AI agents and recovery

In the note this review sits beside, The only thing that stops me is a usage bar, I describe a working pattern I like and have reasons to question. I can keep several agents busy. There is usually something ready to check, fix or start. The work is interesting. Stopping is largely my responsibility.

I wanted to know what happens to recovery when someone works like that.

The evidence assembled for this review did not identify a study pairing the timing and intensity of heavy, self-directed AI-agent use with recovery, sleep or burnout. The search has limits, which I set out below. That finding leaves a specific question open.

There are studies of AI use and exhaustion, trials of clinical AI tools, and a substantial body of research on detachment, work addiction and autonomy. Some findings give me reasons for concern. Others suggest that AI can reduce strain.

AI use and recovery

Liu and colleagues surveyed 6,726 radiologists across 1,143 Chinese hospitals. Those reporting regular or consistent use of AI for image interpretation had higher adjusted odds of burnout than those reporting infrequent use or none. The odds ratio was 1.20, with a 95% confidence interval of 1.10 to 1.30. Weighted burnout prevalence was 40.9% versus 38.6%. That is a 2.3 percentage-point difference; the odds ratio does not mean 20% more radiologists had burnout. This was a cross-sectional survey, so it cannot establish whether AI use preceded burnout or caused it. It concerns radiology workflows, with self-reported use C1.

A Finnish study started with 2,109 employees and collected three waves of survey data. Its main analyses found no significant association between AI use and work exhaustion. The frequency analysis grouped daily users together. A short daily task and hours of daily use could therefore land in the same category. An exploratory analysis found an association among people with higher social-comparison tendencies, but the main interaction tests were null. These were self-reported, observational data; the models did not test whether earlier use predicted later exhaustion C2.

Those findings come from different jobs, different exposure measures and different outcomes. They give us no settled dose-response relationship for agent use, and no basis for treating daily use as uniformly harmful or harmless.

The clinical trials also deserve space. In a 24-week randomised trial involving 66 healthcare practitioners, an ambient AI scribe reduced a combined measure of work exhaustion and interpersonal disengagement by 0.44 points on a five-point scale. The 95% confidence interval was a reduction of 0.25 to 0.62 points. The tool drafted clinical notes. The abstract supports a benefit in that setting; it cannot tell us what happens when someone uses agents to keep expanding their own workload. I have checked this finding at abstract depth C3.

Sleep has also appeared in AI research. Tang and colleagues studied 794 employees across four studies using surveys, a field experiment and a simulation. They reported general support for a pathway involving AI interaction, loneliness and after-work insomnia. I have read the abstract, which gives no effect sizes. It does not answer the question about logged agent sessions or late-night checking C4.

For recovery itself, the evidence is more established. Wendsche and Lohmann-Haislah’s meta-analysis covered 91 independent samples and 38,124 employees. Mentally switching off from work was associated with less exhaustion and better sleep. Much of the evidence was observational and based on self-report, so the associations cannot establish that improving detachment caused those benefits C5.

A five-day diary study of 116 employees, with 476 daily observations, found that more work-related smartphone use after work was associated with poorer detachment. Use and detachment were reported together in the evening, leaving their order unresolved. Someone struggling to switch off might check their phone more. Checking might also keep work in mind. The study cannot settle that direction C6.

That is a reasonable basis for investigating agent notifications and repeated checking. Applying it to agents is still an inference. The studies above did not observe someone moving between several autonomous jobs while trying to finish their day.

Difficulty stopping

In the note, I call the loop addictive. That describes how it feels to me. Establishing a behavioural addiction takes more than a long session or a large bill.

Kardefelt-Winther and colleagues proposed that research on behavioural addiction should centre persistent behaviour causing substantial impairment or distress. Their paper is a conceptual argument, rather than a diagnostic instrument. It is useful here because it challenges the habit of treating intense involvement as sufficient evidence of a disorder C7.

There is also a reason to examine the working pattern someone brings to a tool. A two-wave survey of 1,100 full-time workers found reciprocal predictions between work addiction and social-media addiction. Its network analysis found the stronger prospective associations running from work addiction towards social-media addiction. This was observational research on social media. The abstract does not supply effect sizes, and it does not establish a cause C8.

That possibility is personally inconvenient. Some of the behaviour I might blame on an agent could have arrived with me.

A 2026 preprint gets closer to the tools in question. Across samples in China, Germany and the UK, totalling 1,490 participants, compulsive working was associated with reported LLM dependency, with a weaker pattern in the UK. The study was cross-sectional. I have checked its abstract; the cited version is a preprint, and the reported findings do not test recovery or sleep. It raises a question about existing work habits and AI dependency without establishing which produces which C9.

The unpredictable payoff from an agent is another possible explanation for repeated checking. Sometimes the result is useful, sometimes it needs fixing, and sometimes it opens another task. I would treat the suggestion that this sustains compulsive use as a hypothesis. This review does not establish a reinforcement schedule or a dopamine mechanism behind it.

The research question I would take forward is whether someone repeatedly continues despite consequences they recognise and want to avoid. Hours alone cannot answer that.

Choosing the extra work

I choose much of my workload. I would like that to settle the question of whether it is good for me. The evidence leaves more to consider.

Dettmers and Bredehöft followed a final sample of 236 employees over two survey waves, three months apart. Requirements to find new tasks and initiate projects predicted an increase in cognitive irritation, meaning persistent thoughts about work. Requirements to plan work and regulate effort did not show that longitudinal association. The sample consisted of graduates in full-time work with some control over their working arrangements. It was an observational study of job design, with no AI exposure measured C10.

A cross-sectional analysis of 2,582 employees at the Swedish Transport Administration found that greater control over working time was associated with less need for recovery. More after-hours work-related technology use and overtime were associated with greater need for recovery. The data support associations between these features of flexible work; they cannot establish the order of the proposed pathways C11.

My reading is that control and workload both need measuring. Freedom to choose the next task does not tell us how much work someone takes on, how often they return to it, or whether they recover afterwards.

For agent use, the question is whether saved effort becomes time away from work, a larger queue, or some mixture of the two. The studies here do not settle that. I know which answer I am inclined to give myself.

ADHD and hyperfocus

The ADHD evidence needs careful handling. A large sample can still answer a narrower question than the one we want to ask.

In an online survey of 16,426 Norwegian workers, 32.7% of those meeting the study’s workaholism cut-off also met an ADHD screening cut-off, compared with 12.7% of the other respondents. Those percentages describe ADHD screening results within the two workaholism groups. They are not the proportion of people with ADHD who are workaholics. The sample was self-selected, both measures were self-reported, and a positive screen is not a clinical diagnosis. The cross-sectional design cannot establish direction C12.

Hyperfocus also needs a definition. In a questionnaire validation study involving 347 adults, self-reported hyperfocus correlated with ADHD symptoms at r = 0.53, and with flow at r = 0.12. That supports treating the measured experiences as distinguishable, while allowing some overlap. It does not establish what happens during an agent session or whether that session contributes to exhaustion C13.

These findings justify asking whether people experience and regulate agent use differently. They cannot give us an ADHD-specific risk estimate for heavy agent use. They also cannot stand in for evidence about every form of neurodivergence.

What I would measure next

I would start with a prospective study that follows people while they use agents for their own work. This is a research proposal, not a result from the studies above.

The exposure measure would distinguish active human involvement from a job running in the background. I would record when someone submits work, checks a result, responds to it and starts again, alongside their intended working hours. A process running overnight should not automatically count as an overnight human work session.

I would pair those records with daily measures of detachment, sleep and next-day fatigue. The Recovery Experience Questionnaire provides an existing starting point for measuring detachment: its original validation, involving 930 participants, distinguished detachment, relaxation, mastery and control. I have checked that paper’s abstract. Any diary adaptation would still need its wording, recall period and measurement properties checked for the proposed use C14.

Workload, deadlines, control over working time and reasons for continuing would need recording too. Otherwise, a difficult week could produce both more agent use and worse sleep, leaving the tool to take the blame for the deadline. I would also measure work completed and effort saved, so the study could detect benefits alongside costs.

Following the same people across days would help distinguish a person’s unusually heavy-use days from differences between people. It would still leave questions about cause. A separate experiment could test a defined change, such as when results become available or notifications arrive. Longer follow-up would be needed to examine sustained exhaustion; next-morning tiredness would not settle a claim about burnout.

Where that leaves me

I have grounds to take recovery seriously. I also have evidence that particular AI tools can reduce exhaustion in particular settings. The studies reviewed here do not establish a safe amount of agent use or show that heavy, self-directed use causes burnout.

My working hypothesis is that the pattern matters: what the tool removes, what I add back, and whether work continues occupying the time I intended to leave free. That is something to test.

The usage bar cannot tell me whether I slept, switched off or finished the week in good shape. Those are the things I would want beside it.

Scope and source checks

This is a focused narrative review, checked on 16 September 2026. It draws on a supplied research pack containing 89 extracted records, followed by targeted checks of the primary sources used here. The pack reports searches through PubMed, OpenAlex and Crossref, with limited access to arXiv. I have not independently reproduced those searches. Computer-science and information-systems coverage may be incomplete, so the absence finding applies to this review’s evidence set.

Relevant full-text sections were checked for Liu, Savolainen, Wendsche and Lohmann-Haislah, Van Laethem, Kardefelt-Winther, Dettmers and Bredehöft, Edvinsson, Andreassen and Hupfeld. Afshar, Tang, Zhai, Barajeeh, and Sonnentag and Fritz are used at abstract depth. The Barajeeh source is a preprint. This review makes those different levels of access explicit because they limit what the claims can carry.