Botsitting: the work we don’t count

Written in a personal capacity; views expressed are my own. For transparency, I am Chief Scientific Adviser at the Department for Digital, Culture, Media and Sport (DCMS).

In August, writing about Which games not to play, I discussed some of the longstanding problems of automation and human oversight, including a paper we had recently submitted on the emerging phenomenon of botsitting. That paper, Botsitting: Conceptualizing Value-Enabling Hidden Human Labor in AI-Enabled Work, has now been published in the Journal of Computer Information Systems. Led by Yogesh Dwivedi, with colleagues from across the international information systems and AI research community, it examines the frequently overlooked human labour required to turn AI-generated potential into something reliable, contextually appropriate and organisationally valuable.

The 2026 Work AI Index, published by Glean’s Work AI Institute, helped bring the term botsitting into wider discussion. Drawing on a survey of 6,000 digital workers across the US, UK and Australia, it estimated that respondents spent an average of 6.4 hours each week providing context for AI systems, checking their outputs and correcting errors. Rebecca Hinds and Paul Leonardi, both co-authors of the Index, subsequently examined the implications for organisational productivity in Harvard Business Review. The findings also attracted wider coverage, including in The Register, which highlighted the time UK employees reported spending supervising and correcting AI outputs.

These estimates should be interpreted with some caution. The Work AI Index was produced by Glean, a commercial provider of enterprise AI software, and although the research combined survey responses with interviews and anonymised, aggregated platform data, the headline figures on time spent botsitting are self-reported. They are therefore subject to recall bias and differences in how respondents understand and classify their activities, and should not be treated as independently measured working time. Nevertheless, the underlying human work is familiar from earlier generations of automation, where people have had to interpret, check and correct system outputs. Generative AI has brought renewed attention to these activities as more sophisticated outputs are incorporated into everyday professional and organisational processes.

The question our paper addresses is how this work contributes to the value organisations realise from AI. Drawing on established information systems research into technological value, organisational capabilities and human–AI collaboration, alongside scholarship on hidden human labour, we develop a conceptual account of botsitting as a distinct form of value-enabling work, particularly where its contribution to successful AI adoption has not been anticipated, resourced or measured.

From technological capability to organisational value

The relationship between technological capability and organisational performance has been a longstanding concern in information systems research. Brynjolfsson’s influential 1993 paper on the productivity paradox of information technology examined why substantial investment in computing did not appear to produce corresponding improvements in measured productivity. Subsequent research, including Melville et al.’s model of IT business value (2004), established the importance of complementary organisational capabilities, management practices, skills and processes in determining whether technological investment delivers meaningful benefits.

Recent research on generative AI illustrates how closely productivity gains depend on the nature of the task. Noy and Zhang’s experimental study of generative AI and professional writing (2023) demonstrated significant improvements in task performance, while Dell’Acqua et al.’s research on the jagged technological frontier (2026) shows how AI assistance can improve or worsen performance depending on whether the task falls within the system’s capabilities.

Consider the preparation of an analytical briefing for a policy decision: an AI system may generate a plausible draft in minutes, drawing together evidence, arguments and recommendations that would previously have required considerably more time to assemble. However, a professional still needs to establish whether the sources are appropriate, the analysis is defensible, relevant context has been understood and the recommendations are suitable for the decisions being made. The reliability of the finished briefing depends on the wider socio-technical arrangements through which evidence is scrutinised and advice validated, an issue I also explored in my post in June on seven years of AI research.

The paper describes the distinction between technological potential and realised organisational value as the AI Value Realisation Gap. We argue that botsitting represents one complementary form of human labour through which that gap may be narrowed. This draws on wider research into human–AI collaboration and the automation–augmentation paradox (2021), recognising that the benefits of increasingly capable systems depend in part on how technological capabilities interact with human expertise and organisational practice.

A central distinction in our conceptualisation concerns the difference between deliberately designed human oversight and the additional work that emerges when those arrangements prove incomplete. Human involvement may be specified within an AI-enabled process, with allocated responsibilities, review points and appropriate resources. Botsitting concerns the initially unplanned and frequently hidden labour required when the practical demands of verifying, interpreting, correcting or managing AI-generated outputs have not been fully anticipated. It is therefore more specific than routine AI use, casual prompting or human–AI collaboration in general.

Gray and Suri’s Ghost Work (2019) examined the often unrecognised human labour supporting digital services presented as automated. Scholarship on invisible work has frequently addressed the people involved in producing, training and maintaining AI systems. Our paper shifts attention towards the complementary labour required during organisational use, after deployment, when AI-generated outputs must be made dependable in particular professional and institutional settings.

Verification and correction should not necessarily be regarded as wasted effort. A specialist who identifies an important error, recognises a missing assumption or exercises professional judgement about an AI-generated recommendation may be contributing considerable value. The difficulty arises when that contribution is neither anticipated nor recognised, while the apparent efficiency of the automated process is evaluated independently of the human work that makes its outputs trustworthy.

Counting the work

When AI is introduced into organisational work, the time required to produce an initial output represents only one part of the productivity calculation. If a system reduces that time but creates additional effort in reviewing or reworking what it produces, the return needs to be assessed across the complete process. Substantial human review may still be worthwhile where it improves the quality of the outcome or reduces material risk, but those benefits and costs need to be included in the same assessment.

We argue that botsitting should be considered explicitly in AI investment decisions and assessments of return on investment, including during pilots and proofs of concept when the practical demands of AI-enabled workflows first become apparent. The amount of time spent on human review matters, but so does how effectively that work is undertaken. Our paper distinguishes efficient botsitting, which concentrates on the evidence and criteria needed to establish whether an output is trustworthy, from inefficient botsitting, where a lack of transparency or traceability forces somebody to reconstruct the underlying analysis. This makes the design and effectiveness of review as relevant to productivity as the time spent performing it.

Productivity improvements attributed to one team or process may depend upon checking and rework undertaken elsewhere, sometimes by people who had little involvement in deciding how the technology should be introduced. Conventional measures of efficiency can miss this displaced work, particularly where responsibilities are absorbed into existing roles without explicit recognition or changes to workload.

The limits of human oversight

As AI systems move beyond producing individual outputs towards managing workflows and undertaking actions with increasing autonomy, the practical limits of human oversight become more significant. In Human rights in the loop, I examined whether human involvement in AI-enabled decision-making provides genuine controllability, authority and contestability, particularly where decisions affect people’s rights and access to public services. Botsitting approaches a related problem through the work required to make those arrangements function in practice. Assigning somebody responsibility for checking an automated system offers limited assurance if that person lacks the information, expertise or time needed to exercise meaningful judgement.

Bainbridge’s Ironies of Automation (1983) identified how reliable automation can leave people responsible for rare and difficult interventions while reducing their opportunities to maintain the skills needed to perform them. Green’s study of human oversight of government algorithms (2022) similarly questions whether formal requirements for human involvement can deliver the protections expected of them without sufficient evidence about how oversight operates in practice. Elish’s (2019) concept of the moral crumple zone identifies a related risk: responsibility for the consequences of an automated system may fall upon a person who had little practical control over its behaviour.

Anthropic’s October 2026 study of its own workforce describes the demands placed on human attention when employees supervise multiple AI agents. One engineer described facing “an army of Claudes, with every one asking for my attention”. Anthropic is an unusually AI-intensive organisation, which limits how far its experience can be generalised. The account nevertheless illustrates the problem developed in Proposition 7 of our paper: a widening gap “between nominal human oversight and its meaningful exercise”.

The limits of effective oversight are consequential for public institutions, where AI adoption must sit alongside established responsibilities for accountability, transparency and the lawful exercise of authority. The AI Playbook for the UK Government (2025) addresses the importance of appropriate human control and oversight, while the Algorithmic Transparency Recording Standard, first published in 2021, provides a mechanism for documenting relevant uses of algorithmic tools in public-sector decision-making. The publication of guidance and transparency records establishes expectations for responsible adoption; effective oversight also requires evidence that reviewers can intervene in practice.

As a conceptual contribution, our paper advances a definition, model and seven research propositions rather than reporting new empirical findings. Further research is needed to establish where botsitting contributes value, how its costs are distributed across different forms of work and which approaches to AI system design allow human review to remain effective.

Our current research on agency and delegation in information systems extends these concerns to AI systems that plan and act on behalf of people and organisations. We are examining the relationship between the authority formally delegated to an AI system, the capabilities it exercises in practice and the expectations of those responsible for its oversight. One question I want to pursue empirically is how organisations establish the limits of effective human oversight before delegating authority to autonomous agents operating across interconnected processes. This also raises questions about which judgements can legitimately be delegated and which responsibilities must remain with accountable human decision-makers and institutions.

For organisations, the implication is to recognise the work that botsitting reveals and bring it into the design and resourcing of AI-enabled processes, with its contribution reflected in assessments of productivity. What begins as unplanned verification may become part of an established assurance capability. Human review capacity must also shape how much authority an AI system is permitted to exercise. Where agents can act faster than their actions can realistically be examined, assigning additional supervisory responsibilities will not, on its own, restore meaningful oversight.


Reference

Dwivedi, Y. K., Crick, T., Michael, K., Samothrakis, S., Sarker, I., Naeem, R. M., Raman, R., Malik, T., Hughes, L., Raman, R., Dutot, V., Jeon, I., & Helal, M. Y. I. (2026). Botsitting: Conceptualizing Value-Enabling Hidden Human Labor in AI-Enabled Work. Journal of Computer Information Systems. Advance online publication. https://doi.org/10.1080/08874417.2026.2749411

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.