Cutting your mental load with AI: the personal-effectiveness method

By Paul-Antoine TUAL, AI Transformation Leader, Croissance et Transitions, August 2026.

Stance. This piece is a personal-effectiveness manual for the age of agents, and it opens with a correction. I thought AI was saving me time; the measurements say it lightens the effort while the clock barely moves. The whole method that follows is built on what the data actually support, including where they proved me wrong. I kept my practices, changed their justifications, and dropped two figures I had been citing in good faith for months.

Where the gain actually is

The subjective experience is unanimous among people who work seriously with agents: the day weighs less. The measurement is far less clear, and it has to be faced squarely before building a method on it.

The largest study available on the question was published in May 2026 by a team from Stanford, MIT and New York University, on 1,237 participants [1]. The protocol compares the time actually spent completing a task alone and with an assistant. Result: on simple tasks the assistant saves no time at all while still reducing perceived effort; it genuinely speeds up only the difficult tasks, three of them in the protocol. Participants, for their part, predicted the assistant would be markedly faster across the board, to the point that actual time exceeds their prediction by nearly a minute. The same bias disappears when they are asked to imagine help from another human: they then anticipate a modest saving, on the order of seventeen seconds, and land about right. The illusion is specific to the machine.

The METR lab produced the most rigorous measurement on the software side, and its story deserves telling in full rather than in extract. In July 2025, a randomised controlled trial on 16 experienced developers and 246 tasks, in repositories they had known for years, concludes at 19% more time with AI tools. Those same developers expected a 24% gain before the trial and still estimated, afterwards, that they had gained 20% [2]. Seven months later, METR publishes the sequel: 57 developers, 143 repositories, more than 800 tasks, and a result that flips the other way, an 18% gain for the original cohort. Except that the confidence interval runs from a 38% gain to a 9% loss, so it crosses zero, and METR spells out the reason for its own caution: “we have observed a significant increase in developers choosing not to participate in the study because they do not wish to work without AI” [3]. The lab has struck through its own 2025 result with a banner reading “these results are out of date”.

These measurements converge on a point that is not the expected one. Speed resists measurement: depending on the year, the protocol and the population, it rises, it falls or it does not move. The drop in perceived effort, by contrast, shows up in every study. For the leader of an SME or a mid-cap, the practical consequence is direct: stop promising shorter deadlines, promise lighter days, and organise yourself to bank that gain. It is a personal-effectiveness objective in the literal sense: what you recover is available judgement capacity, the exact resource a mid-sized company lacks most at the top.

The measurable gain is your mental load.

Why the gain evaporates without organisation

If the clock does not move while perceived effort drops, the effort has shifted to an activity that is less costly to live through. It has had a name on this site for a year: the verification tax. I wrote in Asynchronous work: everyone becomes a manager that everything an agent produces must be reviewed by someone who answers for it. Research adds a detail that changes practice: reviewing is work of a different nature, not a lighter version of the first. Writing mobilises production; verifying mobilises error detection inside syntactically flawless text, something the brain does badly because fluency serves it as a quality signal.

Hence the sense of ease. It is real and it is misleading: the fluency of an output is a poor indicator of its correctness. A leader who approves quickly because “it reads well” has just paid the verification tax in counterfeit currency.

The method that follows holds in three techniques, one per component of the load. The decision session attacks attention switches. Team management applied to agents attacks the verification tax. Channel tuning attacks interface friction. None requires a new tool; all require discipline.

Technique no. 1: the decision session

The principle is the heart of the method, and it opposes two working regimes: the decision session versus autonomous execution. In my own practice, the session takes the morning, 8 a.m. to noon. It concentrates everything that commits: judgement calls, architecture choices, plan approvals, anything irreversible. It produces written mandates. After that, and until the next day, no more decisions: agents execute autonomously, and I review everything in a single window at the end of the afternoon.

Before the mechanics, a necessary piece of honesty, because this is where I correct my own notes. I justified this morning window for years with two arguments I believed solid, and both are fragile. The first is decision fatigue, the idea that a stock of willpower depletes across the day’s calls. It comes from the ego depletion model, which a replication run in 2016 by 23 laboratories on 2,141 participants failed to find (d = 0.04), and which a second preregistered replication, 36 laboratories and 3,531 participants, published in 2021 in Psychological Science, failed to find again: d = 0.06, with the data four times more likely under the null hypothesis than under depletion [4]. The stock of willpower probably does not exist. The second is the morning cognitive peak: a study published in 2023 in Collabra: Psychology concludes, at the latent-variable level, that there is no general and robust synchrony effect between chronotype and time of day, and leans towards a methodological artefact [5].

The argument that holds is structural, and it is more than enough. In 2009, Sophie Leroy formalised attention residue: when you leave an unfinished task A for a task B, part of your attention stays anchored on A and degrades performance on B, the effect being strongest when A was under time pressure [6]. And in 2005, direct observation of 24 knowledge workers at the University of California, Irvine measured the real structure of an office day: 11 minutes and 4 seconds on average in a single working sphere before switching, 57% of spheres interrupted, and for work resumed the same day, 25 minutes and 26 seconds on average before returning to it, after passing through 2.26 other topics [7].

One detail is worth flagging in passing, because it illustrates what happens when you cite without checking. The famous figure of “23 minutes and 15 seconds to refocus”, repeated in hundreds of articles and slide decks, appears nowhere in that paper. I searched the original text line by line. The published measurement is 25 minutes and 26 seconds, with a standard deviation of 54 minutes and 48 seconds, that is, a spread twice the size of the mean. The to-the-second precision of the popular figure is exactly what should have raised suspicion.

The decision session therefore owes nothing to a brain that would be better at 9 a.m. Batching decisions removes switches, and every switch costs an attention residue plus a resumption delay whose variance is enormous. The morning has only one solid property: it is the part of the day you can protect, because the rest of the world has not caught up with you yet.

The session produces mandates, not tasks. That is what distinguishes it from a plain “deep work” block. A written mandate names the objective, the authorised scope, the limits, and above all the stop point before any irreversible action. It is the same requirement as the decision-complete plan described in We removed Git from a product in production: the document freezes the choices and leaves the mechanical residue to execution. Writing a mandate takes twenty minutes and avoids three days of misdirected work. And if your organisation’s bottleneck has moved from execution to decision, something I documented with figures in Decision-Gated Development, then the decision session is not a personal comfort: it is where all the value now flows.

Technique no. 2: in the afternoon, manage your agents like a team

The second technique follows from the first. Once the mandates are issued, someone has to make them live, and that someone has changed jobs without always noticing. Supervising agents that deliver asynchronously is team management: setting the intention, distributing the work, reviewing, answering for it. That is the thesis of the asynchronous work article, and it carries a reassuring consequence for a business leader: you already know how to do this. Everything you have learned about managing people transfers to managing agents, flaws included. Four moves concentrate the essentials.

The mandate is a job description. An agent without a clear mandate is an employee without a job description: you can neither steer it nor audit it. Objective, scope, limits, escalation cases. It looks bureaucratic; it is the opposite, because it is what lets you delegate more, in confidence, and state the instruction once rather than patching it turn by turn, the shift from conversation to structured command described in The end of prompt engineering.

Micro-management works no better with agents than with people. Reviewing as outputs trickle in, answering every notification from an agent that has finished something, is manufacturing precisely the switches the morning session has just removed. An agent notification that interrupts a contract review costs the same attention residue as a phone call. The afternoon is a decision-free zone: the agents run, you do something else, including deep work of your own.

The batched review is your management ritual. Everything produced is read in a single block, in one window. It is also the only moment when two outputs can be compared against each other, which is the best known antidote to the approval reflex: reading two versions forces a decision, reading one invites approval.

This ritual answers a precise risk, and this is where a word is needed about “Claude addiction”, since the term circulates. The occupational risk of the agent manager is not dependence, it is complacency. A scoping review published on 9 July 2026 in Frontiers in Psychology, covering 37 empirical studies that had passed peer review, concludes that most instruments meant to diagnose AI addiction “index frequency or preference (e.g., ‘I always prefer to use AI to find information’) rather than impairment, thereby conflating efficient, augmentative use with dependence” [8]. Using an assistant every day because it works describes an adopted tool, not a pathology. Complacency, by contrast, has been documented for a long time: the reference synthesis by Parasuraman and Manzey establishes that automation bias is observed among novices and experts alike, and that it is corrected neither by training nor by instruction [9]. Telling your teams to “stay vigilant” produces no measurable effect. Only process design does: a stop point, a comparison, a signature.

Keep your hand in the craft. A manager who can no longer do the work can no longer judge what they hand out. A randomised study published in the British Journal of Educational Technology in 2025 shows that generative assistance improves immediate performance while degrading self-regulation of thought, what the authors call “metacognitive laziness” [10]; the MIT Media Lab EEG study on “cognitive debt” points the same way, with the caveats it deserves, 54 participants, only 18 of them in the final session, and no peer review [11]. The remedy holds in one manager’s habit: at least one task a week done entirely without assistance. Not as asceticism. To preserve the internal yardstick that lets you judge other people’s work.

This technique has a limit I must name, the same as for a human team handed the signature. It assumes nothing irreversible can happen between two reviews. If an agent can send a client email, trigger a payment or modify a production database with no stop point, the review window arrives too late and the ritual protects nothing. That is the design constraint detailed in Engineering agentic systems: human in the loop. Autonomous execution holds only when the time to harm is longer than the time to control.

Technique no. 3: voice to emit, writing to approve

That leaves the most everyday friction, and the one on which everyone must find their own setting: through which channel to speak to the machine, and through which to receive its answer. The figures are more decisive than one might expect, provided the two directions are kept separate.

For emission, voice wins. The founding study on the subject, run at Stanford in 2016 by Sherry Ruan and her coauthors, measures voice entry 3.0 times faster than the keyboard in English, with an error rate 20.4% lower [12]. The measurement is old, and it covers miniature phone keyboards, which is not the situation of an executive at a full keyboard. A much more recent measurement widens the finding: across more than 1,000 clinicians in more than 60 facilities and 15 countries, median typing sits at 21.4 words per minute against 93 with automatic dictation, a factor of 4 to 5 [13]. Two reservations to state immediately: this work is a preprint that has not been peer reviewed, and the protocol has participants recite a 74-word passage already written. Transcribing a known text is not composing new thought, and dictating a vague intention produces text that needs reworking, which takes back with one hand the time saved with the other.

For reception, writing wins, and for a reason that is not throughput. Marc Brysbaert’s reference meta-analysis, aggregating 190 studies and 18,573 participants, puts silent reading of non-fiction at 238 words per minute against 183 for reading aloud [14]. The throughput gap is real but modest. What matters more is that text can be scanned, skimmed back over, and compared against another text open alongside, whereas an audio or video stream is strictly linear. A linear channel suits familiarisation and forbids control.

The setting rule that follows fits on one line. Dictate what you want, read what you must approve. Voice serves to get an intention out of your head while it is still warm, including while walking, including badly phrased, because the machine can tidy it up; it is the natural channel of the decision session, where mandates are emitted. Writing and diagrams serve the moment you commit your signature; they are the mandatory channel of the review window. As for generated video and audio, they have a real place, but downstream: broadcasting a decision already taken to a team, not taking it.

Three questions for your Monday morning

This method is not installed by changing tools; it is installed by answering three questions whose answers can be checked within a week.

How many times a day does your attention switch because an agent has finished something? If you do not know, count it for two days before changing anything.

What was the last AI output you approved without really reading it, and what made you believe it was good? If the answer is “it read well”, you now know the name of the phenomenon.

What can you still do without assistance, and when did you last do it?

The real gain from AI is not on the clock. It is in what you have left mentally available at 6 p.m. to settle the one thing nobody else can settle for you. Provided you have not spent it reviewing, as it trickled in, what you mandated in the morning.

To install this discipline (decision session, written mandates, batched review, stop points before the irreversible) in your organisation, see the Junyr Method™ and the free 30-minute audit.

Further reading

• The general frame: Asynchronous work: in the age of AI agents, everyone becomes a manager, the origin of the verification tax and the deskilling risk.
• The decision mechanics: Decision-Gated Development, on why the bottleneck moved from execution to decision, and We removed Git from a product in production, the plan registry that follows from it.
• The safety constraint: Engineering agentic systems: human in the loop, to size the stop points before irreversible actions.
• Writing the mandates: The end of prompt engineering, on the shift from conversation to structured command.

Paul-Antoine TUAL, AI Transformation Leader · Croissance et Transitions (SAS) · Junyr Method™ · Junyr Mail™ · Junyr Agents™

---

Sources

[1] Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins, “Cognitive offloading and the speedup illusion in human-AI interaction”, arXiv:2605.23177, 22 May 2026, accepted at the 48th Annual Meeting of the Cognitive Science Society. N = 1,237 participants (401 in the prediction sample, 836 in the completion sample). Assistance shortens time only on difficult tasks, three in the protocol, and not at all on simple ones, where it still reduces subjective effort; actual time exceeds predicted time by nearly a minute in the assisted condition; the bias is absent when the imagined help is human, participants then anticipating a saving of about seventeen seconds, close to reality. Affiliations: Stanford (Yu, Cheng, Jabbar, Jurafsky, Hawkins), New York University (Sucholutsky), MIT and Princeton AI Lab (Collins). A preprint accepted at a conference, to be distinguished from a journal article. https://arxiv.org/abs/2605.23177

[2] METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, 10 July 2025, arXiv:2507.09089. Randomised controlled trial, 16 experienced developers, 246 tasks, repositories of more than 22,000 stars to which they had contributed for years: 19% more time with AI tools, while they expected a 24% gain before the trial and still estimated a 20% gain afterwards. METR has since placed a warning on this page: “These results are out of date.” https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

[3] METR, “We are Changing our Developer Productivity Experiment Design”, 24 February 2026. Continuation of the study: 57 developers, 143 repositories, more than 800 tasks, including 10 developers from the original study. For those, an estimated 18% gain with a confidence interval running from a 38% gain to a 9% loss; for newly recruited developers, a 4% gain with an interval from a 15% gain to a 9% loss. Both intervals cross zero. Quotation used in the article: “We have observed a significant increase in developers choosing not to participate in the study because they do not wish to work without AI, which likely biases downwards our estimate of AI-assisted speedup.” https://metr.org/blog/2026-02-24-uplift-update/

[4] Kathleen D. Vohs, Brandon J. Schmeichel et al., “A Multisite Preregistered Paradigmatic Test of the Ego-Depletion Effect”, Psychological Science, vol. 32 no. 10, 2021, pp. 1566-1581: 36 laboratories, 3,531 participants, non-significant effect (d = 0.06), Bayesian analysis finding the data four times more likely under the null hypothesis. Follows the registered replication report by Martin Hagger et al. (2016), 23 laboratories, 2,141 participants, d = 0.04. https://journals.sagepub.com/doi/abs/10.1177/0956797621989733

[5] Alodie Rey-Mermet, Nicolas Rothen, “The Interplay of Time-of-day and Chronotype Results in No General and Robust Cognitive Boost”, Collabra: Psychology, vol. 9 no. 1, article 88337, 2023. No general and robust synchrony effect at the latent-variable level; the authors conclude a probable methodological artefact. https://online.ucpress.edu/collabra/article/9/1/88337/197502/

[6] Sophie Leroy, “Why is it so hard to do my work? The challenge of attention residue when switching between work tasks”, Organizational Behavior and Human Decision Processes, vol. 109 no. 2, 2009, pp. 168-181. A founding classic, cited as such. No percentage drop in performance is advanced in that paper: the “30 to 40%” figures circulating online are not in it and are not used here.

[7] Gloria Mark, Victor M. González, Justin Harris, “No Task Left Behind? Examining the Nature of Fragmented Work”, CHI 2005, University of California, Irvine. Direct observation of 24 knowledge workers: 11 minutes 4 seconds on average per working sphere before switching, 57% of spheres interrupted, 77.2% of interrupted work resumed the same day, on average 25 minutes 26 seconds later (standard deviation 54 minutes 48 seconds) and after 2.26 other spheres. Checked against the original text: the figure of “23 minutes and 15 seconds” widely attributed to this study does not appear in it. https://ics.uci.edu/~gmark/CHI2005.pdf

[8] Francesca Maria Dagnino, Chiara Fante, Vittorio Guerrieri, Marcello Passarelli, “Addicted, attached, or just delegating? A scoping review on ‘problematic artificial intelligence use’”, Frontiers in Psychology, 9 July 2026, DOI 10.3389/fpsyg.2026.1900953. Scoping review of 37 peer-reviewed empirical studies. Quotation used in the article: “instruments index frequency or preference […] rather than impairment, thereby conflating efficient, augmentative use with dependence.” https://doi.org/10.3389/fpsyg.2026.1900953

[9] Raja Parasuraman, Dietrich H. Manzey, “Complacency and Bias in Human Use of Automation: An Attentional Integration”, Human Factors, vol. 52 no. 3, 2010. Reference synthesis, cited here as a founding classic: automation complacency and automation bias are observed among novices and experts alike, and are prevented neither by training nor by instruction. https://journals.sagepub.com/doi/10.1177/0018720810376055

[10] Yizhou Fan, Luzhen Tang, Huixiao Le, Kejie Shen, Shufang Tan, Yuan Zhao et al., “Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance”, British Journal of Educational Technology, vol. 56 no. 2, 2025, pp. 489-530, DOI 10.1111/bjet.13544. Randomised experimental study comparing four support conditions on a writing task. https://bera-journals.onlinelibrary.wiley.com/doi/10.1111/bjet.13544

[11] Nataliya Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”, MIT Media Lab, 2025. 54 participants across sessions 1 to 3, only 18 of them in session 4; EEG, three groups (assistant, search engine, no tool). Not peer reviewed, small sample, single task in an educational setting: flagged here with those reservations. https://www.media.mit.edu/publications/your-brain-on-chatgpt/

[12] Sherry Ruan, Jacob O. Wobbrock, Kenny Liou, Andrew Ng, James Landay, “Speech Is 3x Faster than Typing for English and Mandarin Text Entry on Mobile Devices”, arXiv:1608.07323v1, 25 August 2016. Voice entry 3.0 times faster in English and 2.8 times faster in Mandarin than a miniature smartphone keyboard; error rate 20.4% lower in English and 63.4% lower in Mandarin. Deep Speech 2 recognition, iOS QWERTY and Pinyin keyboards. A founding classic of the field, cited as such: the measurement covers phone keyboards, not a full keyboard. Referencing note: version 2, filed in January 2018, carries a different title, “Comparing Speech and Keyboard Text Entry for Short Messages in Two Languages on Touchscreen Phones”. It is version 1 that is cited here. https://arxiv.org/abs/1608.07323v1

[13] “A multi-country study comparing typed to automatic speech recognition-based medical documentation speeds among Low- and Middle-Income Country Trained Clinicians”, medRxiv 2025.05.11.25327386, 13 May 2025. More than 1,000 clinicians and health workers, more than 60 facilities, more than 15 countries: median typing of 21.4 words per minute against 93 with automatic dictation, a factor of 4 to 5. Preprint, not peer reviewed; the protocol rests on reciting a standardised 74-word passage, which measures transcription rather than composition. https://www.medrxiv.org/content/10.1101/2025.05.11.25327386v1.full

[14] Marc Brysbaert, “How many words do we read per minute? A review and meta-analysis of reading rate”, Journal of Memory and Language, vol. 109, 2019. Meta-analysis of 190 studies and 18,573 participants: 238 words per minute for silent reading of non-fiction, 260 for fiction, and 183 words per minute for reading aloud (77 studies, 5,965 participants). The author notes that these values are lower than the figures commonly cited. https://biblio.ugent.be/publication/8647789