The AI Daily Brief: Artificial Intelligence News and Analysis - Botsitting: The Work Draining AI Gains

Starting point is 00:00:00 Today on the AI Daily Brief, we're talking about bot sitting and the hidden labor that comes with the AI transformation of work. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Super Intelligent, Mission Cloud, and Out Systems. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. If you want to learn more about sponsoring the show, send us a note at sponsors at AIDailybrief. Two other quick notes. First of all, check out training.Besuper.a.i. for the newly updated enterprise-grade versions of the executive catch-up and the executive agent leadership program. The executive agent leadership program is a six-week intensive that kicks off on Monday. So last chance to get in on that.

Starting point is 00:00:52 And finally, just to let you know, I am recording this a couple days early because of some end of the school year travel. So this will be a main only episode, but we will be back with our normal format on Monday. Today we're talking about a new report from Glein and the Work AI Institute that's part of their Work AI Index for 2026, and it's all about something called bot sitting or the hidden human labor of AI at work. Now, one of the things that you may or may not have noticed this year is that I've done a little bit less coverage of studies from, for example, consulting firms or enterprise-focused research houses, and there is an actual specific reason for that. There's actually a couple of reasons, but they all come back to my feeling that the paradigm has shifted so much between non-agentic and

Starting point is 00:01:37 agentic work that anything that's interacting with non-agentic work is largely irrelevant. Now, of course, if you are an enterprise AI leader, that's not the case. There are still lots of use cases that are non-agentic that are going to be valuable and productivity enhancing. But you guys know that I have a very strong bias towards being interested in opportunity AI, not just efficiency AI, and the big changes that I see happening. in terms of how we work, not just doing the same stuff we've always done a little bit faster. This report, however, starts to get into and name some new types of work that surround AI and

Starting point is 00:02:11 agents that I think is really valuable to call out and start to explore. So that's what we're going to get into. Now, let's start with the statistics that they use to set everything up. Their big banner tweet-worthy statistics are that 87% of digital workers now use AI at work with 75% saying it makes them more productive, saving them 11 hours per week through automation. Yet only 30, 13% say their organization is performing significantly better as a result. And these numbers are almost a perfect encapsulation of what I was just talking about. 11 hours per employee is nothing to sneeze at. AI reaching nearly full penetration with the vast majority of people saying it makes them more productive are also interesting things.

Starting point is 00:02:48 And of course, the contrast between that individual performance and organizational performance reinforces a story that will be very familiar and that we hear over and over again, which is that translating individual AI gains into larger organizational gains is very difficult and not at all implied just by using AI well individually. Now, interestingly, I disagree with the report's key argument about why only 13% of those workers say their organization is performing better. I believe that on a fundamental level, individual productivity gains, wherever they come from, do not inherently translate to organizational gains unless there is a mechanism to actually facilitate that transformation. The question is what specifically are people using those 11 hours

Starting point is 00:03:31 per week four, and how much does it have to do with actually advancing key company missions? The report's argument is that the gains are, in their words, being swallowed by a new, largely invisible form of labor. They continue, we call it bot sitting, the work required to make AI usable, including feeding it missing context, checking its outputs, debugging its mistakes, rerunning prompts and cleaning up the confident but wrong answers AI leaves behind. Workers now burn an average of 6.4 hours a week bot sitting. So basically the argument is the reasons that organizations aren't getting gains is that those 11 hours per week that people are saving

Starting point is 00:04:05 are being eaten in large part by the 6.4 hours a week that they now have to bot sit. My position is that even if they weren't spending those 6.4 hours a week bot sitting, you still wouldn't see a direct translation from individual productivity to organizational performance. But, and this is important, I think that this new, largely invisible form of labor that they're calling out is extremely important to understand as we figure out how to integrate AI. So let's talk more about bot sitting. One important note, certainly, is that this is a byproduct of AI being successful on an individual level. As the report puts it, workers are handing over bigger parts of their jobs to AI and want to hand over even more. Workers they surveyed said that AI now automates about 27% of their work output, and those workers,

Starting point is 00:04:50 workers on average expected it to climb to 35%. 57% said that they want AI to automate even more of their job than they think it actually will ultimately be able to. But that success comes with some challenges. Maybe the most interesting chart in the entire report is the chart about where AI time actually goes. The Work AI Institute organized people's AI use into three categories. The first was learning and building agents,

Starting point is 00:05:15 which is exactly what it sounds like. It includes building workflows, reading online discussions about how others are building and experimenting with new models. That represented about 27% of people's time. 36% of people's AI time was spent actively using AI, in other words, completing work with AI. But 37% was on what they are calling bot sitting. Now, interestingly, they break bot sitting into two different categories, unproductive and

Starting point is 00:05:39 productive. The productive part just kind of looks like what it means to manage increasingly autonomous AI. So that includes verifying high-stakes output. to ensure they're correct, iterating on a prompt to make the output meaningfully better, or adding domain context the AI couldn't have known. The report argues that these are basically good versions of bot sitting. The unproductive versions will be familiar to many of you, reloading the same context into multiple AI tools, comparing outputs across tools because the

Starting point is 00:06:06 first answer wasn't good enough, or cleaning up AI generated work. Now, of the 6.4 hours spent bot sitting per week, 2.3 hours of that, representing 14% of the total overall, AI time was in feeding the AI context. Another 2.2 hours a week went to supervising outputs, 1.7 hours a week went to debugging, with about 10 or 15 minutes a week on other things like cleanup or switching tools. Now, interestingly, they also introduced the concept of an exhaustion multiplier, writing that for every 10% more time workers spend feeding AI context, they're 25% more likely to report feeling worn out by it. In other words, bot sitting is unfun, unglamorous grunt work, and it can have negative consequences. In fact,

Starting point is 00:06:48 In fact, the report found that frequent bot sitters, which they defined as respondents who spend 40% or more of their AI time on bot sitting activities, i.e. above the median for AI users, 73% were more likely to be actively hunting for another job. Now, in terms of who is bot sitting more, one culprit is simply a higher volume of work. Heavy AI users are more likely to report frequent bot sitting than light users, which I think makes intuitive sense. But the report argues that actually tool sprawl. In other words, the number of different AI tools workers use is another big culprit with workers who use multiple AI tools 35% more likely to report frequent bot sitting. They also found that right now, 60% of workers are

Starting point is 00:07:28 rerunning the same prompt across multiple tools because the first output wasn't good enough, all of which adds up to what they call the AI toggle tax. Now, outside of just feeling overwhelmed or burning out, there is another even more pernicious outcome of bot sitting, which the report labels bot shitting. Now, I had my editors bleep it there, but I don't think it takes too much imagination, to add the H to bot sitting to understand the term. Effectively, the process of bot sitting turning into bot shitting is when workers start to cognitively offload too much to AI. Now, this can happen even without bot sitting. We've all seen others and felt probably on ourselves the temptation to hand over more and more of our thinking and judgment to the AIs we use.

Starting point is 00:08:09 As we start to trust the outputs more, especially around common or routine tasks, we stop checking the outputs, we stop verifying the sources. And in an enterprise context, many admit that they start to ship the first output that looks good enough instead of pushing for one they can actually explain, defend, and stand behind. Now, of course, these things can happen even if we're not spending a ton of time bot sitting, but when you add the additional layer of burnout and frustration that comes with that bot sitting work, this can become even more likely. The report writes, bot shitting is rarely a single bad decision or a reckless click. It's usually a slow surrender of agency, one shortcut at a time. First, workers stop fully understanding the output, then they

Starting point is 00:08:48 stop interrogating it. Eventually, they stop feeling responsibility for it at all. And not only does this come with offloading understanding or offloading judgment, people also offload responsibility. The report writes that when AI generated work fails, 40% of workers blame AI and only 29% admit that it was their own fault. They call this an example of moral disengagement, writing, it's the gradual mental process by which people stop holding themselves accountable for harmful or careless behavior. Heavy AI users they found are 3.4 times more likely than light users to blame the tool when something goes wrong. So the cycle that they're identifying looks something like this. First, the organization deploys AI for whatever combination of good

Starting point is 00:09:29 reasons and signaling reasons that might be. Next, bot sitting rises as workers absorb the labor of making AI usable. Think feeding it all that context and checking its mistakes. Third, fatigue sets in. People realize that their work is no longer doing the work, but checking to make sure that AI has done the work. That leads to the bot-shitting phenomenon as the fatigued workers take shortcuts. And as that behavior goes up, we get to the fifth stage where those unverified outputs move upstream. And finally, the cleanup piles up as bad AI-assisted work creates more rework downstream. One of the most important AI questions right now isn't who's using AI. It's who's using it well. KPMG in the University of Texas at Austin just analyzed one of the one of

Starting point is 00:10:14 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers. They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news, these behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at KPMG.com slash us slash sophisticated. That's KPMG.com slash us slash sophisticated. Today's episode is brought to you by the new executive agent leadership program, produced by superintelligent and by frequent AIDB operators guest, Newfar Gaspar.

Starting point is 00:10:56 To tell you a little bit more about the executive agent leadership program, here is Newfar. The best predictor of agent adoption in an organization is how hands-on their leaders are. Talking about agents is completely different than building them. Our participants, ICs, all the way to C-Suite have built working agent fleets, governance frameworks, and the playbooks to scale it. Executive Agent Leadership is the evolution of Enterprise Claw. Everything we've learned across three cohorts rebuilt for right now. The token economy, security, vendor resilience, and architecture to lead agent adoption at scale.

Starting point is 00:11:34 The next cohort of the Executive Agent Leadership Program is signing up now and we'll launch on June 29th. You can find out more at training.b super.com. The average enterprise is spending $11.5 million on AI this year, and most of them can't prove a single dollar came back. What does AI actually look like when it produces ROI? Ask the healthcare company that just made their payment processing 320 times faster. Or the law firm whose document research went from three months to 10 minutes. Or the contact center who reduced wait times by 99%. These are real Mission Cloud customers with real results.

Starting point is 00:12:08 Mission Cloud is a CDW company and an AWS premier tier partner. They're the AI-first outcomes-obsessed AWS experts who build AI solutions that drive your business forward. Whether you're flooded with AI ambitions but no idea where to start or six months into a deployment that's going sideways, they've seen it and they've fixed it. Stop burning your budgets on AI that doesn't produce results. Start at missioncloud.com. This episode of the AI Daily Brief is brought to you by OutSystems, a leading agendic systems platform built for the enterprise. Organizations all over the world are building, orchestrating, and governing agendic systems. on the OutSystems platform and with good reason.

Starting point is 00:12:45 OutSystems open and unified platform allows teams to architect, deliver, and scale governed agenic systems with agility. Teams of any size and technical depth can use OutSystems to build, deploy, and manage AI apps and agents quickly and cost effectively without compromising reliability and security. Without systems, you can rapidly launch ideas from concept to completion. It's the leading Agendic Systems platform that is unified, agile, and enterprise proven, allowing you to accelerate growth, reduce operational friction, and deliver real enterprise impact with AI. OutSystems. Build your agentic future.

Starting point is 00:13:22 Now, one reason that I think this is interesting to explore now is that another part of my disinterest in many of the studies that have been published this year is that at least some meaningful portion of their data collection happened before the full ascendancy of agentic work. Remember, even the most vanguard AI users weren't really using agents fully and agentic coding tools and things like that until the very end of last year or, beginning of this year. And in fact, the 6,000 people that they surveyed for this report were answering those questions in December of last year and January of this year, meaning that this probably wasn't people using Claude Code to go run autonomous agents. There is an interesting

Starting point is 00:13:58 argument, though, that in this particular case, rather than nullifying these results, or making them not relevant, a repeat of this study might find this behavior even more amplified. One of the things that the study found is that the smarter the tool, the sloppier the worker. They write that among ChatGPT, Claude, Gemini, and Microsoft co-pilot, the tools whose workers report the biggest productivity gains, which was ChatGPT at 67% and Claude at 59% were also the tools whose users reported the most bot shitting, at 71% and 92% admitting to it at least monthly, respectively.

Starting point is 00:14:30 As we really start to move from efficiency AI to Opportunity AI, where people are doing things that were not possible for them before, it creates this whole new category of potential bot shing, where it's not that people are being lazy, but that they don't actually have the capability to verify or check the output. I feel this acutely every time a new model comes out and people race to figure out if it's better at coding or not,

Starting point is 00:14:53 and I try it out as well, but can only use very impressionistic views of the outputs as my way to understand because I've never coded before. Now, very clearly, in my estimation, that does not mean that only the people who previously could do what the AI can do now should be using those tools. Democritization of skills is a key value of this entire shift.

Starting point is 00:15:11 and yet it does inherently point to this new challenge that is going to be a part of integrating agents at scale. So how to deal with this? I think that the right way to look at these problems is not at core about people doing something wrong. I think that they are important consequences and artifacts of the transition that we're all experiencing and we're in effect always going to happen in some way, shape, or form. Part of our work in this transition is figuring out how to deal with these exact challenges. The Work AI Institute in Glean argue that we need a new human infrastructure of AI that can't be bought but has to be built.

Starting point is 00:15:48 They argue that it has to be built at three levels, how individuals work with AI, how teams manage with it, and how organizations design with and around it. First of all, they designate a group called high AI achievers. These are people who report that AI has improved both their productivity and the quality of their work. So what do they do that's different? The first thing they do is that they're a little bit more particular about where they use AI. While it's the case that basically everyone uses AI for the core of their work, i.e.

Starting point is 00:16:16 developers using it to write code or analysts using it to crunch numbers, low AI achievers spend roughly half of their AI time performing their core job tasks, whereas high AI achievers spend closer to a third. It's the difference between 48% and 38%. Now, frankly, right away, I think we're going to see why these problems are so challenging and they don't have easy answers. The report is arguing that the difference between effectively better AI users and worse AI users is better AI users still doing more of their core work themselves, but this feels to me like a very, very temporary state of affairs. Take for example in coding. They found that low AI achievers in coding used 46% of their AI time on core tasks of coding versus high AI achievers using 37% of their

Starting point is 00:17:00 AI time on core tasks. But again, this was back in December of last year in January of this, this year. For a significant number of advanced coders at this point, AI is doing all of their previous core task. When the coders who created Claude code basically don't code anymore, it kind of throws this whole idea into a tizzy. Also, the entire paradigm of AI time spent kind of ceases to make sense in an agentic paradigm. In other words, if you give Codex or ClaudeCode a slash goal command that's going to work overnight, do you count all of its time? working as part of your AI time spent, the entire artifice of this construction just sort of ceases to make sense in an agentic world. And yet I do think that if you are taking away an

Starting point is 00:17:46 important part of this, it's that people who are good at the stuff that they're doing are still being discerning about what AI is going to be used for and trust themselves and their judgment to still lead the AI whatever that leadership actually means. The second characteristic of high AI achievers was actually bot-sitting more, but orienting it towards the productive. They found that high AI achievers were more than twice as likely to rate AI itself as a valuable teacher. Effectively, they use bot sitting not only as a way to stop bot-shitting, but as a way to improve their own work with AI. Which gets to the third point, they reinvest that AI dividend, i.e. the 11 hours they saved on average, into new skills, not just more work. Now what about

Starting point is 00:18:30 team architecture. Once again, high-achieving AI teams, i.e. those who report both productivity increases and quality of work increases, do things differently than low-achieving AI teams. One difference is while they treat AI as a teammate, they keep accountability on the people. For example, when AI underperforms, rather than just giving up on AI and doing the task themselves, high AI achieving teams are far more likely to run the same prompt in other AI tools, or add more context and try again. High-achieving AI teams see AI adoption spreading peer-to-peer. here not just top down. In a buried but incredibly important statistic they found, when a leader uses AI, it makes the average employee 2.4 times more likely to adopt it. When a direct teammate uses it,

Starting point is 00:19:13 it makes them 3.2 times more likely to adopt it. But when a cross-functional teammate adopts it, that makes the average employee 5.6 times more likely to adopt. This is worthy of an entire exploration all on its own. The report starts to explore this asking why do cross-functional teammates carry so much weight, and basically argues that it's because this is where the rubber of AI hits the road of actual organizational messiness. They continue, because they're painfully aware of the coordination tax of work, the bottlenecks, the silos, the duplicated efforts, the dropped balls. So when they build an AI workflow or an agent, they aren't designing for some tidy fantasy version of the work. They're designing for the messy version that they have to deal with,

Starting point is 00:19:52 where the marketer needs the data the analyst hasn't pulled yet, and the engineer needs the spec the product manager hasn't written yet. Their workflows spread because they survive contact with real work. The third distinguishing feature of high-achieving AI teams is managers using AI to cut coordination costs and reinvest that time in people. High AI achieving managers delegate 32% more of their time on coordination to AI. As the report puts it, they aren't using AI to replace management. They're using it to clear away the administrative sludge that gets mistaken for management. And with the take that I love the report writes, that sounds like bad news for managers, but we think it's the opposite. The best managers don't try to compete with AI

Starting point is 00:20:32 on coordination work. They delegate the coordination work to AI, using AI to draft the status update, route the request, summarize the meeting. And they reclaim precious time for the work they ought to be spending more of their time doing, coaching, developing, and inspiring their people. And the dividends of this are enormous in terms of employee trust. When asked whether they were comfortable with AI playing a role in performance reviews, pay decisions, and termination decisions, Workers with good managers, which the report defines as managers whose direct reports would recommend them as managers, are basically twice as trusting as those with bad or average managers. For example, 53% of workers with good managers are comfortable with AI playing a role in performance reviews.

Starting point is 00:21:15 That number drops to 26% for workers with bad or average managers. On pay decisions and termination decisions, it's the same story, with workers with good managers being about twice as likely to be comfortable with AI playing a role as those with bad or average managers. coverage managers. Finally, on the third level, the organizational level, the report tries to understand what transformative organizations do differently. They define transformative organizations as those organizations whose employees report that AI has significantly improved their organizations' performance and outcomes, not just their individual performance and outcomes. Basically, this is the 13% who say their organization is performing significantly better because

Starting point is 00:21:51 of AI, and the Work AI Institute is asking, what are the 13% are doing that the other 87% aren't. The first thing is that they have a bias towards relevant metrics versus vanity metrics. Transformative organizations spend much more time trying to measure things like quality of work, productivity and output, and time saved. And here's a small one that's really interesting. Employees of transformative organizations are much, much more likely to have visibility into their own AI usage. While only 40% of workers in non-transformative organizations report that they can see their own AI usage data, 71% of workers in transformative organizations report that they have access to,

Starting point is 00:22:27 that information. This makes AI feel like a feedback mechanism to improve rather than just a surveillance mechanism for who should be fired. Relatedly, transformative organizations make governance a living system. For example, while 55% of workers at non-transformative organizations say that those organizations review their AI policy regularly, that number jumps to 93% of workers at transformative organizations. When it comes to explaining the rationale behind AI policy, Only 57% of workers at non-transformative organizations say their org explains that rationale, and that number jumps to 91% for transformative organizations. This all translates to trust.

Starting point is 00:23:06 57% of workers in non-transformative organizations report trusting their company's AI strategy, and that number jumps to 93% for workers in transformative organizations. Now, in bad news for companies who think that AI strategy is basically just a vendor selection choice, that point of view is actually a hallmark of non-transformative organizations. as opposed to the ones who actually get value out of AI. Unsurprisingly, transformative organizations are much, much more focused on building good systems for enterprise context. And finally, and totally unsurprisingly, transformative organizations actually invest in their people, not just in the AI tools themselves. When asked whether their organization formally rewards

Starting point is 00:23:44 AI skills, the number was only 48% for workers in non-transformative organizations and jumped to 84% for workers in transformative organizations. When the question was whether their organization, provides enough AI training and support. Only 52% of workers in non-transformative organizations said they did, and that number jumped to 90% for transformative organizations. Now, as I mentioned before, part of why I wanted to take the time to actually dig into this report is that contra up many, many reports that I've seen throughout 2026.

Starting point is 00:24:12 I think that if you redid this one right now, with increased agentic adoption, you'd actually find, rather than contradicting what they found, I think almost everything in here would be significantly amplified. And the reason for that is simple. The work of AI transformation is transformation, not just implementation. It is about complex, messy change, not just in how you do, but what you do. It requires new systems, not just new tools. Bot sitting and bot shitting are all part and parcel of the transition that we will all be experiencing for many years to come. There are not

Starting point is 00:24:48 short, easy answers to these problems. There are only organizations willing to do the work and those that aren't. Great work to the teams at Work AI Institute and Clean. For now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always. And until next time, peace.

The AI Daily Brief: Artificial Intelligence News and Analysis - Botsitting: The Work Draining AI Gains

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.