UK Workers Lose Nearly Six Hours a Week 'Botsitting,' and the Compliance Risk Is Worse Than the Lost Time

A report from the Work AI Institute, the research arm of enterprise search vendor Glean Technologies, puts a number on a problem that compliance and operations teams have been circling for a while: the time UK staff save with AI is being clawed back by the effort of keeping those tools accurate, current, and defensible.

The firm calls it "botsitting," and on average it costs British workers 5.8 hours a week.

What the report found

The Work AI Institute surveyed 1,500 digital workers for "The Work AI Index: UK 2026." The headline adoption figures are striking. Ninety percent of respondents are now required to use AI in their roles, 80 percent use multiple AI tools each week, and 39 percent use four or more. Workers estimate that AI automation saves them roughly 12 hours a week, close to a third of their working time.

The catch is what happens to those saved hours. Only 18 percent of respondents agreed that AI has significantly improved their organization's performance. For every hour spent getting output from an AI tool, workers reported spending roughly another hour making that output usable. More than a third of AI sessions, 36 percent, were said to fail outright, requiring a full restart or substantial rework.

The botsitting hours go into loading context the tool should already have, reviewing answers for errors or omissions, and the re-prompt cycle: add context, swap models, prompt again, repeat until something usable comes back. The report frames workers as the de facto integration layer between their company's AI systems, manually telling each tool which sources to trust and which documents are current.

Standards meant to solve this, including ordinary APIs and the Model Context Protocol, let tools exchange data but do not resolve the question of which data is correct and authoritative in a given moment. That judgment still falls to a person.

Why this matters beyond productivity

For a compliance function, the productivity math is the less interesting part of this report. The more consequential finding is behavioral. Seventy percent of UK AI users admitted to passing on the first output that looked "good enough," without verifying sources or checking whether the recommendation made sense. When diligence slips, the report notes, the error does not disappear. It moves downstream to colleagues who never touched the original work and now have to find and fix a problem they did not create.

That pattern matters because of where AI is now being applied. The report states that more than half of UK workers are comfortable with AI playing a role in performance evaluation, and nearly 40 percent say it is already used in reviews. British workers reported being more comfortable than their American counterparts with AI involvement in hiring, promotion, compensation, and even termination.

This is the point where the productivity story becomes a regulatory one.

What UK law already requires

Automated decision-making about staff does not sit in a legal vacuum. The relevant obligations come primarily from the UK General Data Protection Regulation (UK GDPR) and the Data Protection Act 2018, both already in force and enforced by the Information Commissioner's Office.

Article 22 of the UK GDPR restricts decisions based solely on automated processing, including profiling, where those decisions produce legal or similarly significant effects on a person. A termination, a pay decision, or a promotion denial clearly clears that bar. Where Article 22 applies, employers must generally secure a lawful basis, give the individual the right to obtain human intervention, allow them to contest the decision, and provide meaningful information about the logic involved.

The phrase that should concern compliance officers is "based solely on." A worker who rubber-stamps an AI recommendation because it looks good enough has, in practical terms, removed the meaningful human review that keeps a decision out of the Article 22 category. The human is nominally in the loop but not actually exercising judgment. Regulators and tribunals look at substance, not the org chart.

The forthcoming reforms under the Data (Use and Access) Act 2025 adjust the framework around automated decision-making, easing some restrictions for non-special-category data while preserving safeguards for the most sensitive cases. Employers should not read that as a relaxation of the duty to keep humans genuinely engaged in significant employment decisions. The safeguard requirements, the right to human intervention, to contest, and to an explanation, remain central.

What this means in practice

The report's own conclusion is that adoption alone does not equal transformation, and that the value of AI investment depends on operational discipline. From a compliance standpoint, that discipline translates into a few concrete obligations.

First, document where AI touches employment decisions. If you cannot say which tools feed into hiring, performance, pay, or dismissal, you cannot demonstrate the human oversight Article 22 expects. A data protection impact assessment is required where automated processing is likely to result in high risk to individuals, and HR decision support generally qualifies.

Second, make human review real rather than nominal. A reviewer who approves the first plausible output is not providing the meaningful intervention the law contemplates. The 70 percent "good enough" figure is, in effect, a compliance failure rate. Build in steps that force a reviewer to record what they checked and why they agreed or overrode the system.

Third, keep records of logic and sources. The right to an explanation means an employer must be able to describe, in understandable terms, how a decision was reached. If the AI's reasoning cannot be reconstructed because nobody captured which documents or model versions were used, that explanation cannot be produced after the fact.

Fourth, treat the botsitting time as a signal, not just a cost. Sessions that fail 36 percent of the time, and outputs that need an hour of correction for every hour of generation, indicate that the tool is not yet reliable enough to lean on for consequential decisions. The same unreliability that wastes time also produces the errors that create legal exposure.

Dr Rebecca Hinds, head of the Work AI Institute, put the operational version of the point plainly: if employees spend the productivity dividend on botsitting, companies have not eliminated work, they have created a new layer of overhead. The compliance version is that the overhead is not only time. It is the obligation to prove, when an employee challenges an AI-influenced decision, that a person actually thought about it.

The UK has moved faster than most countries on AI adoption, including in higher-stakes areas where the law is most demanding. That combination, deep adoption inside a tightly regulated employment regime, is exactly where speed without verification becomes a liability. The fix is not less AI. It is the documentation, genuine human oversight, and source discipline that the existing rules already require, applied before a tribunal asks to see them.