Botsitting Takes Back the AI Workweek
On June 10, 2026, Glean gave enterprise AI a less flattering time sheet.
Its Work AI Index 2026 said digital workers save roughly 11 hours a week through AI automation. The same workers spend 6.4 hours a week making AI usable: feeding tools missing context, checking outputs, debugging mistakes, rerunning prompts, comparing answers across systems and cleaning up confident errors before work reaches a customer, manager or colleague.
Glean called the work “botsitting.”
Three weeks later, Business Insider reported a smaller version of the same problem. Sol Rashidi, chief strategy officer at Cyera and a senior fellow at Harvard Kennedy School, had been running four AI agents. She cut that number to two after the agents demanded too much supervision and replaced some work with human virtual assistants. It was an anecdote, not a benchmark. It still explained why the word traveled so quickly.
A company can count seats. It can count prompts. It can count how many employees used Copilot, ChatGPT, Claude, Gemini or an internal assistant in a given week. It can show that a worker feels faster. It can show that a draft arrived in seconds.
It may still miss the worker who spent 25 minutes pasting policy context into three tools, the manager who checked an AI-written analysis because the author could not explain it, the HR business partner who reworked a performance summary that sounded polished but missed the employee history, or the engineer who validated a code change because the agent did not know the team’s release rules.
The saving shows up first. The maintenance labor usually shows up later, in review queues and manager calendars.
Botsitting deserves to move from workplace slang to an operating metric. It sits between individual AI speed and organizational AI performance. Glean found 87% of digital workers using AI at work and 75% saying it makes them more productive, while only 13% said AI had significantly improved their organization’s performance. That gap is where AI adoption starts to feel strange: workers like the tools, then spend part of the saved time making the tools safe enough to use.
Picture a Monday morning staff meeting. Finance wants to know why AI spend is rising. IT says employees are using too many tools. HR says early-career workers feel pressure to adopt AI before the company has a clear standard. Managers say drafts are faster but review work is heavier. Employees say the tools save time, then hand part of the saved time back as cleanup.
The meeting should not start with a debate over whether AI works. It should start with a work map: which tasks got faster, which checks got added, which handoffs moved to managers and which costs never reached the dashboard.
June 10 named the work behind the tool
Glean’s report is not a universal labor-market census. It surveyed 6,000 full-time digital workers in the United States, United Kingdom and Australia between December 2025 and January 2026, and it also drew on aggregated workplace AI interaction data from Glean’s own platform. The sample is about computer-based work, not every occupation.
The boundary makes the finding more useful for enterprise buyers. Digital work is where AI was supposed to move fastest. If the hidden labor bill is already visible there, less digitized sectors should not expect adoption to be cleaner.
Workers said AI saves them about 11 hours a week through automation. But 6.4 hours of the same week goes to botsitting, which Glean defines as the work required to make AI usable. In the time allocation, botsitting made up 37% of workers’ AI time, slightly more than the 36% spent using AI to produce work.
For a leadership team, that is a productivity problem before it becomes a fatigue problem.
An organization that sees only the 11-hour saving may assume it has created capacity. An organization that tracks the 6.4-hour maintenance load sees a different balance sheet. Some of that labor belongs in the workflow: verifying a high-risk answer, improving a prompt, adding domain context that a model could not know. Some of it is waste: reloading the same context into multiple tools, comparing outputs because the first answer was not good enough, fixing handoff breakage, or correcting work that should not have been generated in the first place.
Good botsitting and bad botsitting lead to different management decisions.
Good botsitting belongs in the work design. If a financial analyst uses AI to draft a forecast note, the company should expect a verification step. If a recruiter uses AI to summarize interview feedback, the hiring manager should own the final quality bar. If a support team uses an agent to classify cases, exceptions should route to a person who understands the customer impact. That is human judgment attached to a faster workflow.
Bad botsitting is a tax on fragmented adoption. It appears when employees must paste confidential context into tools because official systems are not connected. It appears when a manager asks for AI usage but does not define quality. It appears when a worker ships a plausible answer without knowing whether the answer came from current policy, old documents or a hallucinated pattern. It appears when three AI tools produce three different answers and the employee becomes the arbiter without authority, training or time.
Glean found the risk turning into quality failure. Sixty-nine percent of AI users admitted to shipping AI-generated work they had not verified, did not fully understand or could not confidently stand behind. Forty-one percent said they sometimes delivered work they could not explain if asked. Twenty-eight percent said they had blamed AI for mistakes they caused.
Rebecca Hinds, who leads Glean’s Work AI Institute, framed the problem as a vanity-metric trap: more seats, more prompts and more usage can still leave employees spending the dividend on overhead. The tool can make a worker faster. It can also make a weak handoff look finished.
Leaders need a harsher metric than usage. High usage can mean employees are productive. It can also mean employees are compensating for weak systems.
Glean counted the hidden workday
The hidden workday starts with context.
AI systems often need information that sits in another application, another folder, another chat history or another person’s head. Glean found that 53% of workers say needed information is not accessible from their AI tools. Among workers using context-poor AI, 50% felt worn out by AI compared with 18% among workers whose tools had better context. They were also more likely to clean up after AI weekly, ship work they could not explain and use unapproved tools.
Here the enterprise AI promise meets the enterprise IT backlog. The best model in the world still fails if it cannot see the policy, customer record, codebase, spreadsheet, performance history or approval rule that makes the answer correct. When that happens, the employee becomes the connector.
The second part is verification. Verification is not a defect; it is the price of using probabilistic systems in real work. But the price changes by task. A grammar pass on an internal note needs light checking. A compensation recommendation, legal memo, code change, candidate screen, safety report or financial forecast needs a heavier review path. If all AI output gets the same review standard, the company either wastes time on low-risk work or under-checks high-risk work.
The third part is stitching. Glean found that 77% of AI users juggle multiple AI tools each week, and 33% use four or more. Sixty percent rerun the same prompt across multiple tools because the first output was not good enough. Workers have become the routing layer between disconnected systems. The company may call that flexibility. The worker experiences it as tool management.
Trust leaks out of those small repairs.
When AI output passes through too many systems and too few standards, nobody knows what “good” means. Workers may hide how much AI helped because they fear being judged. They may use unapproved tools because the official tool is weaker. They may ship work faster because speed is rewarded, even when explanation is not. They may avoid AI on tasks where the tool would help because the review burden feels worse than doing the work manually.
Glean’s high-performer finding cuts against a simple automation story. High AI achievers, defined as people reporting both productivity and quality gains, spend more of their AI time botsitting than low achievers. They are also more likely to work around the rules: 54% use unapproved tools or approved tools in noncompliant ways, 38% downplay AI’s help to their manager and 36% hide how much AI helps them.
The best users are not the people who accept every output. They are the people who know when to supervise, when to reject and when to keep the work human.
Companies already reward prompt fluency, model experimentation and tool adoption. They should also reward context judgment, verification discipline and task refusal. A worker who says, “This should not go through the agent,” may protect more value than a worker who uses AI on every assignment. A manager who protects review time may create more reliable productivity than a manager who only pushes usage.
Seat adoption tells leaders who touched AI. Botsitting tells them who paid for that touch.
Workers want agency before automation
Stanford SALT Lab’s Future of Work with AI Agents project built a worker-centered audit using 1,500 workers across 104 occupations and 844 tasks. Workers expressed positive desire for AI agent automation on 46.1% of tasks after considering concerns such as job loss and reduced enjoyment. The project also found mismatches between what workers want automated and where AI research or startup investment appears to concentrate.
Workers are sorting tasks by agency.
Some work is tedious and low identity. Workers may welcome help with scheduling, data cleanup, document formatting, first-pass summaries or routine status updates. Some work is complex but frustrating. Workers may want an agent to prepare options, retrieve context or draft alternatives while a person keeps authority. Some work defines professional judgment, care, creativity or accountability. Workers may resist handing it over even if the model can produce something that looks adequate.
Workday’s global agent research makes the boundary visible in enterprise language. In its 2025 release, 75% of workers said they were comfortable teaming up with AI agents, but only 30% said they were comfortable being managed by one. Only 24% were comfortable with agents operating in the background without human knowledge.
Kathy Pham, Workday’s vice president for AI, described the trust challenge as keeping people at the center of every decision. That matters for HR, finance and operations teams because agent deployment often begins in gray zones. A tool may recommend skills, rank candidates, summarize goals, flag expenses, route cases or generate manager prompts. The employee may be comfortable with an agent assisting the workflow but not with the agent deciding the outcome, hiding its involvement or becoming the boss.
Workday also found that 82% of organizations were expanding their use of agents, while workers worried about pressure, weaker critical thinking and less human interaction. That combination is where botsitting turns into an employee-experience problem. The tool does not need to fire anyone to change the job. It only needs to make workers responsible for explaining, correcting and absorbing work they did not design.
SHRM’s Navigating AI in the Workplace: 2026 report adds the adoption baseline. SHRM surveyed 5,875 U.S.-based workers in March and April 2026. Forty-one percent of workers reported using AI in their work. Forty-seven percent said their organizations had implemented AI, while an equal share said their organizations had not. Entry-level and early-career professionals felt the most pressure to adopt AI tools, with 45% reporting pressure to use AI in their roles.
The result is messy. Some workers are using AI before their organizations know how to support it. Some organizations are rolling out AI before workers trust the boundaries. Some early-career workers feel pressure to use tools at the same time they are still learning the job itself. SHRM also found that just under half of AI-using workers described some of their output as “AI slop.” That phrase belongs in the same operating conversation as Glean’s “botsitting”: employees are making AI output presentable, explainable and safe enough to ship.
Botsitting sits in the middle of that tension. If the company gives an employee a tool, a metric and no support, the employee owns the uncertainty. If the company gives a worker clear task boundaries, review standards, disclosure rules and a path to escalate bad output, AI feels more like assistance and less like surveillance.
The difference shows up in the daily workflow. A worker who knows which agent tasks are advisory can use the output without anxiety. A worker who knows which decisions require human approval can preserve authority. A worker who knows when to disclose AI involvement can avoid shame and hidden workarounds. A worker who knows where to report repeated AI failure can improve the system instead of silently cleaning up after it.
Context-poor tools send work back to people
AI vendors describe tools that understand the enterprise. They connect to documents, email, chat, tickets, repositories, HR systems, finance systems and customer records. They promise grounded answers, permission-aware search, workflow execution and agent orchestration. The product direction is real. Glean itself sells an enterprise context layer. Microsoft, Workday, ServiceNow, Google and other platforms are building similar ideas into their agent strategies.
Enterprise context is not a switch. It is a map of permissions, data quality, system ownership, policy rules, workflow history and team norms. A model may technically connect to a repository and still miss why a service is fragile. It may read an HR policy and still miss the local exception. It may summarize a customer account and still miss the relationship context that lives in a manager’s notes. It may draft an answer using old source material because nobody retired the stale page.
When context is poor, workers do the missing integration by hand.
From the outside, that can look like prompt craft. Inside the company, it is organizational debt. A worker copies the latest policy into a prompt because the AI tool cannot retrieve it safely. A manager asks for a more “executive-ready” version because the first draft lacks the real decision history. A recruiter edits an AI candidate summary because the tool did not understand why a role changed. A finance analyst checks an AI variance explanation against the original spreadsheet because the model saw the dashboard but not the adjustment note.
Each correction is small. Across a week, it becomes the 6.4-hour tax.
Lightcast’s Stanford AI Index 2026 research summary shows why the labor market is moving toward this kind of work. AI skills appeared in 2.5% of U.S. job postings, up 55% from the prior year and 297% from a decade earlier. Mentions of the “Agentic AI” skill cluster rose more than 280% year over year, from 0.06% of postings in 2024 to 0.23% in 2025, roughly 90,000 U.S. postings.
The demand signal is no longer “can you prompt a chatbot?” It is shifting toward building and managing systems at scale. Lightcast notes that skills such as workflow management, scalability and deployment-oriented capabilities are becoming more visible. The labor-market version of the botsitting problem is plain: employers need people who can make AI work inside real operating systems.
For HR leaders, this means AI fluency should not be reduced to tool familiarity. The useful skill is knowing how to bring the right context to a task, recognize missing context, set a verification standard, protect confidential data, explain an AI-assisted output and decide when automation should stop.
For CIOs, context access is worker experience infrastructure. If employees spend hours feeding context into AI tools, the company is paying for missing architecture through labor time.
For CFOs, the AI business case needs the context bill. A vendor may claim time savings, but the company should ask how much time employees spend preparing, checking, repairing and rerouting the output. If that time falls on high-cost managers or specialists, the ROI can change quickly.
Context is the place where AI adoption becomes work.
Microsoft found the handoff gap
Microsoft’s 2026 Work Trend Index gives botsitting a management frame: workers may be ready before the organization around them is ready.
The survey placed only 19% of AI users in the “Frontier” zone, where individual readiness and organizational capability reinforce each other. Ten percent were in “blocked agency,” meaning individuals had built strong AI skills but lacked the systems to apply them. Sixteen percent were stalled. The largest share sat in the middle, where personal practice and organizational conditions were still forming. Only 26% of AI users said leadership was clearly and consistently aligned on AI.
That misalignment is where handoffs fail.
An AI workflow is rarely a single act. Someone asks for an output. The system retrieves context. The model produces a draft, recommendation or action. A person checks it. Another system consumes it. A manager signs off. A customer, candidate, employee or regulator may later ask why the result happened. If the organization has not defined the handoff points, botsitting becomes improvised.
Microsoft says organizations need an evaluation infrastructure for agents because risk compounds when bad outputs pass through at scale. Its Work Trend Index points leaders toward three questions: who reviews agent performance, who can update agent workflows and how a local improvement gets captured across the organization. Those are botsitting questions, even if the report uses different language.
Consider a simple HR example. An AI assistant drafts a manager response to an employee’s leave question. The answer depends on policy, jurisdiction, team coverage, employee history and manager discretion. If the system has current policy and the manager has a checklist, the workflow is manageable. If the system gives a generic answer and the manager must remember every exception, the tool saves typing while moving risk to the manager.
Engineering, finance and sales show the same structure. A coding agent opens a pull request; a reviewer must know whether the change touched a high-risk system. A finance assistant drafts a variance explanation; the analyst must know whether the source data reflects the latest adjustment. A sales agent drafts account research; the account executive must know whether the customer history is complete.
Handoff standards turn invisible labor into planned labor. They answer three questions before the work begins: which outputs are advisory, which require approval, and which should never be automated without a human owner?
BCG’s 2026 analysis raises the stakes. It estimates that 50% to 55% of U.S. jobs could be reshaped by AI over the next two to three years. BCG’s point is not that every job disappears. It is that many people will keep the same or similar roles while facing new expectations for how they work and what they produce. Workforce strategy, in BCG’s framing, cannot sit downstream of automation.
If half of roles are reshaped, the company cannot rely on employees to absorb the redesign informally. It needs a work map.
A botsitting cost map for AI rollouts
The following map is a practical way to bring botsitting into AI rollout governance without turning it into another abstract control word. It asks where the hidden work appears, who owns it and which signal shows that AI is moving work instead of reducing it.
| Cost surface | Worker symptom | Management question | Default owner | Failure signal |
|---|---|---|---|---|
| Context feeding | Employees paste policies, files or history into prompts before useful output appears | Which source should the AI tool access directly, and who owns source freshness? | CIO / data owner | Workers repeat the same context setup across tools or teams |
| Output checking | Workers verify every answer because quality varies by task | Which outputs need light review, expert review or no automation? | Manager / risk owner | Review time rises while AI usage rises |
| Tool switching | Employees rerun prompts across multiple AI systems | Which tool is authoritative for each workflow? | IT / function leader | Four tools answer the same question differently |
| Handoff repair | Managers rewrite AI output before it can move downstream | Where should the workflow pause for human decision or approval? | Team manager | AI speeds drafting but slows final decision |
| Explanation burden | Workers cannot defend AI-assisted work when challenged | What evidence must travel with the output? | Legal / quality owner | Work is accepted until someone asks why |
| Skill pressure | Early-career workers feel forced to use AI without knowing the job standard | Which AI tasks teach judgment, and which skip apprenticeship? | HR / manager | Juniors ship polished work they cannot explain |
| Shadow tooling | High performers use unapproved tools because official tools lack context or capability | What workaround reveals an unmet product or workflow need? | CIO / security / function owner | Policy violations cluster around high-value tasks |
| Reward mismatch | Employees hide AI use or overuse AI because incentives are unclear | What counts as high-quality AI-assisted work? | HR / performance owner | Usage rises but trust and quality fall |
| Business proof | Teams report time saved but the organization does not perform better | Which business outcome should saved time improve? | CFO / COO | AI metrics stop at seats, prompts or hours saved |
The map separates adoption from transformation. Adoption means workers use AI. Transformation means the company can say which work changed, which human effort remained, which quality bar improved and which business result moved.
It also prevents leaders from treating all botsitting as failure. Some supervision is part of responsible AI. A legal team should check AI-drafted language. An engineer should review agent-generated code. A hiring manager should read the interview summary before using it in a decision. Human review is not the problem. Unplanned, unrewarded and unmeasured review is.
HR and finance should sit in the same conversation here. HR sees employee fatigue, skill pressure, trust and role redesign. Finance sees seat cost, tool overlap, manager time and productivity claims. IT sees context, permissions, security and tool architecture. The manager sees the actual handoff. None of those views is complete alone.
The useful operating question is sharper than “How do we remove people from the loop?” Leaders need to ask which human work is worth preserving, which human work exists because the system is bad, and which human work should be budgeted because the decision is high stakes.
That question can change purchasing decisions. A company may buy fewer generic AI seats and invest more in context integration. It may slow one agent rollout because the handoff standard is weak. It may fund manager training before expanding usage. It may create a new quality metric for AI-assisted work. It may reward employees who prevent bad automation as well as employees who automate more.
AI ROI gets easier to discuss once the hidden labor has an owner.
Managers need time back, not another shadow queue
Botsitting ultimately lands on managers because managers own the point where work leaves the tool and enters the organization.
The employee may generate the draft. The AI system may produce the summary. IT may approve the tool. Finance may pay the bill. But the manager is often the person who has to decide whether the output is good enough, whether the worker understood it, whether the handoff is safe and whether the team can rely on it next time.
AI adoption can quietly increase manager load. A manager used to review work. Now the manager reviews work plus the worker’s relationship with the tool. Did the employee use the right context? Did the model invent a fact? Did the worker understand the answer? Did the output skip a stakeholder? Did the team learn anything, or did it only move faster?
Without a time budget, the bargain turns bad. The worker gets pressure to use AI. The manager gets pressure to approve AI-assisted work. The organization gets a usage metric. The hidden review queue grows in the spaces between meetings.
A better bargain treats botsitting as a design signal. If a team spends too much time feeding context, fix the context layer. If a team spends too much time checking low-risk output, automate or narrow the review. If a team spends too much time repairing agent handoffs, rewrite the workflow. If early-career workers rely on AI without understanding the work, rebuild the apprenticeship path. If high performers use shadow tools, find out what the official stack cannot do.
Companies cannot solve this by pausing usage. Workers are already using it. SHRM, Glean, Microsoft, Workday and Lightcast point in the same direction: AI has entered the workplace faster than many organizations have rebuilt work around it. The bigger risk now is unmanaged adoption, not empty seats.
The advantage will go less to companies that bought the most tools and more to companies that reduced the unpaid work around them.
The AI workweek has been rearranged into prompts, review, context feeding, manager handoffs and skills that job postings are only beginning to name.
Glean’s 6.4-hour number gives leaders a place to start. Count the hidden labor. Ask which parts are human judgment and which parts are system debt. Then redesign the workweek around that answer.
If companies ignore botsitting, AI will still make people faster in moments. It will also leave them carrying a shadow queue that no dashboard owns.
Published July 6, 2026.