.png)
The Part of Your Role AI Will Never Replace: What the Research Says Generative AI Changes in Assistant and Chief of Staff Work, and What It Leaves to You.
Updated: Sep 28
It is twenty to eight, and the meeting summary is already in your inbox. The note-taking tool sent it at 6.02, before anyone was awake: seven tidy bullet points, three actions with owners, a neat line about next steps. It is accurate, as far as it goes. It does not mention that the finance director agreed to the budget far too quickly, that the chief executive's "let's park that" meant "I am not having this conversation in front of these people", or that the one action with no owner is the one that will be on the front page of the board pack in six weeks if nobody picks it up by Friday.
You know all of that. The tool does not, and there is a reasonable chance it never will. That gap, between what was said and what it meant, is where a great deal of this job has always lived. I wrote about the part of the role AI will never replace earlier this year, and I am still learning to embrace AI without making it my whole personality. This time I wanted to put the evidence underneath the argument: what the research actually says about which parts of assistant and chief of staff work generative AI changes, which it leaves alone, and what that means for how you shape the role from here. The field moves fast, so every study below comes with its date, and I have tried to say only what each one found.
What you will get out of this
Why the estimates of how many jobs AI will take range from 47% to under 2%, and which number is the useful one for you
What happened in the field experiments when people were given generative AI at work, who gained most, and the task where it made people worse
The evidence that skills fade when a tool does them for you, including a study of doctors whose own performance dropped after they got used to AI help
When people prefer a machine's judgement to a person's, when they prefer yours, and the meta-analysis of 163 studies that explains the difference
What a quiet penalty for using AI means for how you use it
The six kinds of human work the research keeps pointing back to, and why accountability is the one nobody talks about
A way to redesign your own role on purpose, using the four areas of the Irreplaceable Audit
The irreplaceable check: a tool that takes one piece of your work and tells you whether to hand it over, share it or keep it
The Irreplaceable Pages: three pages for sorting your week, checking what AI gives you, and having the conversation with your executive
How much of this job AI can reach, in numbers
Start with the headline figure, because you have almost certainly seen it. In a study first circulated in 2013 and published in 2017, two Oxford researchers estimated that about 47% of total US employment was at high risk of computerisation, and administrative support roles sat near the top of their list. That number did a great deal of travelling, and it frightened a lot of people who do exactly your job.
The same year, a second team went back and asked a different question. The 2017 estimate had judged whole occupations. When the researchers looked instead at the spread of tasks inside each occupation, because two people with the same job title rarely do the same work, the share of US jobs at high risk fell from 38% to 9%. Nothing about the technology had changed. The unit of analysis had.
That shift, from jobs to tasks, is the most useful idea in this whole literature, and it has held for more than twenty years. A 2003 study of the US labour market set out the pattern: computers substitute for routine tasks that follow explicit rules, and complement the non-routine ones that involve problem-solving and complex communication. A 2018 analysis applied a rubric for "suitability for machine learning" to 18,156 tasks and found that most occupations contain at least some suitable tasks, very few can be fully automated, and that getting the benefit usually means redesigning the job.
Generative AI changed which tasks are in reach, and the newer studies say so plainly. A 2024 analysis published in Science estimated that with the simple chat interfaces available at the time, about 1.8% of jobs could have more than half their tasks affected. Once you add the software being built around those models, that figure rose to a little over 46%. The International Labour Organization's global index, first published in 2023 and refined in May 2025, found that clerical occupations have the highest exposure of any group, in both versions. Globally, one worker in four is in an occupation with some exposure, and 3.3% of employment sits in the highest category. That share is higher for women than for men, 4.7% against 2.4%, and in high-income countries it reaches 9.6% of women's employment against 3.5% of men's, because so much clerical and administrative work is done by women. The ILO's own conclusion, in both editions, was that transformation of jobs is the most likely outcome, since most occupations consist of tasks that still need a person.
A 2025 study of 200,000 conversations with a workplace AI assistant, released as a preprint and not yet peer reviewed, came at it from the other direction, by looking at what people actually asked the tool to do. Office and administrative support came out among the occupation groups with the highest applicability, because so much of the work is gathering, writing and passing on information.
So the exposure is real, and it is higher for this profession than for most. Now look at what has happened so far.
A study of Danish workers in exposed occupations, linking surveys to official earnings and hours records and released as a working paper in 2025, found that most employers had adopted chatbot initiatives and workers reported productivity benefits. The effects on earnings and recorded hours were precise nulls, ruling out anything larger than 2% two years after the launch of ChatGPT. What had moved was the structure of the work: new tasks in content generation, AI oversight and AI integration. A field experiment across 66 firms and 7,137 knowledge workers, also a 2025 working paper, gave a random selection of people a generative AI tool built into their email, meetings and writing. The four in five who used it spent two fewer hours a week on email and did less work outside normal hours. Beyond those time savings, the researchers could not detect any change in the amount or mix of tasks people did.
There is a harder edge, and you deserve to see it. On a large freelancing platform, freelancers in the occupations most affected by generative AI lost work and earnings after ChatGPT arrived, and the most accomplished of them were, if anything, hit harder, a 2024 study found. A 2025 study of another platform found a 21% fall in job posts for writing and coding work within eight months, while the jobs that remained were more complex and better paid. Freelance writing is a different job from yours, and it is still a warning about the parts of any job that consist of producing a clean first draft to order.
My reading of all this, in the autumn of 2026, is that the tasks are moving faster than the jobs. That is exactly why the question worth asking is which of your tasks, rather than whether your role.
Try this. Take last week's diary and to-do list and put every recurring task into one of two columns: the ones where the output could be judged correct by somebody who does not know the people involved, and the ones where it could not. A formatted agenda goes in the first column. Knowing that the item on page two needs to move before the chair sees it goes in the second. The first column is where the exposure studies are pointing. The second is where you live.
The research behind this section
Frey and Osborne (2017) estimated the probability of computerisation for 702 occupations and concluded that about 47 per cent of total US employment was at high risk. Arntz, Gregory and Zierahn (2017) used detailed task data and showed that, taking into account the spectrum of tasks within occupations, the automation risk of US jobs dropped from 38 per cent to 9 per cent. Autor, Levy and Murnane (2003) showed that computer capital substitutes for workers in routine cognitive and manual tasks that follow explicit rules and complements workers in non-routine problem-solving and complex communication tasks. Autor (2015) argued that commentators overstate machine substitution and ignore the complementarities between automation and labour, and described the resulting polarisation of employment. Goos, Manning and Salomons (2014) documented job polarisation in 16 Western European countries between 1993 and 2010 and explained much of it through routine-biased technological change and offshoring. Brynjolfsson, Mitchell and Rock (2018) applied a rubric of suitability for machine learning to 18,156 tasks in the US occupational database and found that most occupations include at least some suitable tasks, few occupations are fully automatable, and realising the potential usually requires redesigning the task content of jobs. Felten, Raj and Seamans (2021) built and validated a measure of occupational exposure to AI. Eloundou and colleagues (2024) estimated that roughly 1.8 per cent of jobs could have over half their tasks affected by large language models with simple interfaces, rising to a little over 46 per cent when complementary software is taken into account. Gmyrek, Berg and Bescond (2023) predicted that the overwhelming effect of generative AI would be to augment occupations rather than automate them, with the greatest exposure in clerical occupations and highly gendered effects. Gmyrek and colleagues (2025) refined the index and found that clerical occupations continued to have the highest exposure, that one in four workers globally is in an occupation with some exposure, and that 3.3 per cent of global employment falls in the highest category (4.7 per cent of female and 2.4 per cent of male employment; 9.6 per cent against 3.5 per cent in high-income countries). Tomlinson and colleagues (2025), in a preprint, classified 200,000 conversations with Microsoft Bing Copilot and found the highest AI applicability scores for knowledge work groups including office and administrative support. Wang and colleagues (2025) used GPT-4 to evaluate 1,022 core tasks across 51 US public sector occupations and judged 63 per cent resistant to replacement, mainly because of physical presence, emotional intelligence and complex decision-making. Acemoglu and Restrepo (2019) set out how automation displaces labour from tasks while new tasks reinstate it. Acemoglu and colleagues (2022) found that AI-exposed US establishments reduced hiring in non-AI positions and changed the skill requirements of remaining postings between 2010 and 2018, while aggregate effects on employment and wages were too small to detect. Humlum and Vestergaard (2025, working paper) linked adoption surveys to Danish administrative records and estimated precise null effects of AI chatbots on earnings and recorded hours, ruling out effects larger than 2 per cent two years after ChatGPT's launch, alongside task reorganisation including new tasks in content generation, AI oversight and AI integration. Dillon and colleagues (2025, working paper) ran a six-month field experiment across 66 firms and 7,137 knowledge workers and found that the 80 per cent of treated workers who used an integrated generative AI tool spent two fewer hours a week on email and reduced their out-of-hours work, with no detectable shift in the quantity or composition of their tasks. Hui, Reshef and Zhou (2024) found that freelancers in highly affected occupations on a large online platform suffered reductions in employment and earnings after generative AI was released, with suggestive evidence that top freelancers were disproportionately affected. Demirci, Hannane and Zhu (2025) found a 21 per cent decrease in job posts for automation-prone writing and coding work within eight months of ChatGPT's introduction, with the remaining jobs more complex and better paid.
Who gains, and the jagged edge
The exposure studies tell you where AI could reach. The field experiments tell you what happened when real people were handed it and asked to work, and they are some of the most carefully designed studies in modern management research.
The first big one was published in 2023. Researchers gave 453 college-educated professionals writing tasks from their own occupations, the kind of press release, report or email that fills a working week, and randomly let half of them use ChatGPT. The average time taken fell by 40% and the quality of the output, judged blind, rose by 18%. The gap between the weakest and strongest writers narrowed.
A study of 5,172 customer-support agents, published in its final form in 2025, found the same shape in the real world. Access to an AI assistant raised the number of issues resolved per hour by 15% on average. The less experienced agents improved on both speed and quality. The most experienced and most skilled agents saw small gains in speed and small declines in quality. The researchers' reading was that the tool spread the know-how of the most able agents to everyone else, which helped the newcomers a great deal and did very little for the people who already had it.
Then came the study that gave this field its most useful phrase. In 2023, 758 consultants at a global consulting firm were given 18 realistic tasks, from creative to analytical. For the tasks inside what the researchers called the frontier of AI's capabilities, people using GPT-4 completed 12.2% more tasks, 25.1% faster, and at significantly higher quality. For one task deliberately chosen to sit a little outside the frontier, a business problem built so that a plausible-looking answer was the wrong one, the people using AI were 19% less likely to reach the correct answer than those working without it. The frontier was jagged: tasks that looked equally difficult fell on different sides of it, and nothing about the task told you in advance which side you were on.
The most dangerous task is the one where the tool sounds exactly as confident as it does on the tasks it gets right.
A meta-analysis published in 2024 pulled together 106 experiments and 370 effects in which humans alone, AI alone and humans working with AI had all been measured. On average, the human and AI combinations performed worse than the better of the two working alone. The losses came in tasks that involved making decisions. The gains came in tasks that involved creating content. And when the humans were better than the AI to start with, the combination improved on both; when the AI was better, adding a human made things worse.
There is a pattern in that, and it matters for how you use these tools. Drafting, summarising and creating were where people gained. Deciding, judging and choosing were where the combination went wrong, often because people accepted answers they should have questioned. A 2024 study of legal queries found that public AI models produced false information at least 58% of the time and often went along with users' incorrect assumptions. A 2025 study found that people overestimated how accurate a model's answers were when it gave its usual explanations, and that longer explanations made people more confident without making the answers any more accurate. An earlier study found that adding explanations increased the chance people would accept the AI's recommendation, whether or not it was right.
Two more findings belong here. A 2024 experiment found that writers given AI story ideas produced stories rated as more creative and better written, especially the less creative writers, and that the stories were more similar to each other than stories written without help. Each person gained, and the collection became more alike. And a 2025 study of 18,000 Danish workers in exposed occupations found that women were 16 percentage points less likely than men to have used ChatGPT for work, often held back by employer restrictions and a sense that they needed training first. In a profession where most people are women, that gap is worth closing on purpose.
Try this. Before you use AI on a task, ask which kind it is. If the job is to produce something, a first draft, a summary, a comparison, a tidy version of your own notes, it is probably inside the frontier and the gains are real. If the job is to decide something, who comes to the meeting, whether to escalate, which option to recommend, treat whatever the tool says as one more opinion to weigh, and decide first what would change your mind.
The research behind this section
Noy and Zhang (2023) ran a preregistered experiment in which 453 college-educated professionals completed occupation-specific, incentivised writing tasks and half were randomly given ChatGPT; average time fell by 40 per cent, output quality rose by 18 per cent, inequality between workers decreased, and exposed workers were twice as likely to report using it in their real job two weeks later. Brynjolfsson, Li and Raymond (2025) studied the staggered introduction of a generative AI assistant among 5,172 customer-support agents and found that it increased issues resolved per hour by 15 per cent on average, with less experienced and lower-skilled workers improving both speed and quality while the most experienced and highest-skilled workers saw small gains in speed and small declines in quality; the assistant also appeared to help workers learn. Dell'Acqua and colleagues (2023) ran a preregistered experiment with 758 consultants at Boston Consulting Group and found that, for 18 realistic tasks inside the frontier of AI capability, those using GPT-4 completed 12.2 per cent more tasks and completed them 25.1 per cent more quickly at significantly higher quality, while for a task selected to be outside the frontier they were 19 per cent less likely to produce correct solutions than those without AI. Vaccaro, Almaatouq and Malone (2024) meta-analysed 106 experimental studies and 370 effect sizes published between 2020 and mid-2023 and found that, on average, human and AI combinations performed significantly worse than the best of humans or AI alone (Hedges' g = -0.23), with losses in decision tasks, greater gains in content creation tasks, gains when humans outperformed AI alone and losses when AI outperformed humans alone. Doshi and Hauser (2024) found that access to AI story ideas made stories more creative, better written and more enjoyable, especially for less creative writers, and made them more similar to each other. Dahl and colleagues (2024) found that public large language models hallucinated at least 58 per cent of the time on legal queries, struggled to predict their own hallucinations and often uncritically accepted users' incorrect legal assumptions. Steyvers and colleagues (2025) found that users overestimated the accuracy of large language model answers when given default explanations, that longer explanations raised confidence without improving accuracy, and that explanations matched to the model's internal confidence narrowed the gap. Bansal and colleagues (2021) found that AI explanations did not increase complementary team performance and instead increased the chance that people accepted the AI's recommendation regardless of its correctness. Simkute and colleagues (2024) drew on human factors research to set out four reasons for productivity loss with generative AI: a shift in users' roles from production to evaluation, unhelpful restructuring of workflows, interruptions, and a tendency for automation to make easy tasks easier and hard tasks harder. Humlum and Vestergaard (2025) surveyed 18,000 workers in 11 exposed occupations linked to Danish register data and found that women were 16 percentage points less likely to have used ChatGPT for work, and that workers saw a substantial productivity potential but were often held back by employer restrictions and a perceived need for training.
The skill you stop practising
The field experiments measure what happens while the tool is switched on. A smaller and more recent set of studies asks what happens to you afterwards, and the answer should shape how you use it.
The worry is old. In 1983 an engineering psychologist wrote a short paper called "Ironies of automation", which is still cited constantly. Her point was that automating a system leaves the human operator with the tasks the designers could not automate, usually the rarest and hardest, while taking away the daily practice that kept the operator able to do them. Fourteen years later a review set out the ways people get automation wrong: using it too much, using it too little, and systems designed so that people become an afterthought. A 2010 review of automation bias and complacency found that both appear in novices and experts alike, cannot be trained or instructed away, and get worse when people have several things competing for their attention. That last condition describes most assistants' mornings.
The newer evidence comes with numbers. A 2025 study from four endoscopy centres in Poland looked at how well doctors detected precancerous growths in colonoscopies performed without AI, before and after the centres started using an AI detection tool on other procedures. In the three months before, the detection rate in the non-AI procedures was 28.4%. In the three months after, it was 22.4%. The study was observational rather than a trial, and the authors were careful to say that continuous exposure to AI might reduce the rate. It is still a striking result from experienced professionals doing their core work.
A field experiment published in 2025 gave nearly a thousand secondary school students access to GPT-4 while they practised maths. With a standard chat interface, practice grades went up by 48%. When access was taken away for the exam, those students did 17% worse than students who had never had it. A version of the tool designed to give hints rather than answers raised practice grades by 127% and largely removed the harm. The students with the plain version had used it as a crutch.
A 2025 set of four experiments with 3,562 people found the same shape at work. Collaborating with generative AI improved the task in front of people, and the improvement did not carry over to the next task they did alone. Moving from AI-assisted work back to solo work also brought a drop in intrinsic motivation and a rise in boredom.
A 2025 survey of 319 knowledge workers, who shared 936 examples of using generative AI at work, found that the more confident people were in the AI, the less critical thinking they reported doing, and the more confident they were in their own ability, the more they did. It also found that the nature of critical thinking shifted towards checking information, fitting the AI's response into the real situation, and taking stewardship of the task. A separate survey of 666 people found that frequent AI use was associated with weaker critical thinking, through offloading the thinking to the tool; that study was correlational, so it cannot say which came first.
The case study I find most useful for this profession comes from an accounting firm, published in 2023. Over several years, the firm's staff relied more and more on their software's automated functions. Complacency grew at both the individual and the organisational level: people paid less attention to what the system was doing, did less to keep their own competence up and checked its outputs less, and their skills eroded. The erosion could stay hidden, acknowledged by neither the staff nor the managers, while the people remained accountable for work they no longer fully understood. That is exactly the position the 1983 paper warned about.
The risk is less that the tool takes your job and more that it quietly takes the practice that made you good at it.
There is a counterweight, and it is a real one. The customer-support study found that AI assistance seemed to help agents learn, and the maths study found that a tool designed to protect learning mostly did. A 2021 experiment found that small design changes which made people think before seeing the AI's answer, such as committing to their own view first, significantly reduced over-reliance. People liked those versions least, which tells you something about why the habit needs deciding on rather than hoping for.
Try this. Pick the two or three skills that make you most valuable, perhaps reading a draft for what it leaves out, judging tone for a particular reader, or spotting the conflict in a diary, and decide to keep doing them first, by hand, before you look at what the tool suggests. Write your own three-line summary of the meeting before you open the automated one. Then compare. The comparison is where you learn, and it is the practice that keeps the skill.
The research behind this section
Bainbridge (1983) described the ironies of automation, in which automating a system leaves operators with the tasks designers could not automate while removing the practice that maintains their skill. Parasuraman and Riley (1997) set out the use, misuse, disuse and abuse of automation, including over-reliance that produces failures of monitoring and decision biases. Parasuraman and Manzey (2010) reviewed complacency and automation bias and concluded that both occur in naive and expert participants, that automation bias cannot be prevented by training or instructions, and that complacency arises under multiple-task load. Goddard, Roudsari and Wyatt (2012) systematically reviewed the frequency, mediators and mitigators of automation bias in clinical decision support. Budzyń and colleagues (2025) conducted a retrospective observational study at four Polish endoscopy centres and found that the adenoma detection rate of standard, non-AI-assisted colonoscopy fell from 28.4 per cent before AI implementation to 22.4 per cent after (odds ratio for AI exposure 0.69). Bastani and colleagues (2025) gave nearly a thousand high school maths students access to GPT-4 tutors and found that access improved practice performance (48 per cent for a standard interface, 127 per cent for a tutor designed with safeguards), but that when access was removed students who had used the standard interface performed 17 per cent worse than those who never had access, an effect largely mitigated by the safeguards. Wu and colleagues (2025) ran four experiments with 3,562 participants and found that collaborating with generative AI enhanced immediate task performance, that the augmentation did not persist in subsequent solo tasks, and that moving from collaboration to solo work decreased intrinsic motivation and increased boredom. Lee and colleagues (2025) surveyed 319 knowledge workers who shared 936 examples and found that higher confidence in generative AI was associated with less critical thinking and higher self-confidence with more, with critical thinking shifting towards information verification, response integration and task stewardship. Gerlich (2025) surveyed and interviewed 666 participants and found a significant negative correlation between frequent AI tool use and critical thinking, mediated by cognitive offloading. Rinta-Kahila and colleagues (2023) studied an accounting firm and modelled how reliance on automation fostered complacency at individual and organisational levels, weakening activity awareness, competence maintenance and output assessment, and producing skill erosion that neither workers nor managers acknowledged. Macnamara and colleagues (2024) set out how AI assistants might accelerate skill decay among experts and hinder skill acquisition among learners without either noticing. Natali and colleagues (2025) reviewed AI-induced deskilling and upskilling inhibition in medicine. Buçinca, Malaya and Gajos (2021) found in an experiment with 199 participants that cognitive forcing interventions significantly reduced over-reliance on AI compared with simple explanations, and that participants rated the most effective designs least favourably.
When people trust the machine, and when they want you
Whether a task moves to AI depends partly on what the tool can do and partly on whether the people you work with will accept its answer. The research on that second question has swung back and forth for a decade, and it has finally settled into something useful.
The baseline is uncomfortable for anyone who prides themselves on judgement. A 2000 meta-analysis of studies comparing human judgement with statistical formulas, across health and behaviour, found the formulas were about 10% more accurate on average, and substantially better in a third to nearly half of the studies. People were substantially better in only a small minority. A 2018 study of bail decisions estimated that a machine-learning tool could have cut crime by up to 24.7% with no change in how many people were jailed. Where there is a clear right answer and plenty of data, machines are often better than experts.
People did not always act as though they knew that. In a set of experiments published in 2015, participants who watched an algorithm make forecasts lost confidence in it faster than in a human who made the same mistake, and chose the human even after seeing the algorithm do better. The researchers called it algorithm aversion. A 2018 follow-up found a remedy that will ring true to anyone who has ever introduced a new system: people were far more willing to use an imperfect algorithm if they could adjust its answer, even slightly.
Then, in 2019, a set of six experiments found the reverse. Ordinary people followed advice more when they were told it came from an algorithm than from a person. The researchers called it algorithm appreciation. It faded in two situations: when people were choosing between the algorithm and their own view, and when they were experienced forecasters. The professionals relied on the algorithm less than lay people did, and it cost them accuracy.
So which is it? A meta-analysis published in 2025, covering 163 studies, 442 effects and 82,078 people, gave the best answer so far. People preferred AI when two things were true at once: they saw the AI as more capable than a person at the task, and they felt the decision did not need to be personalised to them. In that case there was a modest preference for the AI. When either condition failed, the preference flipped to the human, and more strongly.
The other studies fill in why personalisation matters so much. People trusted algorithms less for tasks that felt subjective rather than objective, according to studies with more than 1,400 participants in the lab and 56,000 in the field in 2019. In medicine, patients resisted AI because they worried it could not take account of their particular circumstances, a concern the researchers named uniqueness neglect; it disappeared when the AI was framed as supporting a human decision rather than replacing it. Across nine studies in 2018, people did not want machines making moral decisions about driving, law, medicine or war, largely because they felt machines could neither fully think nor feel. People also rejected algorithms more in domains that felt inherently uncertain, across nine studies and 4,820 people in 2020, and relied on them more as a task became harder and had a clear right answer.
The newest study in this line is my favourite, because it is about exactly the kind of knowledge assistants carry. A 2026 study of investment experts found that AI advice changed their evaluations when it recommended against an investment and had no effect when it recommended for one. The experts believed the AI was good with explicit, written-down information and poor with tacit knowledge, such as the quality of a founding team, and a positive recommendation sent them back to look at exactly that.
There is one more twist. A 2025 set of five experiments found that ChatGPT gave better personal advice than the average participant, even about dating and relationships, where people said they strongly preferred human advice. The bias against it appeared only when people knew where the advice had come from.
Try this. When you bring AI-assisted work to your executive, notice which quadrant it sits in. If it is objective and generic, such as a comparison of three venues against a fixed brief, say how you produced it; people will accept it readily. If it needs to fit a particular person or situation, lead with your judgement and treat the tool as the preparation behind it. Being able to adjust the answer is what made people comfortable with imperfect algorithms, and being the person who adjusts it is a large part of your value.
The research behind this section
Grove and colleagues (2000) meta-analysed studies comparing clinical and mechanical prediction of human health and behaviour and found mechanical prediction about 10 per cent more accurate on average, substantially outperforming clinical prediction in 33 to 47 per cent of studies, with clinical prediction substantially more accurate in only 6 to 16 per cent. Kleinberg and colleagues (2018) used bail decisions to show that a machine learning tool could have reduced crime by up to 24.7 per cent with no change in jailing rates, or reduced jailing by up to 41.9 per cent with no increase in crime. Dietvorst, Simmons and Massey (2015) found in five studies that people lost confidence more quickly in algorithmic than human forecasters after seeing them make the same mistake, and were less likely to choose an algorithm over an inferior human even after seeing it outperform. Dietvorst, Simmons and Massey (2018) found that people were considerably more likely to use an imperfect algorithm when they could modify its forecasts, even when the modifications allowed were severely restricted, and performed better as a result. Logg, Minson and Moore (2019) found in six experiments that lay people adhered more to advice when they thought it came from an algorithm than from a person, and that this algorithm appreciation waned when people chose between an algorithm and their own estimate and among experienced forecasters, who relied less on algorithmic advice and were less accurate as a result. Prahl and Van Swol (2017) did not replicate a general preference for human advice, and found that after bad advice, use of automated advice decreased significantly more than use of human advice. Castelo, Bos and Lehmann (2019) found in four lab studies with over 1,400 participants and two field studies with over 56,000 that algorithms were trusted and relied on less for tasks that seemed subjective. Longoni, Bonezzi and Morewedge (2019) found that resistance to medical AI was driven by uniqueness neglect and was eliminated when AI care was framed as personalised or as supporting rather than replacing a human provider's decision. Bigman and Gray (2018) found across nine studies that people were averse to machines making moral decisions in driving, legal, medical and military settings, mediated by the perception that machines cannot fully think or feel, and that the aversion was difficult to eliminate. Dietvorst and Bharti (2020) found in nine studies with 4,820 participants that people were less likely to use even the best possible algorithm in more unpredictable decision domains. Bogert, Schecter and Watson (2021) found in three preregistered experiments that people relied more on algorithmic advice relative to social influence as tasks became more difficult. Qin and colleagues (2025) meta-analysed 442 effect sizes from 163 studies with 82,078 participants and found AI appreciation (d = 0.27) when AI was perceived as more capable than humans and personalisation was perceived as unnecessary, and AI aversion (d = -0.50) otherwise. Jussupow, Benbasat and Heinzl (2024) proposed an integrative classification of algorithm aversion and appreciation. Hou and Jung (2021) produced both appreciation and aversion by changing only how the human and algorithmic agents were described, mediated by perceived competence. Basu and colleagues (2026) found in two experiments, a field study and interviews that AI advice influenced investment experts' evaluations when it recommended against investing but not when it recommended investing, because experts believed AI could handle explicit but not tacit knowledge. Osborne and Bailey (2025) found in five preregistered experiments with 1,722 participants that ChatGPT produced better personal advice than the average participant, that a bias against it appeared only when people knew the source, and that seeing ChatGPT's advice before writing their own lowered people's ratings of their own authenticity.
The human work that remains
Put the studies side by side and the same six kinds of work keep coming back as the parts AI changes least. None of them is a surprise to anyone who has done this job well. What the research adds is why each one stays, and in two cases, a warning that it is less safe than we might like.
Judgement. A 2019 paper by three economists offered the neatest way to think about it: AI makes prediction cheap, and when prediction becomes cheap, the value of judgement, deciding what matters and what the trade-offs are worth, goes up. The consultants' study is the proof of the other side, that the tool cannot tell you when you have wandered outside what it knows. As I put it to myself: a calendar full of meetings is data; understanding that the leadership team has no space to think is judgement. A list of delayed actions is data; understanding that accountability is unclear is judgement.
Tacit knowledge. A 2020 study of patent examiners found that machine learning was biased by incomplete inputs, and that the examiners' domain expertise was what corrected it. A 2022 field study in a major US hospital found that radiologists in only one of three departments made real use of AI, and they were the ones who had developed ways of interrogating its results against their own knowledge. The investment study above found experts discounting AI precisely where tacit knowledge mattered. The unwritten version of the organisation, the meeting before the meeting and the stakeholder who needs early warning, is tacit knowledge, and it is yours.
Relationships. Here the evidence is more complicated than the comforting version, and you deserve both halves. In a 2023 study, evaluators preferred an AI chatbot's answers to patients' questions over doctors' answers in 78.6% of cases, and rated them higher on empathy. In 2024, AI-written responses made people feel more heard than human ones did, and the effect shrank when people were told the response came from AI. A 2025 study found that responses people believed came from a human were rated as more empathic, especially when they conveyed shared experience and care. So AI can produce the words of care, sometimes better than a tired human; what people value is care they believe comes from a person who chose to give it.
Trust. People who used AI-generated smart replies in conversation were seen as closer and more cooperative, and seen more negatively if the other person suspected they were using them, according to a pair of experiments in 2023. In 2019, people who saw a mix of profiles on a rental platform found hosts less trustworthy when they suspected their profiles had been written by AI. And a 2025 set of four experiments with 4,439 people found that people who used AI at work expected to be judged as less competent and less motivated, and were. Trust in a close working relationship rests partly on believing the effort is real.
Discretion. A system can store information. A person earns trust through how they handle it, knowing when details should be shared, softened, challenged or held back. The research on how employees come to trust AI tools found three kinds of trust at work, cognitive, emotional and organisational, and that the emotional and organisational kinds mattered alongside whether the tool gave good answers. What your executive tells you in the car is protected by your judgement about who else should hear it, and no tool has been asked to carry that.
Accountability. This is the one I would underline. A 2022 experiment found that decision-makers who delegated a harmful decision to a machine were punished less than those who delegated it to a person. A 2020 study found that when an AI-assisted conversation went wrong, the AI absorbed some of the blame, acting as what researchers call a moral crumple zone. A 2021 analysis named four separate responsibility gaps that open up when AI is involved in decisions. And a 2022 review of 41 policies requiring human oversight of government algorithms found that people were generally unable to perform the oversight the policies assumed. The pull, in other words, is towards letting the tool be responsible. Somebody still has to be, and in executive support that somebody has always been a person.
There is a real disagreement underneath all this, and it concerns you directly. A 2017 study of the US labour market found that jobs requiring high levels of social interaction grew by nearly 12 percentage points as a share of employment between 1980 and 2012, and that the return to social skills rose. That is the basis of the hopeful argument that AI will make human skills more valuable. But a 2025 analysis of seven million job postings at 595 firms, released as a preprint, found that in roles adopting generative AI, the relative demand for social skills fell by 3.4%. The reason was that functional administrative skills were being added to roles outside administration, which the authors read as AI-enabled self-service. Put plainly, some managers are starting to do their own admin with AI. My reading is that both are true at once: routine coordination is becoming self-service, and the coordination that needs judgement, trust and a person's name on it is becoming more valuable. Which of the two describes your role depends a great deal on how you shape it.
One last finding, and a human one. A 2023 set of four studies with 794 employees in four regions found that the more people worked with AI systems, the more they felt a need for social connection, which led to more helping of colleagues, and the lonelier they felt, which was linked to worse sleep and more drinking after work. As the tools take over more of the solitary work, the connecting part of your job, the part that holds people together, matters more to them as well as to you.
Try this. For each of the six, write one example from the last month where it was you, specifically, who made the difference: the judgement call, the unwritten thing you knew, the relationship you mended, the confidence you kept, the thing you held back, the decision you took responsibility for. Keep the list. It is the evidence for the conversation in the last section.
The research behind this section
Agrawal, Gans and Goldfarb (2019) argued that AI reduces the cost of prediction and so raises the value of complementary human judgement. Jarrahi (2018) argued that AI can extend human cognition in handling complexity while humans offer a more holistic, intuitive approach to uncertainty and equivocality in organisational decisions. Choudhury, Starr and Agarwal (2020) found in patent examination that machine learning was biased by incomplete inputs and that users' domain expertise was needed to find the most relevant prior art. Lebovitz, Lifshitz-Assaf and Levina (2022) found in a major US hospital that radiologists in only one of three departments consistently incorporated opaque AI results into their judgements, through practices of interrogating the AI against their own knowledge. Ayers and colleagues (2023) found that licensed health care professionals preferred ChatGPT's responses to 195 patient questions over physicians' responses in 78.6 per cent of 585 evaluations, rating them higher in quality and empathy. Yin, Jia and Wakslak (2024) found that AI-generated messages made recipients feel more heard than human-generated messages, and that recipients felt less heard when they realised a message came from AI. Rubin and colleagues (2025) found that human-attributed responses were rated as more empathic than AI-attributed ones, especially when conveying shared experience and care. Hancock, Naaman and Levy (2020) defined AI-mediated communication and set out its research agenda. Hohenstein and colleagues (2023) found in two randomised experiments that algorithmic response suggestions increased communication speed and positive language and led partners to evaluate each other as closer and more cooperative, but that people were evaluated more negatively if suspected of using them. Jakesch and colleagues (2019) found that people mistrusted hosts whose profiles were labelled as or suspected to be written by AI when they saw a mix of AI- and human-written profiles. Reif, Larrick and Soll (2025) found in four preregistered experiments with 4,439 participants that people who use AI at work anticipate and receive negative evaluations of their competence and motivation, with effects on assessments of job candidates. Glikson and Woolley (2020) reviewed empirical research on human trust in AI. Gkinko and Elbanna (2023) found in a case study of a global organisation that employees' trust in an AI chatbot was emotional, cognitive and organisational, and that emotional and organisational trust complemented cognitive trust. Feier, Gogoll and Uhl (2022) found in a laboratory experiment that decision-makers could avoid punishment more easily by delegating harmful decisions to machines than to other people. Hohenstein and Jung (2020) found that AI-generated smart replies increased perceived trust between communicators and that when conversations went wrong the AI was perceived as a coercive agent, acting as a moral crumple zone. Santoni de Sio and Mecacci (2021) set out four responsibility gaps with AI. Green (2022) surveyed 41 policies prescribing human oversight of government algorithms and found evidence that people are unable to perform the oversight functions these policies assume. Deming (2017) found that jobs requiring high levels of social interaction grew by nearly 12 percentage points as a share of the US labour force between 1980 and 2012, and that the labour market return to social skills rose. Gulati and colleagues (2025), in a preprint analysing seven million job postings from 595 US public firms, found that relative demand for social skills in generative AI-adopting roles declined by 3.4 per cent, reflecting the addition of functional administrative skills to roles outside those functions. Tang and colleagues (2023) found across four studies with 794 employees in Taiwan, Indonesia, the United States and Malaysia that interacting with AI increased the need for social affiliation, which led to more helping, and loneliness, which was linked to insomnia and alcohol consumption after work.
Redesigning the role on purpose
If there is one message running through all of this research, it is that the outcome depends less on the technology than on how the work is designed around it, and in most organisations that design is currently happening by accident, one tool and one shortcut at a time.
The research on work design is clear that this is a choice. A 2022 review argued that new technologies can raise or lower the things that make a job good, autonomy, the use of skill, feedback and relationships, depending on the choices made when they are introduced, and that work design has to be considered from the start rather than repaired afterwards. A 2021 paper described the automation and augmentation paradox: organisations that only automate risk losing the human expertise they will need later, and those that only augment never gain the efficiency, so the two have to be managed together. A 2026 study of how to divide work between people and AI found that the best arrangement had a consistent shape: AI working alone on relatively easy tasks, AI and humans working together on tasks where they performed similarly, and humans working without AI on the hardest ones. A 2022 study found the catch: people were poor judges of what they knew and did not know, and so poor at deciding what to hand to AI. The AI was better at deciding what to hand to people.
The time a tool saves you is not a gift until you decide where it goes.
That last finding is why a structure helps. The one I use, and the heart of the Irreplaceable Audit, looks at a role through four areas.
Task-based work is repeatable, structured and easy to document: routine notes, formatting, simple drafts, trackers, standard agendas. This is where the exposure studies and the field experiments agree that AI helps most, and where the time savings are real. The question to ask: where am I spending skilled human energy on work a tool could support?
Judgement-based work depends on context, consequence and experience: what to escalate, what to challenge, what to leave alone, when speed matters and when accuracy matters more. This is where human and AI combinations went wrong most often, and where you should decide first and consult the tool second. The question: where does my judgement prevent problems, improve decisions or protect leadership focus?
Relationship-based work depends on trust, emotional intelligence and communication: stakeholder expectations, sensitive conversations, working styles, the emotional temperature around a piece of work. AI can help you prepare. The research on AI-mediated communication says the relationship itself needs to be visibly yours. The question: which relationships work better because I understand the people, the history and the context?
Strategy-based work connects activity to outcomes: operating rhythms, better decisions, tracking priorities, spotting gaps, connecting work across teams. This is the area the time saved should flow into, deliberately, or it will be swallowed by more of the first. The question: where am I helping the business move forward, rather than only helping work get done?
Job crafting research, going back to a 2001 paper, shows that people reshape their jobs all the time by changing their tasks, their relationships and how they think about the work. A 2025 study of entry-level consultants found them doing exactly this with generative AI, and identified a third form, which the authors called signal crafting: using the tool in ways that shape how others see the distinctiveness and value of your work. That is a useful name for something assistants need to do now. The language matters. "I manage the calendar" describes a task a tool can do. "I protect the chief executive's time for the three decisions that matter this quarter" describes judgement. I have written about naming your value elsewhere, so I will say only that AI makes the difference between those two sentences more consequential than it has ever been.
Try this. Take one week and sort everything you did into the four areas. For the task-based work, mark each item automate (the tool does it, you check it), augment (you do it together) or keep. Then write one sentence for your executive about where the time you save will go. The sentence is the redesign; everything else is housekeeping.
The research behind this section
Parker and Grote (2022) argued that digital technologies can positively or negatively affect job resources such as autonomy, skill use, feedback and relationships, and job demands such as performance monitoring, depending on work design choices, and proposed that work design be considered proactively during technology implementation. Raisch and Krakowski (2021) set out the automation-augmentation paradox in management, arguing that the two are interdependent and must be managed together. Fügener, Walzner and Gupta (2026) modelled the allocation of judgement tasks between humans and AI and found, in experimental data, that optimal configurations had AI automating relatively easy tasks, AI augmenting humans on tasks with similar human and AI performance, and humans working without AI on relatively difficult tasks. Fügener and colleagues (2022) found that AI was good at delegating work to humans because it knew what it knew, while humans were poor judges of their own metaknowledge and poor at delegating to AI. Wrzesniewski and Dutton (2001) described job crafting as the physical and cognitive changes individuals make to the task and relational boundaries of their work. Mayer, Baygi and Buwalda (2025) found in a qualitative study of entry-level consultants that they used generative AI to recraft their tasks and relationships, and identified signal crafting, using AI to shape others' perceptions of the distinctiveness and value of their work. Willcocks (2020) argued that claims of net job loss are exaggerated but that there will be considerable skills disruption, and that whole jobs are rarely automated. Bankins, Hu and Yuan (2024) reviewed how AI both automates and complements human labour and creates new forms of work, and concluded that AI literacy and human-centred skills will become more important. Bankins and colleagues (2024) systematically reviewed empirical research on AI at work under five themes, including human-AI collaboration and perceptions of algorithmic and human capabilities. Babashahi and colleagues (2024) reviewed 20 studies on AI and skill transformation across industries. Kolomaznik and colleagues (2024) reviewed the role of trust, empathy, rapport and engagement in human-AI collaboration. Anthony, Bechky and Fayard (2023) proposed treating AI as a counterpart in a system of work that includes its design, implementation and use. Sarbani and Hadisoebroto (2025) surveyed 26 Indonesian administrative and secretarial professionals and found them using ChatGPT, Grammarly and Notion AI for document writing, scheduling and information management, with concerns about accuracy, digital literacy and training.
The irreplaceable check
Every study in this post comes down to one small decision you make many times a week: this piece of work, this tool, how much do I hand over? So here is a tool for that moment.
Name one piece of your work and answer eight questions about it: how structured it is, how much it depends on a particular person, what happens if it is wrong, who carries the responsibility, and how often you do it yourself. It will tell you whether the research points towards handing it to the tool with a check, working on it together, or keeping it yours, which of the four areas it belongs to, what to watch for, and it gives you a note to copy.
Where to start
If you take one thing from all of this, make it the difference between producing and deciding. The field experiments, the meta-analyses and the deskilling studies all point the same way: generative AI is good at producing drafts, summaries and first versions, and human and AI combinations go wrong when the task is a decision. Let the tool produce. Keep the deciding, and keep practising it.
After that, there are three moves worth making.
Sort a week of your work into the four areas and mark the task-based items automate, augment or keep. Protect the two or three skills that make you valuable by doing them first, by hand, before you look at what the tool suggests. And have one conversation with your executive about where the time you save will go, in the language of judgement rather than tasks.
I have put the three pages I would use into a download: a week sort for the four areas, a checking routine for anything AI has touched before it leaves your desk, and a page for the redesign conversation with your executive.
If AI has made you pause and wonder what the future of your role looks like, you are not behind. You are paying attention, and in roles like ours, paying attention has always been the work. The research is, on balance, kinder to this profession than the headlines: the tasks are changing fast, the need for somebody who understands people, pressure, timing and context has not gone anywhere, and the choices about how the role is redesigned are still, for now, largely in your hands.
The part of this that shows up in how you are seen at work sits inside what the Connected Leader Assessment calls Judgement and Decision-Making: knowing what to decide yourself and what to delegate, weighing advice from people and tools against what you know about the situation, and taking responsibility for the call. It is the dimension that decides whether AI makes you faster or makes you replaceable.
The assessment is free, takes about twelve minutes and gives you a read across all five dimensions. If you want to go further on this one, the Skill Sprints are short, focused programmes on a single dimension, and the Irreplaceable Audit maps your role across task, judgement, relationship and strategy work, so you can see which parts a tool could support and which parts of your contribution deserve to be seen.
Stay savvy.
Meg ✌️
Free tools
The irreplaceable check, the interactive tool in this post, free
The Irreplaceable Pages, the download in this post, free
The Connected Leader Assessment, twelve minutes, free, with a read on Judgement and Decision-Making, the dimension this post sits in
The Irreplaceable Audit, forty-eight activities across four kinds of work, free, for sorting what a system could do from what only you can
P.S. Want to Go Deeper on AI and Your Role?
If this resonated and you want to go further, here is the evidence the piece is built on. These are the papers that shaped my thinking on what generative AI changes in this work, what it leaves to people, and how to redesign the role on purpose, and they are well worth your time.
Every DOI below has been checked. Where I could not confirm a figure from the paper itself, I have left the figure out rather than guess.
📃 Eloundou and colleagues (2024) About 1.8% of jobs had over half their tasks affected with simple tools; a little over 46% with added software. https://doi.org/10.1126/science.adj0998
📃 Gmyrek, Berg and Bescond (2023) Generative AI will mostly augment jobs; clerical work is most exposed, and the effects are gendered. https://doi.org/10.54394/fhem8239
📃 Gmyrek and colleagues (2025) Refined index. Clerical work still most exposed; one worker in four has some exposure; women more than men. https://doi.org/10.54394/hetp0387
📃 Felten, Raj and Seamans (2021) A validated measure of how exposed each occupation is to AI. https://doi.org/10.1002/smj.3286
📃 Frey and Osborne (2017) 702 occupations. About 47% of US employment judged at high risk of computerisation. https://doi.org/10.1016/j.techfore.2016.08.019
📃 Arntz, Gregory and Zierahn (2017) Counting the tasks inside jobs cut the share of US jobs at high risk from 38% to 9%. https://doi.org/10.1016/j.econlet.2017.07.001
📃 Autor, Levy and Murnane (2003) Computers replace routine, rule-based tasks and complement problem-solving and complex communication. https://doi.org/10.1162/003355303322552801
📃 Autor (2015) Why automation has not wiped out most jobs: it complements labour as well as replacing it. https://doi.org/10.1257/jep.29.3.3
📃 Acemoglu and Restrepo (2019) Automation displaces people from tasks; new tasks bring them back. https://doi.org/10.1257/jep.33.2.3
📃 Brynjolfsson, Mitchell and Rock (2018) 18,156 tasks. Most jobs have some tasks machines can learn; few can be fully automated; jobs need redesign. https://doi.org/10.1257/pandp.20181019
📃 Noy and Zhang (2023) 453 professionals. ChatGPT cut writing time by 40% and raised quality by 18%. https://doi.org/10.1126/science.adh2586
📃 Brynjolfsson, Li and Raymond (2025) 5,172 support agents. AI raised productivity 15%; novices gained most, experts slightly lost quality. https://doi.org/10.1093/qje/qjae044
📃 Dell'Acqua and colleagues (2023) 758 consultants. Faster and better inside the frontier; 19% less likely to be right outside it. https://doi.org/10.2139/ssrn.4573321
📃 Humlum and Vestergaard (2025) Working paper. Danish records: no effect on earnings or hours beyond 2%, but new tasks and reorganised work. https://doi.org/10.3386/w33777
📃 Doshi and Hauser (2024) AI ideas made individual stories better and the whole collection more alike. https://doi.org/10.1126/sciadv.adn5290
📃 Budzyń and colleagues (2025) Doctors' detection rate without AI fell from 28.4% to 22.4% after they began using AI. https://doi.org/10.1016/s2468-1253(25)00133-5
📃 Bastani and colleagues (2025) Nearly 1,000 students. Unguarded GPT-4 raised practice grades, then cut exam grades by 17%. https://doi.org/10.1073/pnas.2422633122
📃 Lee and colleagues (2025) 319 knowledge workers. More confidence in AI went with less critical thinking. https://doi.org/10.1145/3706598.3713778
📃 Gerlich (2025) 666 people. Frequent AI use correlated with weaker critical thinking, through offloading. https://doi.org/10.3390/soc15010006
📃 Parasuraman and Manzey (2010) Automation bias and complacency hit experts as well as novices and resist training. https://doi.org/10.1177/0018720810376055
📃 Goddard, Roudsari and Wyatt (2012) Systematic review of automation bias in clinical decision support. https://doi.org/10.1136/amiajnl-2011-000089
📃 Parasuraman and Riley (1997) How people use, misuse, disuse and abuse automation. https://doi.org/10.1518/001872097778543886
📃 Dietvorst, Simmons and Massey (2015) People abandoned algorithms faster than humans after seeing the same mistake. https://doi.org/10.1037/xge0000033
📃 Dietvorst, Simmons and Massey (2018) People used imperfect algorithms when they could adjust the answer, even slightly. https://doi.org/10.1287/mnsc.2016.2643
📃 Logg, Minson and Moore (2019) Six experiments. Lay people preferred algorithmic advice; experienced forecasters did not, and lost accuracy. https://doi.org/10.1016/j.obhdp.2018.12.005
📃 Castelo, Bos and Lehmann (2019) Algorithms were trusted less for tasks that seemed subjective. https://doi.org/10.1177/0022243719851788
📃 Longoni, Bonezzi and Morewedge (2019) Patients resisted medical AI for fear it would miss what made them unique, unless it supported a human. https://doi.org/10.1093/jcr/ucz013
📃 Bigman and Gray (2018) Nine studies. People did not want machines making moral decisions. https://doi.org/10.1016/j.cognition.2018.08.003
📃 Glikson and Woolley (2020) Review of empirical research on how people come to trust AI. https://doi.org/10.5465/annals.2018.0057
📃 Yin, Jia and Wakslak (2024) AI responses made people feel more heard, until they learned the response was AI. https://doi.org/10.1073/pnas.2319112121
📃 Ayers and colleagues (2023) Evaluators preferred chatbot answers to patient questions over doctors' in 78.6% of cases. https://doi.org/10.1001/jamainternmed.2023.1838
📃 Rubin and colleagues (2025) Empathy people believed came from a human was valued more than the same empathy from AI. https://doi.org/10.1038/s41562-025-02247-w
📃 Reif, Larrick and Soll (2025) 4,439 people. Those who used AI at work were judged less competent and less motivated. https://doi.org/10.1073/pnas.2426766122
📃 Hancock, Naaman and Levy (2020) Defining AI-mediated communication and the questions it raises. https://doi.org/10.1093/jcmc/zmz022
📃 Jakesch and colleagues (2019) People mistrusted hosts whose profiles they suspected were written by AI. https://doi.org/10.1145/3290605.3300469
📃 Hohenstein and Jung (2020) When AI-assisted conversations went wrong, the AI absorbed blame: a moral crumple zone. https://doi.org/10.1016/j.chb.2019.106190
📃 Hohenstein and colleagues (2023) Smart replies made partners seem closer, and users were judged worse if suspected of using them. https://doi.org/10.1038/s41598-023-30938-9
📃 Raisch and Krakowski (2021) The automation-augmentation paradox: the two have to be managed together. https://doi.org/10.5465/amr.2018.0072
📃 Jarrahi (2018) AI for complexity, people for uncertainty and ambiguity in organisational decisions. https://doi.org/10.1016/j.bushor.2018.03.007
📃 Fügener and colleagues (2022) People were poor at deciding what to hand to AI; AI was good at deciding what to hand to people. https://doi.org/10.1287/isre.2021.1079
📃 Vaccaro, Almaatouq and Malone (2024) 106 experiments. Human and AI together did worse than the better alone, especially on decisions. https://doi.org/10.1038/s41562-024-02024-1
📃 Wrzesniewski and Dutton (2001) Job crafting: how people reshape the tasks, relationships and meaning of their work. https://doi.org/10.5465/amr.2001.4378011
📃 Parker and Grote (2022) Why work design decides whether new technology makes jobs better or worse. https://doi.org/10.1111/apps.12241
📃 Goos, Manning and Salomons (2014) Job polarisation in 16 European countries, 1993 to 2010, from routine-biased technology and offshoring. https://doi.org/10.1257/aer.104.8.2509
📃 Deming (2017) Jobs needing high social interaction grew by nearly 12 points of US employment, 1980 to 2012. https://doi.org/10.1093/qje/qjx022
📃 Humlum and Vestergaard (2025) 18,000 Danish workers. Women were 16 points less likely to have used ChatGPT for work. https://doi.org/10.1073/pnas.2414972121
📃 Anthony, Bechky and Fayard (2023) Treating AI as a counterpart in a whole system of work. https://doi.org/10.1287/orsc.2022.1651
📃 Lebovitz, Lifshitz-Assaf and Levina (2022) Radiologists used opaque AI well only where they had practices for interrogating it. https://doi.org/10.1287/orsc.2021.1549
📃 Dahl and colleagues (2024) Public AI models gave false legal information at least 58% of the time. https://doi.org/10.1093/jla/laae003
📃 Agrawal, Gans and Goldfarb (2019) AI makes prediction cheap, which raises the value of human judgement. https://doi.org/10.1016/j.infoecopol.2019.05.001
📃 Choudhury, Starr and Agarwal (2020) Patent examiners' domain expertise corrected machine learning bias from incomplete inputs. https://doi.org/10.1002/smj.3152
📃 Grove and colleagues (2000) Mechanical prediction was about 10% more accurate than human judgement on average. https://doi.org/10.1037/1040-3590.12.1.19
📃 Bansal and colleagues (2021) Explanations made people accept AI recommendations whether or not they were right. https://doi.org/10.1145/3411764.3445717
📃 Buçinca, Malaya and Gajos (2021) Making people think first cut over-reliance on AI; they liked it least. https://doi.org/10.1145/3449287
📃 Babashahi and colleagues (2024) Review of 20 studies on how AI is changing the skills work needs. https://doi.org/10.3390/admsci14060127
📃 Bankins, Hu and Yuan (2024) AI automates and complements work, and raises the value of AI literacy and human-centred skills. https://doi.org/10.1016/j.copsyc.2024.101828
📃 Willcocks (2020) Net job loss claims are exaggerated; skills disruption is not. https://doi.org/10.1177/0268396220925830
📃 Gkinko and Elbanna (2023) Employees' trust in a workplace chatbot was cognitive, emotional and organisational. https://doi.org/10.1016/j.jbusres.2023.113707
📃 Fügener, Walzner and Gupta (2026) AI alone on easy tasks, together on similar ones, humans alone on the hardest. https://doi.org/10.1287/mnsc.2024.05684
📃 Kolomaznik and colleagues (2024) Trust, empathy and rapport in human-AI collaboration. https://doi.org/10.3389/fpsyg.2024.1369957
📃 Acemoglu and colleagues (2022) AI-exposed firms hired fewer people for non-AI roles; aggregate effects too small to detect by 2018. https://doi.org/10.1086/718327
📃 Kleinberg and colleagues (2018) Bail decisions. An algorithm could have cut crime up to 24.7% with no more people jailed. https://doi.org/10.1093/qje/qjx032
📃 Dillon and colleagues (2025) Working paper. 7,137 workers in 66 firms. AI users spent two fewer hours a week on email. https://doi.org/10.3386/w33795
📃 Feier, Gogoll and Uhl (2022) Delegating a harmful decision to a machine attracted less punishment than delegating to a person. https://doi.org/10.1007/s11948-022-00372-7
📃 Wu and colleagues (2025) 3,562 people. AI help lifted the task at hand, did not carry over, and lowered motivation. https://doi.org/10.1038/s41598-025-98385-2
📃 Tang and colleagues (2023) 794 employees. Working with AI raised helping and loneliness, with worse sleep and more drinking. https://doi.org/10.1037/apl0001103
📃 Steyvers and colleagues (2025) People overestimated AI accuracy, and longer explanations raised confidence without raising accuracy. https://doi.org/10.1038/s42256-024-00976-7
📃 Bogert, Schecter and Watson (2021) People relied more on algorithms than on peers as tasks got harder. https://doi.org/10.1038/s41598-021-87480-9
📃 Dietvorst and Bharti (2020) Nine studies, 4,820 people. People rejected even the best algorithm in uncertain domains. https://doi.org/10.1177/0956797620948841
📃 Jussupow, Benbasat and Heinzl (2024) An integrative account of algorithm aversion and appreciation. https://doi.org/10.25300/misq/2024/18512
📃 Hui, Reshef and Zhou (2024) Freelancers in the most affected occupations lost work and earnings after generative AI arrived. https://doi.org/10.1287/orsc.2023.18441
📃 Demirci, Hannane and Zhu (2025) A 21% fall in freelance writing and coding job posts within eight months of ChatGPT. https://doi.org/10.1287/mnsc.2024.05420
📃 Santoni de Sio and Mecacci (2021) Four responsibility gaps that open when AI is involved in decisions. https://doi.org/10.1007/s13347-021-00450-x
📃 Macnamara and colleagues (2024) How AI assistance may speed skill decay and hide it from the people affected. https://doi.org/10.1186/s41235-024-00572-8
📃 Rinta-Kahila and colleagues (2023) An accounting firm whose reliance on automation quietly eroded its people's skills. https://doi.org/10.17705/1jais.00829
📃 Natali and colleagues (2025) Review of AI-induced deskilling, and skills never learned, in medicine. https://doi.org/10.1007/s10462-025-11352-1
📃 Qin and colleagues (2025) 163 studies, 82,078 people. People preferred AI only when it seemed more capable and personalisation was unnecessary. https://doi.org/10.1037/bul0000477
📃 Basu and colleagues (2026) Experts accepted AI advice against an investment and discounted AI advice for one, where tacit knowledge mattered. https://doi.org/10.1287/isre.2023.0635
📃 Osborne and Bailey (2025) ChatGPT gave better personal advice than most people; the bias against it appeared only when people knew. https://doi.org/10.1038/s41598-025-86623-6
📃 Simkute and colleagues (2024) Why generative AI can cost productivity: from producing to checking, interruptions, and hard tasks made harder. https://doi.org/10.1080/10447318.2024.2405782
📃 Mayer, Baygi and Buwalda (2025) Entry-level consultants used generative AI to recraft their work, including how their value was seen. https://doi.org/10.1007/s12599-025-00959-x
📃 Bankins and colleagues (2024) Multilevel review of empirical research on AI at work. https://doi.org/10.1002/job.2735
📃 Bainbridge (1983) The ironies of automation: people are left the hardest tasks and lose the practice that kept them able. https://doi.org/10.1016/0005-1098(83)90046-8
📃 Green (2022) 41 policies. People were generally unable to perform the human oversight these policies assumed. https://doi.org/10.1016/j.clsr.2022.105681
📃 Wang and colleagues (2025) GPT-4 judged 63% of 1,022 public sector tasks resistant to replacement. https://doi.org/10.1145/3676281
📃 Sarbani and Hadisoebroto (2025) 26 administrative professionals using AI for writing and scheduling, with concerns about accuracy and training. https://doi.org/10.36914/nv3y1473
📃 Prahl and Van Swol (2017) After bad advice, people dropped automated advice faster than human advice. https://doi.org/10.1002/for.2464
📃 Hou and Jung (2021) Aversion or appreciation could be produced by changing only how the agents were described. https://doi.org/10.1145/3479864
📃 Tomlinson and colleagues (2025) Preprint. 200,000 AI conversations; office and administrative support among the most applicable occupations. https://doi.org/10.48550/arxiv.2507.07935
📃 Gulati and colleagues (2025) Preprint. In roles adopting generative AI, demand for social skills fell 3.4% as admin became self-service. https://doi.org/10.48550/arxiv.2503.09212
P.P.S. Why Do I Even Have the Nerve to Write This?
Fair question.
I am someone who has spent years working alongside remarkable executives, navigating chaos, translating vision into structure and figuring things out, sometimes beautifully and sometimes the very hard way. I have had the privilege of being the steady right hand to leaders who move fast, take risks and expect a lot. That environment has shaped me more than any classroom ever could.
I also happen to have an MBA, and I am studying psychology because people fascinate me. How we work. Why we disconnect. What actually holds us together when the pressure rises.
But really, none of that is the point.
I am here because I have lived the erosion and rebuilt from it. I have seen the cost of carrying too much and the relief that arrives when you finally stop disappearing inside your own competence.
So I share what I have learned in case it helps someone else. Take what is useful. Leave the rest. And remember this one truth that professionals like you often forget:
You are probably doing far better than you give yourself credit for.
Comments