you know? Enterprises have the How do you know? Enterprises have the money to go after those kinds of money to go after those kinds of money to go after those kinds of problems for their specific house style, problems for their specific house style, problems for their specific house style, for their specific insights. This is why for their specific insights. This is why for their specific insights. This is why you see reports on PowerPoint and you see reports on PowerPoint and you see reports on PowerPoint and insight creation from financial firms, insight creation from financial firms, insight creation from financial firms, from McKinsey, from others because they from McKinsey, from others because they from McKinsey, from others because they have the capital to go after that.

have the capital to go after that. have the capital to go after that. Another lesson that I would call out is Another lesson that I would call out is Another lesson that I would call out is that regardless of scale, you should that regardless of scale, you should that regardless of scale, you should assume assume assume that agents are going to have more and that agents are going to have more and that agents are going to have more and more of an agentic management layer over more of an agentic management layer over more of an agentic management layer over time. We are going to start to see time. We are going to start to see time. We are going to start to see agentic management harnesses, just as we agentic management harnesses, just as we agentic management harnesses, just as we see agentic work harnesses today, but see agentic work harnesses today, but see agentic work harnesses today, but even there, because of Jeeves' paradox, even there, because of Jeeves' paradox, even there, because of Jeeves' paradox, we're still going to have plenty of work we're still going to have plenty of work we're still going to have plenty of work to do because what we're going to find to do because what we're going to find to do because what we're going to find is that the management harnesses allow is that the management harnesses allow is that the management harnesses allow us to abstract up a level, make it a us to abstract up a level, make it a us to abstract up a level, make it a little bit easier to manage agent little bit easier to manage agent little bit easier to manage agent execution runs, and that allows us to execution runs, and that allows us to execution runs, and that allows us to manage more agents, and then in turn, we manage more agents, and then in turn, we manage more agents, and then in turn, we will be expected to do that. The will be expected to do that. The will be expected to do that. The implications for all of us are profound. implications for all of us are profound. implications for all of us are profound. We have to develop the skillset of being We have to develop the skillset of being We have to develop the skillset of being able to manage agents. If we are able to manage agents. If we are able to manage agents. If we are vendors, we have to develop the skillset vendors, we have to develop the skillset vendors, we have to develop the skillset of recommending of recommending of recommending where agents should be placed with small where agents should be placed with small where agents should be placed with small medium businesses with accuracy, so medium businesses with accuracy, so medium businesses with accuracy, so we're not generating more work for them. we're not generating more work for them. we're not generating more work for them. If we are entrepreneurs, we have to If we are entrepreneurs, we have to If we are entrepreneurs, we have to learn how to build agents for specific learn how to build agents for specific learn how to build agents for specific use cases that are verifiable, that use cases that are verifiable, that use cases that are verifiable, that regardless of scale, businesses are just regardless of scale, businesses are just regardless of scale, businesses are just going to eat up and say, "I I can get going to eat up and say, "I I can get going to eat up and say, "I I can get that. I can get behind it. I can get that. I can get behind it. I can get that. I can get behind it. I can get value out of it." And if we are simply value out of it." And if we are simply value out of it." And if we are simply working in a big company, and we're working in a big company, and we're working in a big company, and we're expected to somehow use AI, what this expected to somehow use AI, what this expected to somehow use AI, what this video should remind you is one, video should remind you is one, video should remind you is one, these things aren't taking your job.

these things aren't taking your job. these things aren't taking your job. Two, you will have to learn how to Two, you will have to learn how to Two, you will have to learn how to manage agents to get promoted. And manage agents to get promoted. And manage agents to get promoted. And three, if you are managing agents, the three, if you are managing agents, the three, if you are managing agents, the problems you're facing as a manager in a problems you're facing as a manager in a problems you're facing as a manager in a large business are extremely analogous large business are extremely analogous large business are extremely analogous to the problems in a small medium to the problems in a small medium to the problems in a small medium business. In a sense, you can think of business. In a sense, you can think of business. In a sense, you can think of it as a fractal problem. Enterprises it as a fractal problem. Enterprises it as a fractal problem. Enterprises have many small medium business units have many small medium business units have many small medium business units that they're all managing at once, and that they're all managing at once, and that they're all managing at once, and although you may have more resources although you may have more resources although you may have more resources than an SMB owner, you still have the than an SMB owner, you still have the than an SMB owner, you still have the same strapped for time, got to show same strapped for time, got to show same strapped for time, got to show value pressures that an SMB owner has. value pressures that an SMB owner has. value pressures that an SMB owner has. And so, there's things we can learn And so, there's things we can learn And so, there's things we can learn across different scales. I hope this has across different scales. I hope this has across different scales. I hope this has been helpful to you. If you want to dive been helpful to you. If you want to dive been helpful to you. If you want to dive farther, if you want to understand farther, if you want to understand farther, if you want to understand better how agents are actually showing better how agents are actually showing better how agents are actually showing up, I have a full article on that on the up, I have a full article on that on the up, I have a full article on that on the Substack. I also have a guide to Substack. I also have a guide to Substack. I also have a guide to managing agents day-to-day, which I managing agents day-to-day, which I managing agents day-to-day, which I think everybody should pay attention to. think everybody should pay attention to. think everybody should pay attention to. So, if you want to go deeper on what I So, if you want to go deeper on what I So, if you want to go deeper on what I think about managing agents and how to think about managing agents and how to think about managing agents and how to get into the application from today's get into the application from today's get into the application from today's video, that's where you can head. I'll video, that's where you can head. I'll video, that's where you can head. I'll see you next time. I hope this has been see you next time. I hope this has been see you next time. I hope this has been helpful. I hope the specifics have been helpful. I hope the specifics have been helpful. I hope the specifics have been helpful. I love telling these individual helpful. I love telling these individual helpful. I love telling these individual stories of companies. I love talking stories of companies. I love talking stories of companies. I love talking about what I hear from leaders. I hope about what I hear from leaders. I hope about what I hear from leaders. I hope this has given you insight today.

this has given you insight today. this has given you insight today. Cheers.