Botsitting, botshitting, and the hidden human labor of AI at work

The gap between what we call high AI achievers (AI users who report that AI has improved both their productivity and the quality of their work) and everyone else isn’t how much AI they use. It’s where they use it. And what they refuse to hand it.

The flip side of pointing less AI at the core is that high AI achievers spend a greater share of their AI time on the work around it. They feed AI context, supervise its outputs, and debug the results. That extra effort pays off in two ways.

The first is quality control. High AI achievers are far more likely to catch botshit. 79% of high AI achievers caught and fixed an AI error in the past month, compared to 64% of low AI achievers.

The second is learning. Most botsitting is grunt work, but some is productive. And that’s where high AI achievers do much of their learning. Every prompt is practice, and every bad output is feedback. Over enough cycles, the worker builds a working theory of what the tool can be trusted with, and what still needs a human at the wheel. They are more than twice as likely to rate AI itself as a valuable teacher (68% vs. 28%). They don’t wait for an org-wide training program. They learn from the tool itself in real time and on real work.

This is why AI tool design matters more than most organizations realize. Workers who say their AI tools are easy to use are 110% more likely to rate AI itself as a valuable learning source (61% vs. 29%). A clunky tool eats the attention that would otherwise go to learning. People spend their time wrestling with the interface, troubleshooting the workflow, and figuring out why the agent stopped halfway. Organizations often treat AI usability as a nice-to-have. It isn’t. When the tool is also the teacher, a clunky UI isn’t just a UX problem. It’s a tax on learning — paid in attention, focus, frustration, and skills workers never get to build.

When AI gives a worker an hour back in time savings, what happens to that hour matters.

Low AI achievers don’t get much of a dividend to begin with. What they do get gets reabsorbed into more of the same work. High AI achievers spend theirs differently. They’re more likely to reinvest it into higher-quality work and building new and stronger AI skills — running agents, connecting tools through APIs, and debugging workflows when something breaks.

And, most importantly, they’ve learned when not to use AI at all.

Knowing when to skip the tool is often the hardest skill to build. It takes restraint, and that kind of restraint is rarely taught in a formal L&D session. It’s earned the hard way — when a polished output falls apart downstream, when a shortcut creates more botshit cleanup than it saved, when the AI sounds confident and the worker doesn’t believe it. You don’t read about that in a training module. You collect the scars. And the next time the tool offers an easy answer, you pause.

Researchers call this constructive deviance: rule-breaking by people who believe the outcome matters more than the process. These workers often aren’t reckless. They’re committed to the organization’s goals. But when policy can’t keep up with how the work actually gets done, they work around it and choose the work.

Men and executives are more likely to work around the rules. 52% of men (vs. 46% of women) and 54% of executives (vs. 48% of non-executives) admit to using unapproved tools or using approved tools in noncompliant ways. Those gaps are consistent with a long-standing pattern in organizational research: the higher your status in a workplace, the more confident you are that the rules are negotiable.

‍The smart move for organizations isn’t to hunt down every workaround like a crime scene. It’s to ask what the workaround is telling you. When a high AI achiever routes around a sanctioned tool, it’s usually because the sanctioned tool doesn’t fit the work. It’s too slow, too generic, or too disconnected from the information the worker needs.

At one of our Work AI Institute founding members’ universities, they have enterprise agreements with Google and ChatGPT, but not Claude. When faculty need Claude for work, they pay for it through their research budget. Their colleagues without research accounts simply go without. The tool isn’t banned. It’s just not funded, which means access depends less on what the work demands than on whether the worker has a discretionary budget.

Good leaders treat both kinds of workaround as feedback on their AI strategy. The goal is to make the best tool the easiest, safest, and most approved path. Otherwise, workers get stuck with a rotten choice: follow the policy, or do the work well.

For decades, the team has been a stable unit — a manager, a group of human coworkers, and the institutional scaffolding around them: org charts, performance reviews, compensation bands, hiring pipelines, succession plans. All of that scaffolding assumes teams are made of humans.
‍
That assumption is cracking. The team is becoming a hybrid unit, with humans plus a growing cast of AI assistants, copilots, agents, and digital twins. They attend meetings, draft deliverables, make recommendations, and hand work back to the humans who’ll be evaluated on it. Workers are no longer sure their human coworkers are the most useful members of the team.

High-achieving AI teams aren’t waiting for their organizations to redesign the org chart. They’re already figuring out how to manage a team where not every teammate is human.

Here’s what they’re doing.

High AI achievers don’t just use AI differently. They relate to it differently. Less like software, more like a colleague — one that’s useful, fast, available at 2 a.m., and still willing to bluff when it doesn’t know the answer. High AI achievers are 2.3x more likely than low AI achievers to say they trust AI as a teammate.

High AI achievers are more than 2x more likely than low AI achievers to trust AI as a teammate.

George Lakoff’s work on categorization shows that the labels we attach to things don’t just describe them — they shape how we behave toward them. Call AI a tool and the relationship turns transactional. You use it, it gives you something back and when it disappoints, you move on. But call it a teammate and the dynamics change. You assign it work, push back on the first draft, explain what fell short, and try again. Workers who think of AI as a tool are 26% more likely to give up after a poor output. Workers who view it as a teammate keep pushing — rerunning prompts, swapping models, adding context, and hashing the problem out with the AI.

This may also help explain why 64% of high AI achievers say collaborating with AI is easier than collaborating with human colleagues. Not because AI is smarter, but because it is easier to work with. It’s always available. It doesn’t get defensive on the tenth revision of the same paragraph. And, as far as we know, it has never scheduled a meeting that could have been a three-line email.

But the teammate frame has limits. Earlier in this report we showed how humanizing AI can inflate trust and breed more botshitting. Some organizations are blowing past that limit altogether — putting agents on the org chart, assigning them roles, and counting them as headcount. That can go sideways quickly. In a recent randomized experiment, researchers at Boston Consulting Group found that when AI was framed as an employee rather than a tool, workers felt less accountable for what it produced and reviewed its output less carefully.

The point of the teammate metaphor isn’t to make AI feel more human. It’s to give workers a working mental model for getting useful work out of it. Treat AI as a teammate, but keep the accountability where it belongs — with the human. And don’t put a bot on the org chart until you’ve built the guardrails that keep humans responsible for whatever it produces.

Many leaders try to push AI adoption from the top and assume gravity will handle the rest. They announce the strategy, bless the tool, fire up a dashboard with up-and-to-the-right vanity metrics, and declare victory at the next all-hands meeting.

Top-down change matters. It tells employees what’s allowed, what carries real risk, and which behaviors get rewarded when promotion time rolls around. But memos, mandates, and other top-down change efforts don’t, by themselves, change how people work. What changes how people work is watching their peers do it.

The average employee is¹⁶:

Why do cross-functional teammates carry so much weight? Because they’re painfully aware of the coordination tax of work: the bottlenecks, the silos, the duplicated efforts, the dropped balls. So when they build an AI workflow or an agent, they aren’t designing for some tidy fantasy version of the work. They’re designing for the messy version that they have to deal with, where the marketer needs the data the analyst hasn’t pulled yet, and the engineer needs the spec the product manager hasn’t written yet. Their workflows spread because they survive contact with real work.

This is why cross-functional workers belong at the center of your AI strategy. Identify them, give them early access to tools and training, ask for their feedback on what’s working and what isn’t, and invest in AI tools that work across functions. If your engineers and marketers can’t build or run AI workflows together in the same tool, you’re probably feeding the coordination tax you hoped AI would fix.

For decades, management has been swallowed by coordination work: chasing updates, routing information, scheduling meetings, forwarding context, keeping the work moving. AI is finally taking some of it off managers’ plates. McKinsey estimates that less than 30% of middle managers’ time goes to actual people leadership. The rest disappears into individual execution and admin work.

Managers who are high AI achievers spend about the same share of their time on coordination as less effective managers (11% vs. 9%), but they delegate 32% more of it to AI. They aren’t using AI to replace management. They’re using it to clear away the administrative sludge that gets mistaken for management.

That sounds like bad news for managers. We think it’s the opposite. The best managers don’t try to compete with AI on coordination work. They delegate the coordination work to AI, using AI to draft the status update, route the request, summarize the meeting. And they reclaim precious time for the work they ought to be spending more of their time doing — coaching, developing, and inspiring their people.

The payoff doesn’t stop at day-to-day work. Workers don’t decide how they feel about AI showing up in a performance review, a pay decision, or a termination the moment HR rolls out the system. They decide it through hundreds, sometimes thousands, of micro-interactions with their manager. The manager who uses AI to develop people teaches the team to see AI as developmental. The manager who uses it to monitor people teaches the team to see AI as surveillance. That’s, in part, why workers with good managers are far more comfortable with AI showing up in the highest-stakes decisions of their working lives, including performance reviews, pay, and termination. By the time AI walks into one of those rooms, the team has already decided whether it belongs there.

When you manage four people, it’s easier to be fair. When you manage ten, or fifty, as some organizations are experimenting with, you start running on shortcuts. You say “yes” to whoever asks first. You keep meaning to circle back to the introverts, and then forget. The result is a system that feels arbitrary and biased toward certain voices. At that point, the cold consistency of an algorithm starts to look a lot fairer than a human stretched too thin.

Workers tell us that AI automation saves them roughly 11 hours every week. Only 13% say their organization is performing significantly better because of it. So what are the 13% doing that the other 87% aren’t?

The reflex inside most organizations has been to push AI usage up: more tools, more licenses, more dashboards measuring more clicks. But the organizations actually getting performance gains from AI aren’t winning that way. Their workers spend a smaller share of their AI time inside the AI tools, not a bigger one.

Workers at organizations reporting no impact, negative impact, or no idea spend nearly half of their AI time (49%) directly inside AI tools, versus just over a quarter (27%) for workers at transformative organizations. The same pattern showed up among the high AI achievers we identified earlier, who cluster in these same transformative organizations.

They aren’t getting ahead by maxing out their tool time. They’re getting ahead on the work around the tool — setting the context, catching what the tool is too confident about, integrating the results into real work, and knowing when to leave AI out of it entirely.

Transformative organizations build management discipline around AI, such as clear rules, quality checks, shared habits, and enough psychological safety for someone in the room to say, “No, we’re not putting that bot-generated mush in front of a customer.”

The discipline shows up in five key ways.

Too many organizations measure AI through vanity metrics: tokens consumed, lines of code generated, tool clicks, login rates, dashboards that glow reassuringly upward. We’ve spoken with leaders at major tech firms that count AI-generated lines of code in performance reviews. These metrics are easy to count. They’re also easy to game. When leaders measure AI activity, workers learn to produce AI activity — and, often, little else.

One engineer at a major tech company described doing exactly that: routinely asking AI questions whose answers were already sitting in the company’s documentation, just to inflate his token numbers:

“I am conscious of not wanting to be seen as ‘uses too little AI,’ and I’m not ashamed to say I need to do tokenmaxxing to do this. Things I do to inflate my token usage metrics: Ask AI questions about the code already in the documentation. The AI pulls up the documentation, processes it, and gives me results 10x slower, but while burning lots of tokens. I could use ‘readthedocs’ [an internal product], but then my token numbers would be lower. Ask the AI to prototype a feature that I have no intention of working on. Prompt it a few more times, then throw the whole thing away. Default to always using the agent, even when I know I could do the work by hand much faster. Then watch it fail.”

That’s Goodhart’s Law in action: “When a measure becomes a target, it ceases to be a good measure.” If the metric is token volume, workers produce token volume. Whether the work is any good — or just expensive botshit — becomes somebody else’s problem.

Steven Kerr warned about this fifty years ago in his essay, “On the Folly of Rewarding A, While Hoping for B.” Organizations say they want one thing and reward another. They hope for long-term growth and reward quarterly earnings. They hope for teamwork and reward individual heroics. They hope for candor and reward the people who tell the boss what the boss already believes. Now they’re doing it again with AI.

There’s a heavy cost tied to measuring the wrong thing. 74% of workers in organizations that measure only productivity admit to botshitting. Where both productivity and quality are measured, it drops to 64%. And when organizations track quality alongside productivity, 83% of workers say AI has improved the quality of their work, compared with just 68% in organizations that track productivity alone.

Metrics don’t just measure behavior. They tell employees what the organization values and, in turn, they incentivize certain behaviors. When organizations measure quality, people slow down where human judgment matters. When they measure only speed, people speed up, pumping out faster drafts and leaving downstream cleanup to someone else.

Transformative AI organizations build measurement systems that reward the behaviors they actually want. At one organization we studied, AI success is evaluated across three dimensions: efficiency, quality, and employee experience — so the company doesn’t reward fast work at the expense of good work or good people. One Fortune 100 executive told us his company measures “intent diversity”: how many different use cases employees apply a given AI tool to. It’s a better proxy for real learning because it shows whether workers are expanding their grasp of where AI is actually useful, not just clicking the tool more often.

Transformative organizations know no single metric tells the whole AI story. They evaluate AI initiatives across five different dimensions on average, compared with three at other organizations.

Workers in transformative organizations say their organizations track an average of five different metrics, compared with three in other organizations.

In transformative organizations, employees are far more likely to say they have visibility into their own AI usage data. The data flows in two directions: up to the people running the strategy, and back to the people doing the work. Where it only flows one way (upward), it stops being feedback and starts feeling like surveillance. Workers figure out the message: This isn’t here to help me learn; it’s here to judge me. So they game the numbers, hide the messy parts, and perform whatever version of “AI adoption” leadership seems to want.

Most organizations have an AI policy by now. That doesn’t mean they have AI governance.

A policy is a document someone uploaded to the corporate intranet and hoped people would read. Governance is what actually happens when the deadline is breathing down someone’s neck, the approved tool is clunky, and the unapproved one works better. Right now, there’s a wide gap between the two.

Transformative organizations don’t just write a policy, post it, and wander off. They review it regularly, explain why it exists, and enforce it when somebody breaks it. Only 60% of workers say they’ve read their organization’s AI policy. For everyone else, it’s just another prop in the AI governance theater — written for the auditor, displayed for the board, ignored by everyone with actual work to do.

Transformative organizations also define who’s allowed to build and deploy AI agents. Without that line drawn, you get agent sprawl — three different teams have built three different bots to do the same thing, two of them are running on data the company never sanctioned, and nobody can remember who approved any of it.

The payoff isn’t just compliance. It’s confidence. Workers in transformative organizations are far more likely to say they trust their company’s AI strategy (93% vs. 57%). And that confidence drives retention.

Three actions build that confidence most:

When leaders communicate AI goals clearly, workers stop filling in the gaps with worst-case scenarios. When the information they need is actually accessible through AI, the tool stops producing generic answers and starts producing useful ones. And when AI is used to redesign the work rather than load more onto it, the credibility shows up in visible relief — not in another round of promises about what AI will eventually deliver.

Transformative organizations don’t treat governance as a communications exercise. They give it decision rights, accountability, and budget. They’re more likely to put a named AI owner in the C-suite. And their CEOs don’t just talk about AI in town halls. They use it — in front of their employees. Nvidia CEO Jensen Huang has been publicly clear that he uses AI daily, “as a tutor, a research assistant, a coach, a thought partner.”

Employees who have seen their CEO personally use AI use it 67% more than those who haven’t. Employees whose executives describe AI as a teammate are 51% more likely to treat it as one themselves, and 65% more likely to report productivity gains.

Employees watch what their leaders do with AI, not what they say about it. When leaders hype AI from the stage, monitor token counts from a safe distance, and never use the tools in the daily work themselves, employees figure out the real rules quickly: look AI-forward, keep the dashboard pointed up, and leave someone else to clean up after the bots.

For many organizations, their “AI strategy” is little more than a roll-up of their existing contracts with technology vendors. If they’re a Microsoft shop, they add Microsoft Copilot. Salesforce shop? Add Einstein. Already paying for Google Workspace? Welcome to Gemini.

This is an AI version of what James March and his colleagues called the garbage can model fifty years ago. In real organizations, problems and solutions don’t arrive in that order. Often, the solution is already rattling around inside the can, waiting for someone to find a problem to attach it to. That’s how most AI strategies unfold. The company buys an AI tool from an existing vendor, or gets one bundled into a contract they already have, and then goes hunting for places to use it. The tool comes first, the problem comes later.

Transformative organizations invert the order. They start with the work before they start with the shopping. They map where employees are stuck doing low-value drudgery, where customers keep getting frustrated, and where handoffs keep dropping. Then they pick tools that solve those problems. Employees in these organizations are much less likely to say existing vendors are limiting their AI strategy (33% vs. 49%).

Starting with the work also changes what these organizations expect from their vendors. The vendor has to understand how the work actually happens, connect to the data and context the work depends on, and stay involved long enough to define what success looks like. Employees at transformative organizations are nearly twice as likely to describe their AI rollouts as collaborative with vendors (78% vs. 44%).

This is part of what’s driving the rise of forward-deployed engineers, people whose job is to make AI work inside the actual organizations that buy it. Job postings for the role, which originated at Palantir, grew more than 800% between January and September 2025. Salesforce alone has committed to hiring 1,000 of them. In May 2026, OpenAI launched a $4-billion deployment company and acquired the AI consulting firm Tomoro, instantly adding 150 forward-deployed engineers to its team.

Starting with the work changes what matters in the technology. It’s not enough for an AI tool to be connected to the company’s data. It has to use that data the way the people doing the work actually use it.

Enterprise context is the layer between the raw data and the usable answer. It’s what tells a system which file is the actual current version of the deck, whether a spike on a dashboard means something or is a known artifact, which workaround the operations team has been using when the official process breaks down, and what “Q3” means at your company this week versus last quarter.
‍
When context is missing, employees pick up the slack. They smuggle work into unapproved tools, hide their usage, and run the same prompt across multiple platforms, hoping one finally coughs up something usable. In our data, we treat workers who say the information they need to do their work is accessible through their AI tools as “context-rich,” and those who say it isn’t as “context-poor.” The gap between the two groups is striking.

AI transformation is disruptive. Roles shift. Some disappear entirely. And when companies cite AI in layoffs, the fear cuts deeper: 73% of workers worry their own role could be next.

The impact of layoffs doesn’t end when the laid-off employees leave the building. Research on “layoff survivors” shows that downsizing increases job insecurity, erodes trust in management, and raises turnover intentions among the people who remain. When AI is cited as the reason, workers start optimizing for self-protection over careful judgment — moving fast, hiding how they use AI, pushing through work they don’t fully trust. Many start looking for the exit. At organizations that cited AI in layoffs, 62% of workers are actively job-hunting.

Both botsitting and botshitting are higher in organizations that have done layoffs, and higher still when AI is named as the reason. Layoffs put more pressure on the people who stayed, so they lean harder on AI — and they ship more work they wouldn’t defend if asked.

Transformative organizations break some of this cycle. They don’t just hand workers new tools and hope they’ll magically become AI-fluent. They reward AI skills, recognize people for good AI work, celebrate well-intentioned failures, and provide enough training and support that AI fluency becomes a normal part of professional growth, instead of a precondition for keeping the job.

One of the most important things transformative organizations do is treat AI as a chance to redesign the work — not as a shiny excuse to squeeze more output from fewer people. 90% of workers at transformative organizations say their employer treats AI as a chance to redesign work, compared with 54% at other organizations.

While botsitting and botshitting are pervasive across all three countries, they show up in different ways. In the US, AI adoption is largely driven by bottom-up experimentation, with workers absorbing much of the botsitting burden on their own. In the UK, organizations appear to give workers more structure, training, and support, which may help explain why UK workers report the strongest productivity gains despite high levels of AI use. Australia matches the UK on adoption, but weaker training and poorer context infrastructure push more cleanup and verification work back onto employees, contributing to higher fatigue, less consistent gains, and greater exposure to botshitting.

Percentages reflect the share of digital workers in each country (US n=3,000; UK n=1,500; Australia n=1,500) who selected the corresponding response. Cross-country differences of three percentage points or more are statistically significant at p<.05.

Adoption also differs by function. IT is the most confident and structurally prepared for AI, but also the most likely to test beyond approved tools. Engineering uses AI deeply, while trusting it cautiously. Marketing is scaling output faster than impact. Support has some of the clearest AI use cases, but the lowest adoption and trust. HR is adopting AI at above-average levels, even as AI moves into higher-stakes people decisions.

Percentages reflect the share of digital workers in each function (total n=6,000 across US, UK, and Australia) who selected the corresponding response. Function-level samples range from n=89 (Design/UX) to n=2,028 (IT). Comparisons in parentheses show the cross-function global average and a peer-anchor function (the top or bottom performer on that dimension). Differences of four percentage points or more are statistically significant at p<.05.

AI is landing differently across industries. Adoption ranges from 63% in government to 97% in technology, but usage is only one part of the story. Technology workers report the strongest productivity gains and the highest adoption. Healthcare workers trust AI less because the stakes are higher. Government workers use AI the least. Media and entertainment workers feel the sharpest loss of ownership. Across industries, the pattern is clear: AI creates the most value where the work is already digital, measurable, and well supported, and the most strain where trust, judgment, and human context matter most.

Percentages reflect the share of digital workers in each industry (total n=6,000 across US, UK, and Australia) who selected the corresponding response. Industry-level samples range from n=312 (Construction) to n=1,104 (Technology). Comparisons in parentheses show the cross-industry global average and a peer-anchor industry (the top or bottom performer on that dimension). Differences of four percentage points or more are statistically significant at p<.05.