给机器当保姆：Glean 的 Rebecca Hinds 谈职场 AI 背后隐形的人力劳动

Today, my guest is Rebecca Hinds, author of the bestseller *Your Best Meeting Ever* and head of the Work AI Institute at Glean, which has just published the new Work AI Index 2026 report. The report draws on a survey of 6,000 digital workers to describe the state of AI as it’s used and experienced by employees at companies operating well outside of the AI bubble.

The headline numbers are genuinely strange. 87% of workers now use AI. 73% say it makes them more productive, and, on average, they report saving 13 hours per week—a third of a full workweek. And yet, only 13% say their organization is performing significantly better as a result.

The report contributes two new terms to the AI discourse: bot-sitting and bot-shitting. Bot-sitting is all the unglamorous, untracked labor required to make AI useful—feeding it context, debugging its outputs, and cleaning up its messes—which the report finds consumes 6.4 hours per week, or roughly half of all the time that AI supposedly saves.

For those who are being asked to automate parts of their work that they’d rather do themselves—for example, a customer service representative who enjoys talking to people but is now being asked to supervise agents—this can be especially painful. Such alienation predicts both reduced engagement and increased turnover and helps explain bot-shitting, which is when people deliver AI-generated work that they can’t explain or defend.

In the extreme, business becomes farce: a perpetual-motion machine of AI slop. And, shockingly, in the survey, 69% admit to doing it—a number that reflects both the incredible progress that AIs have made and perhaps the amount of bullshit work that people are asked to do.

Obviously, one part of the solution is more integrated AI systems that have the context they need. My experience with my own deep-context system is that it’s dramatically reduced my own time spent bot-sitting. We discuss how Glean’s enterprise graph product is playing a similar role for enterprises.

Beyond that, we also consider what organizations can do to create a more functional AI culture, including how to use AI detection to protect the business without discouraging positive use, rewarding people monetarily for effectively collaborating on AI solutions, and perhaps most powerfully, aligning work to a meaningful shared mission.

My mission for this show, as you may know, is mostly to learn and to help others learn as much as possible. But lately, I've also been trying to entertain and delight you with original songs made with Suno, which we've been playing at the end of each episode. I've really enjoyed the comments that people have sent about these, and I encourage you to stay tuned to the end of this episode for a legitimately catchy tune with some outstanding, poignant, AI-written lyrics. It did require quite a bit of bot sitting to get it just right, but I do enjoy the final product, and I hope you do, too. With that, I hope you enjoy this groundbreaking look at AI as it's practiced in large-scale organizations throughout the English-speaking world with Rebecca Hinds, head of the Work AI Institute at Glean.

Rebecca Hinds, head of the Work AI Institute at Glean and author of the new Work AI Index 2026 report, welcome to the Cognitive Revolution.

Thank you so much for having me, Nathan.

I’m looking forward to this conversation. I think people like me who live very much in the AI bubble—which is kind of a social bubble, albeit a very online one, and also a day-to-day work bubble—have diverged pretty significantly from what the rest of the world is doing.

I think people like me run a bit of a risk of getting detached, especially because I work by myself largely these days, from what’s going on in the real world at real companies that are actually driving most of the economy, where not everybody has the luxury or the inclination to be a bleeding-edge early adopter, with all of the—I’d say more ups than downs, certainly, but still a mix of ups and downs—that come with that.

So today’s conversation is going to be a really good exercise in grounding and calibrating myself and understanding, in the bigger world, what the current state of play is.

I’d love to start just by getting a little bit of grounding from you, though. In terms of how you understand AI, I think so many AI conversations diverge early, or people can talk past each other if they’re expecting very different things in the near-term future of AI, and that’s not necessarily put on the table up front.

What is your expectation? Are you an AGI-short-timelines, AGI-soon person? Do you make analogies to other technology waves? Give us your zoomed-out view, and then we’ll zoom in on the report itself.

It’s a great question. I think, in many ways, what we’re seeing with AI is not unlike what we’ve seen with previous technologies, particularly when we think about change management. That is what we’re seeing above all else.

This is a human change just as it is a technology change. As we’ve seen with every other technology change, we underestimate the human piece. We underestimate resistance. We underestimate fear.

Now, AI has this very unique element in that the fear is more visceral. I think, in large part, as trivial as it might seem, it comes back to the naming. The fact that we’ve named this thing artificial intelligence—etymologically, it’s fundamentally in tension with human intelligence—and I don’t think organizations take that seriously enough.

Even if the technology works objectively, if employees don’t see it as a teammate, if they don’t see it in the context of something that can amplify or augment their skill set, they’re going to either resist or symbolically use the technology as opposed to meaningfully using it.

That’s what we’re seeing. We’re seeing pockets of excellence. We’re seeing individuals, teams, and organizations fundamentally transform themselves with this technology in very, very exciting ways. But we’re also seeing the vast majority of teams and organizations struggle.

They struggle to translate the individual productivity gains, which pretty much everyone is getting on some level, into real business- and team-level outcomes.

Makes sense. I might have a couple of follow-up questions regarding your expectations as we go, but tell me about the methodology. Everything we’re going to discuss here in terms of findings is downstream of basically two main data sources, as I understand it.

One is a big survey, which you can tell us more about. And then the other that caught my eye in the report is aggregated telemetry from the Glean platform. I’m interested in what that actually looks like and in understanding a little bit better the substance of that data, as well.

Sure. The main bulk of the data is the survey data. This was something that had been in the works for months and months. We fielded it in December 2025 and January 2026.

What’s novel about how we collected the survey data is that it was 6,000 knowledge workers: 3,000 in the US and 1,500 each in the UK and Australia. It was developed in partnership with our 8 founding members of the Work AI Institute, which is our internal research center at Glean.

What’s special is that each one of these 8 experts comes at the AI conversation a little bit differently. Some are very much focused on the psychology and the mindset around AI. Others are more focused on the technology and digital transformation. Others are more focused on organizational design: how do we think about this from a systematic organizational-design perspective?

We thought it would be important, given the transformative nature of the technology, to have a pulse into all of these different dimensions. So that’s the survey component.

We also had a lot of conversations. We’re talking with our customers at Glean, as well as organizations broadly, in terms of what works and what doesn’t work. That informed a lot of the narrative.

Then the aggregated, anonymous Glean telemetry data is something that’s very exciting because it’s objective. Triangulating both the subjective survey data, which is inherently biased for many different reasons, and pairing that with something objective, we thought, was important.

In particular, we looked at how we’re seeing adoption happen on the Glean platform: the importance of cross-functional adoption, and whether your manager, a team member, or a cross-functional team member adopts. The network effects associated with this technology are massive.

When we think about change management, overwhelmingly, right now we’re seeing top-down change happen: mandates and memos telling employees to use AI or else. The best organizations, the most effective ones, are having a bifurcated strategy.

Yes, top-down change is important. We absolutely need a policy. We absolutely need principles. We absolutely need to see the CEOs and executives using the technology. But bottom-up change is just as important, and finding these AI influencers or champions within your organization is so essential to activate meaningful change—not just symbolic change because the CEO has told us to use the technology.

This is a good moment to take a beat on Glean and what Glean does. I think our audience is generally AI-obsessed.

I think that’s the one commonality that we all share. Everybody has heard of Glean, at least. I will confess, though, that as an individual operator, I’m obviously not in the sweet spot of the target market, so I’ve never actually used it and I don’t know too much about what the experience is like, although I certainly have a sense. I also don’t know how you go to market, whether it’s a sort of enterprise sales model versus product-led, bottoms-up, or maybe it’s a hybrid.

Maybe give us the double-click on Glean so people can understand, in a little bit more of a functional, procedural way, how that data is being generated.

Sure. Glean is a work AI platform. Historically, we started in enterprise AI search, solving the problem of how we find the right, relevant information within an organization. This was pre-AI, pre-mainstream AI.

Now our platform has an intelligent assistant as well. Every employee has that intelligent teammate that deeply understands not only how they work, but also how the organization works. It starts to give you recommendations in terms of what you should be prioritizing each day. When you ask it questions, it knows enough about you and your job function to provide intelligent answers.

Agents are a big part of the platform as well: automated workflows that have that organizational context and are able to streamline tasks and cross-functional tasks across the organization. The real bread and butter of the platform is context. There are a lot of conversations right now around the importance of the context graph. Our data model enables us to surface that context and feed off of it in a really exciting way.

When AI is able to truly understand your work, your team’s work, and the organization’s work, you avoid generic answers and start to get into this really exciting territory of predictive AI and proactive AI, telling you what matters right now in the moment of your day-to-day work.

Cool. Interesting. The mix of company context and individual context is definitely a pretty live discussion. I’m thinking of Dan Shipper and the Every team, and how they’ve experimented with various versions of this and already evolved their approach in a meaningful way.

I have my own pretty elaborate personal AI infrastructure at this point that is kind of an extension of me in one lane, and then also increasingly trying to get it to be an employee that can do its own thing with more and more autonomy over time as well. Maybe let’s come back to that again.

I want to make sure we don’t bury the lead on the headline of the report, which I do think is interesting food for thought, at a minimum. I’ll give you the headline, and you give me the expanded version of it.

Eighty-seven percent of the people that you surveyed are using AI. Seventy-three percent say it makes them more productive. The one that really blew me away: on average, they say they save 13 hours a week thanks to their use of AI. That’s a third of a workweek. Some caveats to come, obviously, but that’s a lot of time—a really strikingly large number.

The flip side of this is that only 13% of the survey respondents say that their organization is performing much better than it used to based on all this AI that’s happening. What are the big headlines that jumped out to you? Give me the double-click on how we should start to understand that.

Sure. I think you’ve hit the nail on the head with your previous setup to the question. We’re seeing this massive disconnect between individual productivity gains and organizational performance.

This is survey data, so it’s inherently biased in terms of how people are reporting. Depending on how we asked multiple questions around time savings, and depending on the question, we saw ranges from 10 hours per week to 14 hours per week. On average, what we’re seeing is that people are reporting that, based on the work output they have fully automated with AI, it’s about 11 hours of time savings.

But when you ask them about organizational performance—whether they see their organization performing significantly better as a result of the technology—we see this gap. Just 13% of employees are saying that their organizations are performing significantly better because of the technology.

It’s a big disconnect that we’ve seen in multiple different studies. We’ve certainly seen it in the headlines: the budgets that don’t pay off, the ROI that doesn’t pay off. What’s novel about this report, this Work AI Index, is that we’re theorizing where the time savings is going, and we’re coining this phrase “bot-sitting” as the hidden human labor that is required to make the technology usable.

I think we all feel it on some level. It’s feeding AI context, overseeing the AI, debugging the AI, and cleaning up after the AI. It’s a massive chunk of labor—on average, upwards of 6 hours per week—that workers are spending on this bot-sitting activity, which is often tedious.

You make a point in the report to point out that not all bot-sitting is negative, certainly. There are healthy forms of bot-sitting, but more often than not, it is negative. It is tedious. It is exhausting. It is not rewarded, appreciated, tracked, measured, or incentivized within the organization.

I think that is feeding a lot of these failed expectations—the ambition-execution gap that we see across the board. We’re not recognizing the human labor, and we’re not recognizing that this isn’t, in most cases, a problem with the individual employee. This is a systemic issue.

We’ve invested in tools that don’t have that context. We’ve invested in tools in silos, and now we’re seeing AI sprawl and agent sprawl. It’s a whole host of different factors coming from multiple directions to feed this bot-sitting, which then turns into what we’re calling bot-shitting.

Bot-shitting is the dissipation of work that hasn’t been checked, that workers can’t explain. We see 40% to 41% of employees saying they ship AI work that they couldn’t explain if asked. This is bot slop, and I think it’s pervasive.

We often see polished nonsense come through. It looks polished, it looks finished, and it looks like we’ve done the assignment. When we dive a little bit deeper, it’s actually quite hollow. There’s very little substance.

That’s part of it—the most visible form—but it’s also the shadow AI, the use of AI that you can’t explain in terms of the output. It’s a whole host of different dimensions as well.

I guess, to give you my personal bot-sitting and near-miss bot-sitting experience in preparing for this conversation: I have an agent that runs every day and looks for new podcast bookings on my calendar. Its job is to—and it does have pretty good access, but this anecdote goes to show just how important connecting all the dots is—because for lack of one connection, the result came out not so well.

I’ve given this agent access to my Drive, and in my Drive I’ve got hundreds of these previous documents where I’ve written up outlines of questions. It can go in and look at all those that I’ve previously done and use them for inspiration. Of course, it can search the web and everything else.

It saw you pop up on the calendar, then went and did its thing. It did a pretty extensive deep dive into your background, your book—which we can touch on a little later as well—and the previous reports and things you’ve been involved with.

It missed one point, which is that there’s this new report that I had gotten a copy of in email from one of your teammates, not specifically sent by you. Basically, the whole thing was not going to work. Maybe 10% to 20% of it was a conversation we actually were going to want to have, but it missed 80-plus percent of the substance of what we were really going to be talking about.

You’d think I know better, and I didn’t send that off without noticing that. But if I had, it would have been textbook bot-shitting behavior, which I do think we should all be very much watching out for.

I got a much better result when I said, “Hey, you missed something really crucial, which is the report we’re actually going to be focusing on.” There’s still another round where I feel like, to do a good job and to be respectful to you as a guest and to the audience who’s trusting me with their time—which is obviously the one resource they can’t possibly get more of—I have to come having internalized that information and ready to make it my own.

Even though I did get a great jump and a very meaningful assist from the agent, and in some sense it was maybe good enough as an artifact that I could have shared and you would have said, “This is fine. Good, thorough outline of questions,” I still need to do that extra work to be able to show up and have the interaction in a way that hopefully allows me to do a good job and not feel like I’m making it up or reading it cold as I go.

So I definitely relate to those 2 different things. Fortunately, I’ve had very few instances where I’ve let something get away from me without noticing that the bots have gone haywire.

Hey, we'll continue our interview in a moment after a word from our sponsors. Today's episode is brought to you by Anthropic, makers of Claude and Claude Code. Over the last few months, Claude has helped me build and refine a personal deep context database that now contains all of my emails, Slack messages, tweets, DMs across platforms, video calls, and podcast transcripts going back a full 5 years. On top of that, we've now layered summary articles describing my relationship with hundreds of contacts, organizations, and ideas. And now that this exists, there's almost nothing that Claude can't help with. For tax season, I asked Claude to help me get organized. It went through my inbox, tracked down 1099s for all 10 of my part-time jobs, and built me a comprehensive report on my expenses and donations. For my angel investing, Claude can now draft investment memos in exactly the form that my venture fund requires, based on the calls I've had and the emails I've exchanged with the founders. And when someone needs a favor, Claude can often do it as well as I can. Recently, a friend reached out to ask if I know anyone who might be a fit for a role that he is currently hiring for. Initially, nobody came to mind, but then I thought to ask Claude, and sure enough, it identified two great leads. Claude is the AI for minds that don't stop at good enough. It's the collaborator that actually understands your entire workflow and thinks with you. Whether you're debugging code at midnight or strategizing your next business move, Claude extends your thinking to tackle the problems that matter. So, for problems worth solving, get started with Claude at claude.ai/tcr. That's claude.ai/tcr. And check out Claude Pro, which includes all of the features mentioned in today's episode. Once more, that's claude.ai/tcr.

And I wonder, what do you think is the root of this problem? Especially the bot-shitting one, where people are just putting work out there into their team environment that they can't even defend. What's the root of that? Is it that they don't understand that the AI is not always going to do a good job?

Is it that they're just alienated in the first place and fundamentally don't care? Obviously, there's an issue in general with not everybody being super conscientious. So maybe this is just a new flavor of not doing a very good job for some people. What can you tell me about the psychology? How do you understand the mistake people are making when they do that?

I think, as most things are, it's multifaceted, and in the report we theorize the cycle at play here. I do think it starts with bot-sitting. I do think it starts with all of this manual work that is often a reflection of all of the AI that's being deployed in our organization and the pressure to adopt.

Organizations are under massive pressure to adopt this technology and implement it. So you start to get more pressure to adopt. You have more of a need to bot-sit the technology. That is exhausting. It's exhausting because it takes up a lot of time, but it's also exhausting because you're not rewarded for it. And there's really not very much incentive in many organizations to bot-sit well.

What we start to see is “good enough.” In the research, they sometimes call it satisficing. Once you see an AI output that is good enough, well, that's often permission to ship it. In your case, you're an expert at your skill, at your craft, right? You're recognizing that good enough isn't good enough.

And so that's a big thing: bot-sitting is a precursor to bot-shitting. It's the exhaustion. Employees hit a breaking point. They can no longer bot-sit, and they start to bot-ship more and more.

I think a big part of this is the lack of context—the lack of context in the AI tools, the fact that so many AI tools don't speak to one another. In many organizations, employees want to use different models and different tools, and that's why I'm excited about Glean as a platform as well, because there's no situation in my mind that I can envision where we have a single-model, single-tool environment.

Employees want that choice. We're seeing models leapfrog each other left and center. Employees want different models, different tools for different use cases.

It becomes very complicated if you don't have that contextual layer to connect them and make sense of them. And not just in terms of how the dots are connected, but also in terms of recency—knowing that this report was published this week, this month, versus a report that was published 2 months ago or 2 years ago. That should be given different treatment. The authoritativeness of the content as well is very hard to discern in an enterprise context.

So I think context is the big feeder as well from a technical standpoint. And then you have all these perverse incentives in organizations. The token maxing, the rewarding, the clicks of the tool—that is a big contributing factor as well.

Yeah, I used to tell people, when I did any sort of AI advisory consulting, that you could do a lot worse than, as a leader, just watching your token consumption. But definitely don't tell the team that's how you're going to be measured.

Yeah, so it's really super easy to cheat on that. It's amazing. That was probably years ago when I was saying that, and it's funny to see that people are still shooting themselves in the foot that way.

I guess, in terms of understanding my fellow human, I struggle a little bit with the idea that this bot-sitting work is so onerous. My attitude—which I don't expect everybody to share, but I'll give it to you for compare and contrast—is that I used to have to do stuff, and now I get to have AI largely do the stuff for me. I still have to make a contribution, but I definitely get a lot more done a lot faster.

I also get to learn a lot more about AI and what it can do, which I find to always be an interesting question unto itself, and just do a lot more because I'm able to take on so many more different things. I'm able to learn much more broadly and satisfy my curiosity in all kinds of ways that I never could before AI.

So if you combine that with—and I would say mine is even more—I was just at an AI event, Recursive, last weekend, and a question for people in the audience was, “If your team had to replace you plus AI as it exists today, how many of you, unaided by AI, would they have to hire to get the same output?”

The median answer in that room was basically 2. In other words, people thought that they were twice as productive thanks to AI as they would be if they were unaided. And that's pretty much where I put myself as well.

Even leaving that aside, okay, so people are reporting 13 hours gained. There's another stat in the report that says—let me make sure I get it exactly right—people are spending 6.4 hours, basically half of that time savings, on this bot-sitting activity: reviewing outputs and connecting things between different products that don't connect.

I think anybody listening to this show has certainly had that experience: “Okay, I got a Claude prompt. I got a Claude report or plan or whatever, but now I've got to copy and paste it.” So, the phrase you just used—the human becomes the integration layer. And I felt that, and it is certainly tedious.

Although it's honestly also just a part of general computer work, right? We've all got Slack, then this other task tracker, and then there's email, and so it's all a little bit disjointed anyway. All that to say, I don't really get it. Why is it so bad to be responsible for babysitting the bots, or bot-sitting? What is it that's really bothering people so much, or alienating them so much, about that?

So it's the taking away from the meaningful work. In the report, we look at 3 categories of interaction with AI. One is bot-sitting; one is using the technology—using the technology to do real work. You prompt it and it gives you the answer, or it asks you a follow-up question, in a way that you're moving work forward and iterating with the technology, as opposed to asking it a prompt, finding that it doesn't have the context, and then reprompting it.

We're seeing about 36% of all AI sessions fail. Meaning, a worker goes to use the technology and it's not successful. They either have to start completely from scratch or do significant rework. Imagine if it was right the first time.

Imagine if you could put those 6.4 hours into either using the technology to drive work forward or the third category, which is learning or building agents. The time savings would be significantly higher.

And so I think, again, there's a small component of bot-sitting that I think is healthy. For people who are curious, it's less of a problem. I love to believe in human curiosity. I think the reality for many—and I think Nathan, you and I are probably an outlier here—is that employees are too exhausted to be curious right now.

They're too overwhelmed with work to spend those 1 to 2 hours tinkering with the technology, prompting for different tools, and picking the right answer. And that's the problem: the fact that it's not meaningful work and it's not meaningful learning with the technology.

We also look at, along the different dimensions of bot-sitting, what is most exhausting. In the report, we call it the exhaustion multiplier. And what we see is that the highest exhaustion multiplier is associated with feeding AI context, right? Because that is, in the best case, something your AI should know.

It should know where the documents are and which documents are authoritative, and you should not be supplying that as a human in most cases. The other one is debugging. You see an output, you know it's wrong, but because of the nature of LLMs, you're not quite sure why it's broken or what's wrong.

You try to tweak one thing, but because the nature of the technology is probabilistic, not deterministic, you're not really sure which tweak worked and which didn't. That is the biggest contributor to this exhaustion multiplier.

Yeah, that's interesting. I do have this experience sometimes. One thing I've been really enjoying doing lately is creating songs for each episode of the podcast, so you can start thinking about if you want to request a genre.

I try, but I can't always promise to be able to make something great in any given genre. It's really striking how sometimes I'll run my kind of produce-an-episode Claude Code skill, which includes coming up with an idea for a song, writing lyrics, and prompting Suno with that. Sometimes I show up and it's a banger immediately, and then other times I find myself sitting there, and it's this kind of black-box thing where you're like, “I tweak the style prompt, and I tweak the lyrics a little bit, go again,” and for some reason it's just not working. It's just not landing. It's just not giving me what I want.

That can definitely be an exhausting thing, especially when I get into this spot where I'm like, “Who even cares about these songs? Am I doing this for anyone?” Although, actually, I do get a remarkably large amount of positive commentary on the songs. Anyway, I can relate to that sort of exhaustion point.

I can definitely also relate to the shoveling-context point. Before my now much more integrated setup, I used to have a single PDF with a bunch of intro essays that I'd previously done for the podcast, and I found it was kind of exhausting—although this is such a baby, first-world thing to say—even just to go find that PDF every time and put it into the web UI so Claude would have it to use as examples. It's like, man, that's really not much to complain about, and yet somehow it feels so much better now that moving files and context around has been mostly automated away.

Just on the time, though, I'm empathizing with some of these problems, some of these pain points, for sure. But it still seems like there's something—I guess one of the hypotheses we should always keep in mind is that people maybe just, in many cases, don't like their jobs that much. I think this is something that the AI discourse broadly should remember much more than it does. I've probably gotten on my soapbox enough times about that already, but that's for sure an ingredient in this overall recipe.

It was just striking that, okay, 13 hours saved, a little under half of that sort of re-consumed by doing this copying-and-pasting and double-checking work, but that still gives you almost a full workday back, right? Am I reading that right? Are people—have we created a 4-day workweek that we're just not ready to talk about?

So here's the problem, Nathan. When we look at the individual level, we're seeing all of this exciting productivity gain. Even if we consider bot-sitting, the net-net is positive. The problem is: Where is that time savings going? That's one aspect. What we do see, especially in organizations that haven't communicated an AI strategy and where employees aren't confident in their organization's AI strategy, is that employees are taking it for themselves. The amount of nondisclosure—the hiding of AI usage from managers—is rampant, and we have some data points in the report.

They're managing the perception of how they're using the technology because, in many cases, if you tell your manager or your organization you're getting 6 hours back, you're probably going to get 6 hours more of work. That is the wrong calculus. But I actually think the bigger problem is that what happens at the individual level often does not translate to the team and organization. In the research, it's sometimes called coordination neglect.

The most concrete example of this is that you can have me, as an individual worker, take a single bullet point and convert it into a 5-page report, then send it to my colleague. The colleague takes that 5-page report, decides it's too long, and turns it back into a single bullet point. Each person looks productive; each person looks like they've saved massive amounts of time. But when you put the pieces together, it's just this hamster wheel of AI slop, workslop, and task-shedding in a way that does very little service to the organization.

I think that's the disconnect here. It's very hard to measure, but it's certainly something we're seeing in practice, in conversations with executives, and in some of the data—certainly in terms of individual productivity gains not translating to real team and organizational gains.

Part of me feels like that might be healthy. It could be more healthy, perhaps, but a 4-day workweek would be a nice stepping stone, perhaps, on the road to the AI future.

I don't want to present the perception that I think we should not be giving time savings back to the employee—to be innovative, to be creative, and to have better work-life balance. The problem is that most organizations, in part because of the context problem, don't understand where the time savings is going or how to extract meaning from it at the team and organizational level.

That's the problem, because we start to see—that's feeding a lot of this “more is more” pressure. You don't know how X translates to Y, with Y being business outcomes and KPIs. The knee-jerk reaction is that the only thing I can associate as more tokens and more clicks of the tool is more output, as a proxy for productivity. That becomes very dangerous.

Do we see harm? It's one thing to say people are saving 13 hours a week. They have to give half of those back to these somewhat annoying tasks. Maybe they save a bunch of time, get more leisure, or do more social media at work or whatever, at least as long as nobody's asking too many questions.

Only 13% say their organization is doing much better, but is there a corresponding statistic for whether some people say their organization is doing much worse? The optimistic read of the results so far would be, as people often say, “It's still the worst it'll ever be.” This survey was done at a point in time that was coinciding with, by many anecdotal accounts at least, another step-change advance in models' ability to search for and assemble their own context on the fly—at least if given programmatic access to do it.

If you were to say, “Hey, only 13% are doing notably better, but we're not really seeing anybody doing terribly worse,” then I'd say, “Hey, that's a pretty good start.” But you might say, actually, no—there are a bunch of people saying their organization is outright suffering as a result of this. What do we see on that side of the ledger?

It's a great question. For this particular question—and again, the time savings range from 10 hours to 13 hours—the average is 11 hours per week. But in terms of the 13%, it's a Likert scale. We essentially ask whether people strongly agree or disagree, so we definitely see all aspects of the spectrum.

The optimism in me says you're right in the sense that we shouldn't expect transformative gains from this technology too quickly. What gives me pause, and what does concern me, is the volume of not just the bot-sitting but also the bot-shitting: the shadow AI, the nondisclosure, and the lack of transparency to the organization. That's not going to get better if all else remains similar. It's only going to get worse.

But I think we shouldn't rush to transform our organizations so quickly. That's also a really dangerous place to be. In the most effective organizations, they're measuring, and as part of that measurement, they're baking in failure. In the best cases, we're seeing executives pinpoint 80% of AI initiatives failing, because failing is part of innovation.

I think we need to keep in mind, too, that a lack of transformation in certain areas is very healthy, because it indicates that, in the best cases, you're taking well-intentioned risks and hopefully learning from those failures as well.

One thing I wanted to follow up on was your comment on meaning. I saw what I would call a striking apparent contradiction in the report: the observation that the people who feel most threatened by AI seem to be most eager to adopt it and use it more and more. They're going so far as to automate work that they'd rather keep.

I'd love to hear a little bit more, because it's actually very connected to this question of meaning versus alienation. Could you give us some examples of things that you heard from individuals about how you end up in a spot where this is the part of my job that I actually like, but I'm feeling pressured to have AI do it? What does that actually look like? I'd love to get a couple of sketches, if you could.

It's such an insightful question, and we unpack several of these paradoxes or contradictions because there are many of them at play, in part because this is such a psychological technology. The fact that we're treating it as a human adds a whole bunch of different complexities.

One of these paradoxes, one of these contradictions, is: Why do we see the people who are most fearful of the technology, most worried about being replaced or displaced, leaning in more? I think a lot of it is the perception. You feel a threat, you don't fully understand the technology, perhaps you don't have the support of your organization, and you want to look AI-native. You want to look like you're transforming.

The natural, immediate reaction in too many cases is to automate as much as possible. Unfortunately, in many cases, if you're looking to automate and visibly show that you're transforming with the technology, you're probably going to point it at parts of work that you're most familiar with. In not all cases, but in many cases, the parts of work that you're most familiar with are probably the parts of work that give you the most meaning.

Not always, but certainly. Concretely, what we're seeing is around relationships with other people—customer service, for example. These amazing customer service representatives have spent years, decades, developing the craft of the personal relationship and the long-term relationship. All of a sudden, you have a technology that can, in theory, automate some of that relationship building, perhaps to get it completely off your plate so that you're no longer interacting with a human.

Well, that's what gives you joy and meaning at work. That is very dangerous, and that is what we're seeing—not just in this survey. There was a fascinating study that came out of Stanford a while back that found 41% of Y Combinator AI startups are automating things that people would prefer to keep human. Again, this boils down to the psychology of this. We can't just assume that because the technology can do something, because it can automate something, it should be automated.

We know that so much of work and so much of the process is meant to be messy. It's meant to be full of friction, because that friction is sometimes called the IKEA effect, right? When we build something ourselves, when we do the hard work of doing the thing, that builds ownership, good judgment, purpose, and pride. These are not feel-good, nice-to-haves; these are hard drivers of performance.

It's a very difficult calculus because it differs for every person, every team, and every organization. But it's absolutely essential that we think about what the division of labor between humans and AI should be. The calculus should not just be, "Can the AI do the thing?" It also needs to take into account this human piece: What does the employee find meaningful? That meaning is going to drive their best work, and it's going to drive situations where, when they do get the time savings, they're reinvesting it into the betterment of themselves, their teams, and the organizations, rather than more clicks of the tool or taking the time savings for themselves and not sharing that with the organization.

That customer service example is a really good one, and I do think it goes to show how tough this is going to be in a lot of ways. I can totally imagine being a person who is in that kind of job because you like talking to people, you're a people person, and that is what gets you going every day. Then to think, "Okay, you're not going to do that anymore, but instead you're going to get to sit in front of this agent-builder UI or whatever and try to string together what you used to do, watch out for its failures, and it's like, 'I didn't sign up for this. I never would have wanted this job in the first place if I had to do this, but here I am.'"

At the same time, if you put yourself in the leadership standpoint—or honestly, in the customer's position as well—I think the logic of it is pretty unavoidable, if only for responsiveness. I've looked at, for example, my company's response times when Finn is active on our Intercom versus when it's a human. We do a great job on customer service, and we do have real people. Our customers have always spoken very highly of our customer success team. Yet the immediate response of Fin is, in many cases, a real value driver for the customer, too, because they get out of there in a couple of minutes instead of the longer back-and-forth of a human, which could be 30 minutes or more doing it the old way.

I think that is really tough. I've seen it said recently that companies are graphs of algorithms, and I think that did not come from Glean, but it very much rhymes with some of the recent releases around the enterprise graph. How do you think people should be thinking about this from a leadership, executive, and competitive-dynamic standpoint? It's hard to say. Maybe we want to keep some humans in customer service because we have super-high-value customers, and they're going to value it. There's something intangible. But I think it's hard for most companies to really make an argument that we shouldn't take a 90% cost savings and the ability to be instantly responsive to all of our customers because people like doing it the old way.

But that's one algorithm in the graph of algorithms that constitute a company. There may be other ones where you do have a better reason to keep it more human. How do you think leadership should be composing their organizations and thinking, for all the different parts of them, about what they have to accept as the tides of history and where they maybe want to hold on to things for special reasons?

This is where I can really geek out because I think it's so exciting. When we think about the enterprise graph in particular, the enterprise graph isn't just a connection of people, tasks, and documents, right? It's a collection of everything from the mission of the company to the goals, the projects, the tasks, the people, the documents, and the technology.

You can easily imagine a world where, if you have an AI platform that understands enough about your organization—and we're starting to see it, certainly with our customers at Glean—we can do two things that are really exciting. One, it can understand, given all the interactions you've had with customers, your sales team has had, or anyone else, what the level of complexity in those interactions is. What did the customer or client want in those interactions? Did they want that fast response, or was it a long-term relationship-building conversation?

It can then make a determination or recommendation to you in terms of, "Okay, this is a predominantly human-warranted interaction. This is a human-in-the-loop type of interaction. This should be completely automated." We should be making the trade-off ourselves. I absolutely don't want to live in organizations that over-index on keeping work meaningful for employees. In the best case, you sometimes automate parts of work that employees do find meaningful, but you absolutely replace that with parts of work that employees can now take on that they find just as meaningful, if not more meaningful, in the best case. So that's one dimension of this.

The second dimension is that, if you have an understanding of not just the enterprise graph but also how individuals work—their skill sets and their career ambitions—you can start to make recommendations using AI in terms of, "Okay, what is the task allocation now for this specific person, given their expertise, given the goals of the organization, given the development of the technology, and given their career ambitions?" It becomes incredibly exciting.

We're seeing this in terms of how some of our customers will now staff project teams in fundamentally different ways. Previously, we've relied on the static org chart to staff our projects. If AI knows enough about your organization, it can make a dynamic recommendation in real time: Who across your 1,000-, 2,000-, or 10,000-employee base is the right mix of people based on skill sets, expertise, career ambitions and passions, and bandwidth? It can do a complex calculus that we as humans could never do.

That is, in my opinion, the North Star of enterprise AI. It's the power of the enterprise graph, and we're starting to see snippets of this come to life in a way that I think is incredibly exciting.

I like that notion. It's been a long time since I've worked at a big company, but when I did, I had a lot of outside-the-box ideas about how to use things like internal markets or auction mechanisms to figure out how to allocate people to the best and highest uses. Those mostly fell on deaf ears, and for understandable reasons, because there was going to be a lot of time spent operating those mechanisms. Even if you assume buy-in, which we didn't have, I could see why it just wouldn't happen. The younger me didn't understand those things quite as well.

But because you can have the AIs grind through this, it's, in a sense, the positive version of the mass-surveillance use case, right? We used to be saved from mass surveillance because there just wasn't enough human brainpower to process all the logs. Now we've got that problem solved in a potentially very problematic way. But here, you can actually imagine going through your full roster and really trying to tinker with all the different assignments and configurations that you might spin up.

You can imagine how that could really unlock a lot of potential that nobody would have had the time, and certainly probably the perspective, to be able to pursue in a pre-AI era. So I do think that is pretty exciting.

There's a small finding in the report that I'm incredibly excited about. We have one of our co-authors, Aruna from Berkeley, and she made a note as soon as we put this in the report: "Wow, this is so exciting." I think so, too. We don't foreground it because there's so much, but what we see in the most effective organizations—the 13% who have employees saying that significant productivity gains are occurring—is that they measure.

They measure a lot of different things. In particular, they don't just measure productivity, but they also put that data disproportionately more in the hands of employees. This is not new. I saw it with collaboration technology, too. I saw it with remote and hybrid work situations: If employees have access to that data, if it's transparent to them, everyone is better off.

Unfortunately, we're seeing cases of that not happening in organizations. Again, another benefit of the enterprise graph is that everyone has access to it, right? Everyone is able to query the graph and understand the state of play within the organization in a way that I think is going to drive much better outcomes for both the organization and the individual.

So, speaking of measurement, one thing that I was pondering myself as I was imagining myself leading an organization through these challenges, based on all the findings in the report—I don't know if we said this yet, but 69% of people admit to some form or some amount of bot-shitting. That has a couple of different definitions, but for me, it's passing off AI work somewhere down the line where you yourself cannot defend the quality of the work. That's a very high rate.

It does strike me that, again, kind of like my earlier thought, in a way, that's maybe a really optimistic thing for how well AI is working. If you already have two-thirds of people just doing AI outputs blindly and sending them down the line, and the wheels aren't falling off entirely, that's kind of an amazing finding. But then, if I'm trying to manage that as a leader, I'm thinking, how do I detect who's doing it? How do I detect where it's actually working?

Maybe I do want to share that with employees. Maybe I do want to partially share it. I'm not sure exactly what the right level of transparency would be there. But have you seen anybody doing that? I've also seen, historically, my attitude, or my synthesis of available information, has been that AI detectors don't work.

When I go to present to a group of teachers, I'll say, “Don't do AI detectors.” Or, if you do, you certainly can't trust them too much; they can be wrong. These days, it does seem like Pangram Labs is getting a lot of praise in the general discourse around being pretty reliable.

So I'm wondering, should I be, if I'm a leader, adopting something like Pangram Labs and having all these intermediate work outputs evaluated, such that I can potentially both realize who's doing this and maybe where? Again, this is kind of an angle on nodes in my graph of activity that constitute my enterprise. Which ones can I actually just have AI do? I potentially already have a lot of answers to that question in the work product that people have—the AI work product that people have—passed off as their own. Do you see anything like that?

Yes, and I think it's so early. There's not going to be a right answer for every organization. A lot of it depends on the psychological safety within the organization. I've long followed Amy Edmondson's great work, and I think psychological safety has never been more important in our organizations because, ideally, you have people raising their hands and saying, “This is bot-shitting,” or, “I contributed to bot-shifting, and here is why.”

I think AI detectors—there is, for sure, and I'm seeing it from multiple different angles—a world where all of our AI tools will flag the level of uncertainty associated with the response, as well as the likelihood of purely AI generation versus human-AI generation. I think what's exciting about having more context as well is that you can feed it more data points. You can feed it not just, generically, “Is this likely to be generated by AI?” If it knows enough about you as an individual, it can know, “Okay, Rebecca's default writing style as a human is this,” versus doing something generic. I think that's why we're seeing a lot of detectors not work.

This isn't necessarily, and certainly isn't strictly, a technology issue or solution. Organizations need to understand the why. The why behind the bot-shifting is just as important as the bot-shifting itself. In particular, when we think about using shadow AI tools, right? That is a form of bot-shitting. You're injecting risk into the organization by using an unapproved tool or a sanctioned tool in an unapproved way.

We're seeing a cohort of organizations crack down and punish employees for doing that. We're seeing a portion that definitely have guardrails and repercussions for unsanctioned use of the technology, but aren't strictly doing that. They're understanding why employees are using unapproved tools or why they're coloring outside the lines, because usually it's our highest performers that are doing that. Usually, they're doing that because the existing tech stack isn't working for them.

They see so much potential in the technology, and they're making that calculus in their head: “Hey, I would rather get the productivity gains and, in the best cases, hopefully get the gains for the organization.” They're making this trade-off. Ideally, organizations recognize that and make the safe path the more efficient path. That is the gold standard. I think part of it is a technology piece, but part of it is deeply human.

Another stat that jumped out at me was that doing more bot-sitting is associated with being more successful with AI—both probably because you need to do it to get good results and because you're naturally going to do more of it the more you use AI, in the simplest analysis. But then also, more bot-sitting seems to be associated with a higher likelihood of being on the job market and actively looking for your next phase of your career. The same is also true on the bot-shitting side: people who are doing more of that are again more likely to be looking for a job.

That sounds scary, and it might be scary. I guess another way I might interpret it, if I think about the person who was in the traditional customer service or customer success role and is now being asked to babysit bots, is that maybe some of this turnover could, in fact, be healthy. If people are doing something they don't want to do, maybe that's part of the story.

A big question in general with AI is, obviously, is it going to create more jobs than it destroys? But even more locally, focused on specific organizations, teams, and individuals, is the new job that gets created—whether it's greater or less than one per job that gets eliminated—something that the person who had the original job can pivot into or would want to pivot into?

My read on some of this stuff is that it suggests that, in a lot of cases, the answer is no. People who are doing a lot of this AI stuff, good and bad, sound like they're kind of voting with their feet that this is not really what they want. Maybe they don't see themselves being successful in this new world. So I wonder, do you see that as a sign that leadership is doing something wrong, or do you see it as an actual signal that change is often not what people want, and so they're reacting in what might ultimately be a perfectly sensible way?

So I'll hypothesize here based on the data, because we certainly don't know causation. These are correlations more than anything else. When we think about bot-sitting, it's very different from bot-shitting. What we're seeing in the data, and what I'm certainly seeing in practice, is that there are 2 big links between the act of bot-sitting and the desire to leave the organization.

One is—and we have an amazing co-author on the report, Polly Annardi, who's done foundational impactful work on what he calls digital exhaustion—the digital employee experience is increasingly the employee experience. We often trivialize it, but the reality is that, overwhelmingly, we're seeing that if employees are exhausted by technology, they're wanting to leave the organization.

When we think about bot-sitting in particular, what I'm seeing is that if employees are spending all of this time manually feeding context to the AI, that doesn't instill a whole bunch of confidence in the employee that their employer, the organization, has a strategy—or a good one—around the technology. I think that's a big part of it. If you're wasting your days bot-sitting when your organization is vocalizing in all-hands meetings and town halls that this is a transformative technology and you're transforming, there's a disconnect.

If you're seeing, all else being equal, another organization that is investing in technologies that do have context, you're going to choose the latter. So I think that's the bot-sitting piece. It's certainly not the whole piece, but I think those are reasonable conclusions to hypothesize.

The bot-shitting, I think, is different. The bot-shitting is, in my opinion, and based on what I'm seeing in conversations, a sign that you've decided you're disengaged from the organization, perhaps because you're spending so much time bot-sitting. Your wanting to leave the organization is a reflection of that broader disengagement, right?

If you're bot-shitting and no longer feeling a sense of ownership over your AI-generated work, probably that's part of a bigger picture of disengagement and sentiment within the organization. I certainly don't think that's the case for every employee, but I do think that's a part of it as well.

Aruna, who I mentioned before, hypothesizes in the report that perhaps people have bot-sat so much, or they've become such experts in the technology, that they've realized their market value is higher outside the organization than inside the organization. That's a reflection of the disengagement, both in terms of bot-shitting and the desire to leave the organization.

Your market value has increased. I certainly think that's the case for a subset of employees as well.

Yeah, so basically, the key is that if you're leading an organization, you've got to tell who the good AI users are and who the bot-shitters are, and you're going to want to take steps to retain your top performers. You need to both use your AI detector to know who's using AI, but then also need a quality score to know who's actually using it effectively. And those people who are using it a lot but using it effectively are the ones that you're at risk of losing if you don't level up your game in one way or another, which could mean better AI tools for them or perhaps better compensation. The ones on the other end who are just passing off AI work that they can't defend, probably in the end you're going to have to make your peace with the fact that parting ways with these people is going to be the cost of transformation.

How do you see people telling the difference, and do you see that upside? Obviously, I'm deep in the AI bubble, though I'm not as central in the AI bubble, but we've got all these stories of basically sports-star and pop-star incomes for top-end machine-learning researchers. I don't expect that's happening in quite as extreme a way in enterprises, but are we seeing retention plays where people are being meaningfully rewarded in terms of compensation if they're really on the cutting edge of helping their organization with AI transformation, if only in terms of embodying it and bringing it to their own role in an elite way?

I love this question. No, not in a meaningful way. Should they? Yes. And this is not new, right? My colleague Rob Cross, who has been an inspiration for me for many years, has research showing that high-performing organizations are up to 5.5 times more likely than less-effective ones to measure and reward effective collaboration. The great organizations do this. They build some sort of measurement. It looks different for every organization. Sometimes it's tied to pay, sometimes it's more informal, but fundamentally, the most successful organizations, the most enduring organizations, do something to reward the collective and not just the individual.

I think that could not be more important with AI when we're starting to see this tragedy-of-the-commons hyperfocus on individual productivity in a way that's not translating. I'm seeing it in bits and spurts. As one example, they take something like a hackathon or an agentathon, and they think very carefully about what they're going to incentivize. They're not just giving prizes for the biggest business impact, as most organizations do; they also give prizes for improvement, the best before-and-after prompt, co-creation, and peer feedback. It's not just, are you using the technology to drive your individual productivity gains, but are you creating value together? Those types of things we're seeing in effective organizations, but we're not seeing them at scale in a way that I think we should.

And again, if I think 10 steps ahead, if you do have a context graph, if you do have an enterprise graph, it can help you make that determination because it will know where value has been created, and it can start to give you recommendations. We see it in the data where a large portion of employees are already saying their organizations are using AI to inform performance management, hiring, and firing, too. When we ask, “What is your expectation?” not surprisingly, it's even higher.

I think where—I hope—this is the hopeful part of me, I hope we see a world where not only can AI help us drive more objective, fair performance reviews, when they've been riddled with biases forever, but it can also help us feed the performance evaluation with a bigger focus on the collective as opposed to the individual. That becomes incredibly exciting, and I think we're going to see it because we're already seeing bits and spurts of it.

So, tell me a little bit more about how you think leaders should approach this culture-building. I've said that many times myself: it's got to be a cultural thing. You mentioned that we need to see executives using AI; that's a very sensible starting point, but it's got to go farther than that. In some cases, it's going to be in a pretty tough messaging environment, let's say, right? Because we do have quite a few rounds of layoffs from tech companies, mostly so far being attributed to AI.

Tell me if you think this will play out differently, but I suspect that we're going to see a pretty similar phenomenon extended to the rest of the economy in probably a few successive waves over a not-too-long time horizon. Certainly, my sense is that public-company CEOs feel like, “Geez, whether I like it or not, I'm under a lot of pressure to figure out how to make this stuff work, and I'm told that my competitors are going to be dramatically cutting their costs. Can I really compete if I don't?”

So, they're presumably going to be in a pretty tough spot where they're like, “Okay, here's me as the CEO using some prompts.” At the same time, though, we're cutting head count by 10% to 20%, or whatever, and we're going to work better together. It seems like it's going to be a very tough messaging plan to land. What advice do you have for people who are facing this difficult challenge?

This is an area where we do know so much about what works. Unfortunately, we don't see it implemented, and employees can call bullshit from a mile away. If they see a talk track that's different from what's happening in practice, that does not instill confidence, and you start to see a whole bunch of this symbolic use of the technology.

Transparency is so important right now for so many reasons, but in particular to understand the why. The why behind organizational changes, in either direction, helps people understand whether they're willing to believe in it and willing to commit to it in a reasonable way.

Now, what I'm seeing overwhelmingly is performative theater at every level of the organization, but in particular with executives. I spend a lot of time with executives, and the number of conversations I've been part of where one executive decides that they're going to cut 15% of head count, and then there's a conversation around, “Should it be 16%? Should it be 14%?” There's this very disturbing narrative that there's a one-size-fits-all approach for every organization. There is not, and we've seen efforts to change organizations, dismantle hierarchy, and flatten the org chart time and time again. Often, they don't work because they're perceived as short-term fixes that don't help the underlying problem.

This is an organizational-design transformation in its fullest sense. Having an understanding of what your culture is, what sorts of behaviors your culture embodies, what it rewards, and what it punishes—that is very important right now. Articulating that to employees in the context of AI is important as well.

What AI is in relation to humans is very important, and what we see overwhelmingly is that in these transformative organizations, employees are significantly more likely to view AI as a teammate in how they interact with it, but not as an employee on the same level as them. That becomes dangerous when we're starting to deflect blame onto the AI tool. The blame, the onus, and the responsibility rest with humans, but ideally the organization is positioning AI, both psychologically and in the tools and technologies they invest in, as a teammate in a way that employees can grok: How might I interact with this technology? How might I start to delegate work to it?

It's a multifaceted problem, but the answer is not to make cuts or sweeping changes without understanding the impact of those changes at a smaller level. I think that's why, coinciding with these layoffs and cuts, there's a lot of regret because we don't fully understand what employees do. We've seen this in customer-service cases where you cut the human agents and realize, “Wow, those human agents did really important work in building long-term relationships with customers.” You then bring them back.

Again, this is the power of an enterprise-type graph: you can do that proactively, and you can understand the real value of the humans in a way that is much more thoughtful and much more long-term thinking than just doing the thing without understanding the ramifications. It's not easy, but the answer is not to move quickly without understanding the foundations.

It strikes me that the importance of mission and mission buy-in might really be at a premium in the near term, because I can just imagine a hugely different reaction if you imagine a scene from The Office where it's a generic widget company, a paper company, indistinguishable from tons of other competitors, and you ask, “If we don't do it, who will?” The answer is, “Well, lots of other people, and they'll do it pretty much just as well, and the world won't really be that much different.”

That seems like a really hard place to manage this sort of transformation from. Whereas, I do know some organizations where they're really, in a very meaningful way, bought into the mission and the success of the organization. When things like this happen, even if it does result in somewhat painful change, if people believe that the mission itself is going to be advanced by this change, then presumably their tolerance for certain pain is dramatically higher.

I guess the tough thing is that it's hard to synthesize a mission where there really isn't a very compelling one, but any thoughts on the mission premium, we might call it?

So, it's super insightful. One of our other contributors to the report and one of our founding members, Bob Sutton, and I were working on a piece right now.

Hopefully, it’ll be in Harvard Business Review: If you are going to flatten, how do you do it well? And what are the pieces you need in place to do it well? This is not necessarily job cuts; it’s strictly looking at flattening and hierarchy. One of the core arguments, backed by research and evidence, is that you can’t flatten well if employees don’t understand the mission, because what you’re starting to do with job cuts, but also just flattening in general, is dismantling hierarchy.

And hierarchy gives employees a sense of what to do in situations of uncertainty. If you don’t have that hierarchy, you need to replace it with something else. And the mission is a key way to do that, right? If employees have a strong sense of what the company’s mission is, and they believe in it and invest in it, they’re going to make the right decisions when no one’s watching or surveilling them. If you don’t have that, even if you have a super-strong company mission, but employees can’t make the connection between their work and the company mission, that’s just as detrimental.

Right? You can’t do any of this well, and I think we underestimate—and I love that you called this out—I think it’s incredibly insightful and important that we need to think about the bigger picture. When we think about the DNA of an organization, the mission is at the top. And again, the enterprise graph, if it knows enough about that, can start to give you proactive insights in terms of, okay, does the work we’re doing show that employees understand the mission? Does the work they do at an individual level ladder up to the mission, or is it completely disconnected from the rest of the enterprise graph?

That again becomes very exciting. But I think we underestimate just how important purpose and pride and meaning at work are. If nothing else, without that, it drives certainly more buck-shifting within the organization.

How much of an impact do you think Elon’s Twitter takeover and subsequent managerial decisions have had on, let’s say, executive culture broadly? I guess my read on that is, it sure looks like it basically worked. I was asking AIs about this in preparation. It’s something like 7,500 employees was what they had when he took over. He took that down to something like 1,000—maybe as an absolute low, maybe 1,500. He did hire some people back and has hired some people. It seems like the team has grown a little bit since then, but it’s maybe a third of what it used to be.

They did lose a lot of revenue. I would say that was mostly because advertisers didn’t like him, his content policies, or just the general vibe, and walked away. But on all the other metrics, they’ve kept their users. The site—you remember, I’m old enough to remember when people said the site wasn’t going to work anymore—continues to work. If anything, I think recently they’ve started to accelerate some things with a really nice new API that I’m building on. I’m like, “Oh, this thing is actually really well done.”

Elon’s obviously a special case in many ways that has options seemingly available to him, financing and otherwise, that not everybody has. But do you think that example entices executive thinking broadly throughout the country to think, “Geez, could I do something similar?” Or is it just so far out that it doesn’t even really register with most company leaders?

I’ll comment generally on this because I think we’re seeing multiple different instances of CEOs making pretty radical changes and executives jumping on the bandwagon and wanting the promised gains for their organization. That is natural. Absolutely. We see it with any high-profile CEO. We’re certainly seeing it with Jensen at NVIDIA: very interesting leadership strategies. I admire him deeply in so many ways.

What works for Jensen and NVIDIA doesn’t work for 99.9% of other organizations. Part of it is “mission as boss,” right? They do a lot of work on meetings. The fact that Jensen doesn’t have one-on-one meetings with his direct reports—that does not work in most organizations. It works at NVIDIA because they have this mission as boss, right? Every employee—at least Jensen’s direct reports, as far as I can gather—has a super, super clear sense of what the mission of the company is, and they’re fully bought in. So they don’t need as much of that one-on-one interaction with the boss.

I think this holds for AI as well. What works for one organization, especially when we think about AI that is, again, so deeply psychological, isn’t going to work for your organization. I think we can take inspiration, and certainly I think we underestimate the power of a leader who motivates and a leader who makes bold changes during this time in a way that employees feel a sense of energy around them. But I do think it’s very dangerous to think that what works, especially in the headlines and the external perceptions, for one organization can be translated or copy-pasted into your organization.

I think I also wanted to double back and get a little more color commentary. You said that in the most successful organizations, you’re seeing this pattern of people relating to the AIs as teammates, but not necessarily peers. I want to hear a little bit more about that, and maybe you could fill in some of the details with your experience of using Glean at Glean today.

You obviously have human teammates, and you have AIs. Presumably, these AIs have as good an integration into the deep context of Glean as any AIs anywhere could possibly have. So what is your mental model for when you go to an AI, when you go to a person, and when you start a group chat between you, another person, and the AI?

What’s the difference, also, between the sort of you, the Rebecca extension AI, versus the—I don’t know, is it a central Glean bot that’s a company-wide version? I guess paint a picture of your present being everybody else’s future, right? Tell us what’s going on at Glean now so people can have a little more concrete sense of what they’re going to be walking into.

All right, and this is a nuanced argument for sure. I think I’ve done enough research on this sort of mental model. In a previous life, I did some research with colleagues and Carol Dweck at Stanford, in her lab, on mindsets and the importance of this psychological framing.

What we’re seeing in practice—and I’m not convinced this won’t change—is that when we think about the mental model of a teammate, that is extremely valuable right now because it gives employees something concrete to explain: How do I think about this thing in my day-to-day work? In particular, we see that when employees adopt that teammate mentality, they’re not treating the technology as transactional, right?

It’s not like a hammer or a calculator, something you can pick up, use, and put down, right? Treating it as a teammate means you’re not expecting the perfect answer. You’re not expecting the technology to work perfectly. The teammate is in the interaction. If it knows enough about me, the Glean assistant is my personal assistant, my teammate that I go to every hour of the day for things I need help with and to move work forward.

The danger of the teammate mentality in practice, when we think about the psychological aspect, is deflecting blame. This isn’t a human teammate where, if the AI makes a mistake, we can blame the human. And Paul Leonardi, who I mentioned, has done fascinating research to show that when a human screws up—if I have an assistant and the assistant makes a mistake, a human assistant—we blame the human assistant.

When an AI makes a mistake, we blame the human as well. If we’re on the receiving end, unfortunately, as we’re deploying these tools, so often we think, “Oh, this AI is agentic. It has agentic capabilities. We can deflect blame.” No, that’s not the case.

But in terms of Glean, the assistant knows enough about me. I also use it all the time when I’m wanting to do a lot of executive-type communication—upward reporting, for example. Every executive likes to consume information in a different way, and so being able to ask my Glean assistant, “Okay, what is the preferred mode of consumption for this executive versus this executive?” is very helpful.

As I was joining the organization, I'm about 11 years 11 months in, feels like it's been 11 years some days—I’m able to query the AI and understand the institutional context that came before me. I wrote an article on this. I’m able to understand not only what decision was made around this, but why we launched this specific feature, who our customers are, and what our product roadmap is. All of those things are super important, but I also asked it, “What makes a successful employee successful at Glean?”

It was able to understand from the culture: This is what the successful employees do differently. And I fundamentally acted a little bit differently because of that, knowing what the culture rewards. So that becomes really exciting.

A little bit out of—it’s too close to the secret sauce—but could you share a little bit more about what Glean AI said makes an employee successful at Glean?

Yeah, I wrote an article because I used my exact queries, but it was something around the team mentality and, I think, what we’ve talked about today. It’s not a culture of rewarding individual performance; it’s a culture of rewarding the team. It’s also a culture of long-term thinking, as opposed to other organizations, and so that innovative spirit—and I feel this every day—is rewarded when you raise your hand and share the weird, wacky idea, share the future of what’s possible with the enterprise graph.

Those types of things that do differ case by case are some of the things I asked it in my early days, and now it knows enough about how I work that it’s able to produce information that is unique to me.

I'm also able to choose different models depending on the task at hand, and it has memory capabilities that are constantly learning with every query and every day what my priorities are. It'll proactively flag when I haven't followed up on an email. I've become horrible at email after working at Asana for many years, so that's really helpful. Again, it proactively tells me when I've missed something or when there's an action item in a document that I've forgotten, helping to make sure I don't drop the ball on too many things.

So, on that note of long-term thinking, a tough question here for sure: If you try to see through the fog and imagine what Glean looks like in a few years, and more broadly, what large companies in general look like a few years from now, what do you see?

I think we have the easiest time imagining a version of the world that's pretty similar to the current one, but where a lot of stuff is automated and the savings get passed on to the consumer. Maybe companies are both more profitable and their stuff is cheaper, and people are able to buy more because everything got so much more efficient. It does seem like, though, that idea of efficiency—both higher profits and higher consumer surplus—is premised on major labor-cost savings, meaning companies are going to be smaller in terms of head count than they are now.

There are a lot of people who want to tell me a different story about how it's not going to be like that, and it's really going to be about growth. I'm always like, well, sure, it's going to be growth for some, but I also see a trend in business that starts again with technology toward winner-take-all dynamics. It seems like we're seeing more and more market concentration in a lot of different places.

I do believe that some companies will really win through growth, but a lot of that is probably going to be taking share from other companies. Is there another vision that you find compelling aside from that default vision, or is that ultimately the future you think we're headed toward?

They're two separate questions. A vision I think is compelling is different from what I think the reality that will play out is. The answer everyone loves to give is that AI frees us up for higher-order work and we're going to have more jobs than ever before. I think we will. There are certainly going to be jobs created that have never existed before, and that's exciting.

I don't think costs are going to decrease in the short term at all when it comes to using the technology. You speak with executives every week whose AI budgets have been blown in a very short time for the year. Sometimes they're blown—we've seen headlines where someone blows their token budget for a year in a month. The costs are not decreasing, and I think that is a very real concern for every organization.

That's why I think the future is not a single-model platform. It's a multi-modal, multi-tool platform. In the best cases, the AI is able to proactively route the task to the right model given a multitude of different dimensions: efficiency, complexity, and cost as well. So, I'm pretty confident in that.

I think we are seeing smaller and smaller teams, and I think that makes sense because, if we're using AI in the right way, A, it should reduce the massive coordination tax of work. We should probably have slightly smaller teams, and we should probably need fewer managers if a core part of the managerial role has been to manage this coordination tax that can now be taken on with AI.

That becomes exciting in many ways because we're seeing an opportunity for specialists to become more generalist. We're certainly seeing roles rebundle. I've seen this in practice, in particular through my PhD: I saw the org chart pressure-tested and organizations start to think about where AI can give them insights across silos in a way that now makes sense to redraw roles—where work that previously lived in 2, 3, or 4 different roles now lives in a single role. I think we're going to see that in the companies that survive and thrive.

I hope for a world where we're spending more time on the deeply human parts of work and we have a much clearer sense of what the right division of labor is. I'm starting to see enough evidence that I think that's a possibility.

I spend a decent amount of time with an amazing CHRO at one of the largest healthcare organizations, and they've used AI to task-map all their roles and determine that 70% of all tasks are overlapping with other roles across the organization. In those sorts of situations, you can easily imagine the enterprise graph making the determination: “Okay, what makes sense to live in a human role versus what makes sense to live in an agent role? What's the right agent-to-human ratio?”

That is certainly going to be in flux, and I think organizations are going to grapple with that. Those are the things that I'm pretty confident in. Beyond that, I think a lot of it is up for grabs, and I think a lot of it depends on the level of intentionality that organizations put into the human aspect.

This is not a technology that's inherently good or bad. It's not going to inherently make our organizations better or worse. It's about how we enact it, and my optimism says that, yes, this is going to make for much better organizations in terms of delivering customer value and delivering employee value. But I think we're seeing enough evidence that, if you're not invested in the human piece, this can easily create a world that's worse for the individual, team, and organization. I hope that's not the case, and I think I'm seeing enough evidence that it isn't if we're intentional about it.

Yeah. There's probably going to be a healthy amount, or perhaps an unhealthy amount, of creative destruction in any case. I do imagine we'll see a lot of organizations that look great, but there's probably going to be a strong survivor bias in the organizations that continue to exist in however many years' time. That's obviously another cycle. All these cycles are getting shorter and shorter.

100%, and I think what I'm also seeing is that what works for an AI-native company does not work for a legacy company. I think there are going to be 2 models. Unfortunately, for many reasons, I think AI-native companies have a leg up. They don't have that existing baggage or the existing legacy org chart.

But don't try to copy and paste an AI-native business model onto a legacy organization. Figure out what has been working for so long, and figure out what needs to evolve versus what shouldn't evolve.

So, thank you for giving me so much of your time today. Just a couple of final questions. One is on your own work. If you imagine the next couple of cycles of this—whether it's the biggest companies on the stock market getting displaced or the model-release cycle—it seems like everything is speeding up.

In work like this, my gut says that you've probably already used AI pretty extensively in producing this report, but it feels like we might be hitting a phase-change moment for you, where the way in which you can study organizations is probably changing pretty dramatically.

I would imagine that you might be planning to use AI interviewers in the future, along with data analysis, and maybe this sort of thing becomes less of a one-off report and more of a rolling index, where there's a monthly update or even a real-time sort of vibe to it. Certainly, everybody is going to want to be as close to the moment as they can in terms of understanding what is really going on. How do you anticipate your own work changing over the next couple of cycles?

We're planning to have this be a pulse-type survey, which I've long been excited about, especially because we're looking at it from multiple different dimensions. Every 6 months or so, the goal is to be able to track this longitudinally. When we think about what the bot-sitting percentage will look like and how we see roles evolving, that becomes incredibly exciting and uniquely possible with AI in many ways, in particular because of the data analysis.

When we were writing this report, the co-authors and I were very passionate about human narrative and human storytelling. There are parts of the research process that are absolutely 100% human, but there are parts that we can now do much more quickly and effectively because of AI. I think that aspect is exciting.

As a researcher, I'm trained in a methodology called ethnography, which essentially means you embed yourself in an organization and watch over many months and years how things change. I actually don't think that's going to change a whole lot with AI. I certainly think field notes and some of the ins and outs of how you do the method will change, but these are long-term changes, and I think there are parts of this change that we will never be able to glean fully from survey data.

We're going to need to have the lived experience of the people on the ground. Interviews can get you some of the way there, but I think there's no substitute for embedding yourself in an organization, doing truly grounded research, and figuring out all of the unexpected ways that this is going to change our organization, because I think we as humans don't even see it a lot of the time.

So, that aspect, I don't think, is going to change much: the importance of really grounded research, where you're taking the time. This is a technology that you can't learn everything from subjective survey data or even the exciting telemetry data that gets you a piece of it. There's no substitute for that sort of multimethod research approach, in my view.

Is there any place—an organization, whatever—that you would love to embed in that would be your dream deep-research environment? Another version of that, I guess, would be: I don't know that there were any experiments. All of this was pretty observational, but an organization obviously has enough on its plate without letting academics like you come in and control experiments.

What do you think would be, if somebody really wanted to offer themselves up as tribute, the kind of big questions that are hardest to answer? If you were really able to get the right kind of access or structure an experiment in the right way, what would shed the most light on the biggest questions that are currently open for you?

Gosh, this is so tempting, and I'll give them a cheesy answer. One of the reasons I joined Glean was that we do have this bias for experimentation. While we don't have any experiments in the traditional sense in the report, we do have a whole bunch of experiments going on, and ones that I'm excited to talk about in the short term in terms of how we're pressure-testing AI alongside some of our practices.

Meetings is one area where we've done experiments around how Glean employees respond to AI now being brought into our meetings in a way that is deeply psychological and exciting. That's an exciting piece. I think the org-chart transformation is top of mind for me, and Bob Sutton, who I work very closely with.

I would love to embed in a traditional organization. Sometimes these are called extreme cases, where organizations will do something completely radical and observe on the ground how this is happening. I would love to go into a legacy organization that has decided it's going to fundamentally rewire the org chart and watch it happen on the ground.

I should also mention your book, “Your Best Meeting Ever,” right? That's what it's called—“Your Best Meeting Ever.”

The layer?

You just reminded me of it in that comment about how bots are being brought into meetings. What's new in meetings? How would you update your advice as we see cultural expectations and practices around meetings evolving?

Oh, gosh, this is an hour question. I think meetings are not unlike any other work practice. In many ways, they're the most dysfunctional work practice, but what we're seeing with every work practice is that AI is helping in some places and making things more dysfunctional in others.

I'm incredibly excited about the potential of AI injected into meetings to help us have a better sense of the health of a meeting. I'm seeing so many examples, working with organizations all across the world, of people using AI in fundamentally different ways: understanding whether the dynamics in the meeting are healthy, understanding whether executives are dominating the airtime, and automatically deleting meetings from calendars when they fail to have a good meeting design or an agenda, or when people don't accept them.

Even measuring collaboration—understanding whether people are in the right collaborative mode for creativity versus coordination. That's all something I could nerd out on forever. Where I'm seeing a lot of dysfunction is in cognitively offloading to AI and sending our digital twins to meetings instead of showing up ourselves.

I don't know if these stats made it into the end report. We certainly have a UK piece where we see this on another level, and there are some stats on the sheer number of people who have sent digital twins to lead meetings on their behalf and the sheer volume of meetings that now have at least one note-taker.

That's where I become very nervous and frustrated, because meetings are so expensive. They're one of these activities where the power and potential of them is that they are deeply human. If we think that we can send a note-taker bot to a meeting instead of showing up ourselves, that is a symptom of the fact that we didn't need the meeting, and that is something that maybe AI can fix. Maybe, if it has enough information, it can fix it.

But some of these things are ground-zero fundamentals. Having an understanding of what deserves to be a meeting in your organization is the first step. It's largely independent of AI in terms of fundamentally rethinking whether a meeting has the right purpose and ensuring that it truly warrants that live, synchronous time that is incredibly expensive.

The last point I'll make around this is what is often being referred to in the research right now as “mental proof” around the technology: the fact that we are watching how people use AI very carefully, and we're starting to be able to tell when people are bot-shitting—not always, but when people are bot-shitting versus not.

When you see someone—certainly a manager, but anyone—bot-shit, that triggers a sense that they don't value your time, right? The opposite is also true. If we see someone who has invested enough time into designing a meeting and understanding how to use AI in a disciplined way, that's a reflection of the person.

I think that's what we're seeing in meetings as well. You send your digital note-taker bot to a meeting, and it signals that you don't value other people's time in a way that is a deeply human phenomenon. I think that's another big piece of this whole picture, inside and outside of meetings as well.

Cool. Thank you. Time, again, is the scarcest resource. I appreciate how much you've been willing to share with me today. Rebecca Hinds, thank you for being part of The Cognitive Revolution.

Thank you for having me.

[music]

Woke up early, opened the tab. 50 [music] windows, half of them back. Fed it context, watched it spin. Ran the whole thing back again. [singing and music] 6 hours gone for noon, making sure the bots knew what to do. Work [music] looks finished, polished, clean, but I don't know what the hell it all means. I'm bot sitting, bot [music] sitting, babysitting the machine. I'm bot sitting, bot sitting, best you've never seen. [music] All this nonsense, hollow sheen. I'm bot sitting. Where's my curiosity? [singing and music] Saved 13 hours, gave 6 [music] right back. Copy the context, cover my tracks. Everybody's faster, nobody moves. [music] Hamster wheel on premium fuel. Boss asked where the hours went. Opened [music] five more tabs, paid the toggler in. Output's perfect, [music] nobody knows whose mind it came from, which way it goes. I'm bot sitting, bot sitting, babysitting the machine. [music] We're bot sitting, bot sitting, best you've never seen. [music] All this nonsense, Hollow sheen. We're bot sitting. Where's my curiosity? [music] I used to LOVE THE HARD PART. I USED TO LOVE THE MESS. NOW I'M THE integration layer, a human work harness. Sodomite my joy. [music] Then wonder why I'm tired. We used to come and burn [singing] in the north. Now we just feed it and let [music] it go 6 hours a week. [singing] Nobody sees. Sowing seeds [music] nobody reads. But oh, the company WE KEEP. WE'RE BOT SITTING, BOT SITTING. Baby sitting the machine. We're bot [music] sitting, bot sitting. Best you've never seen. All this nonsense. [music] Hollow sheen. We're bot sitting. Where's my curiosity? [music and singing] Bot sitting, bot sitting. Ooh, I'm baby sitting [music] the machine. Bot sitting. Bot sitting.

If you're finding value in the show, we'd appreciate it if you take a moment to share with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website cognitiverevolution.ai or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts, which is now part of a16z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at AIpodcast.ing. And thank you to everyone who listens for being part of the Cognitive Revolution.