I can't actually browse the internet live or generate a Word file directly, so I should be upfront about that while still offering a useful table format they can paste into Word, drawing on real, open-access academic articles I know about AI and automation's impact on administrative government jobs post-2020. I need to be careful to only cite articles I'm confident are accurate. I'm recalling candidates like Straub et al. (2023) in Government Information Quarterly, which proposes a framework for AI in government but isn't job-focused, and Bright et al. (2024) on arXiv, a UK public sector survey finding significant generative AI adoption and usage among public servants with implications for their roles. I'm also recalling a related Alan Turing Institute paper: Straub et al. (2024), "AI for bureaucratic productivity," arXiv:2403.14712, which estimated that AI could automate roughly 84% of the 143 million complex repetitive transactions processed by UK central government. This confirms the estimate: about one billion citizen-facing transactions across 400 services, with 143 million complex ones, 84% highly automatable, yielding that 1,200 person-year savings figure from just one minute saved per transaction. The paper also notes high turnover in service types, suggesting automation efforts should target general procedures rather than specific services. Moving to the next reference, I'm checking Gaozhao, Wright & Gainey's 2023 piece on bureaucrat versus AI preferences in Public Management Review—unsure about open access status, so I'll flag it as uncertain and move on to the next citation. Then there's Giest & Klievink (2022), "More than a digital system," also in Public Management Review—I believe this one is open access given the authors' Leiden affiliation and Dutch funding patterns, examining how AI changes bureaucrats' roles across different organizational contexts, likely using Dutch case studies. Busuioc's 2021 piece on algorithmic accountability is close to the cutoff date, so I'll treat it cautiously since its focus is accountability rather than job impacts anyway. Wirtz et al. 2019 is too old to include, and I'm also considering Zuiderwijk, Chen & Salem's 2021 systematic literature review on AI implications for public governance. The review finds mostly exploratory, qualitative research on AI's implications for public governance, and the authors propose a research agenda with eight process-related and seven content-related recommendations — pushing for more empirical, multidisciplinary studies focused on specific AI forms rather than AI broadly. This source seems relevant to AI in government generally, but I'm uncertain it directly addresses workforce or jobs implications, so it may be a weaker fit for that angle specifically. Switching tracks, Eloundou et al.'s "GPTs are GPTs" (2023, arXiv, later published in Science 2024) is a strong candidate—publicly available and directly on topic, finding that roughly 80% of the US workforce could have at least 10% of tasks affected by LLMs, with about 19% of workers facing more substantial exposure. Since this isn't government-specific, though, I'm also considering Gmyrek, Berg, & Bescond's 2023 ILO Working Paper 96 on generative AI and jobs, which specifically flags clerical support workers as the most exposed occupational group—finding 24% of clerical tasks highly exposed and another 58% with medium exposure, which ties directly into administrative service functions. This ILO Working Paper by Gmyrek and colleagues, published August 2023, is a strong candidate since it directly addresses clerical/administrative support jobs and notes the gendered impact—women are affected at more than double the rate of automation exposure compared to men. It's publicly available academic-style research, though not government-specific, and I recall there's also an updated version of this paper worth checking. Looking at other candidates: a 2024 piece by Selten and Klievink on "Organizing public sector AI adoption" in Government Information Quarterly seems potentially relevant, likely open access through Leiden, though I'm not confident on the specifics. I'm continuing to weigh other options too. I'm unsure of the OA status for the Ahn & Chen piece, so I'll set it aside. The Kuziemski & Misuraca 2020 paper falls outside my post-2020 cutoff, so I'm skipping it too. For Neumann, Guirguis & Steiner (2024) in Public Management Review, I'm fairly confident it's open access under CC BY. The findings suggest that technological and organizational factors matter more at different adoption stages, while environmental factors are generally less important, advancing theoretical understanding of AI adoption specificities across stages. This is relevant to organizational readiness for strategic planning, though not jobs-focused specifically, and I recall factors like data availability, IT infrastructure, and top management support being discussed, though I'm not certain of all the specifics. I'm now moving to Wang, Zhang, and Zhao's 2022 paper on understanding determinants. I'm listing several more sources: Chen, Guo, Gao, and Liang's 2021 work on AI-based self-service technology in public service delivery; Wenger and Wilkins' 2009 piece on unemployment insurance automation, which I'm dismissing as too old; and Bullock, Young, and Wang's 2020 paper on AI, bureaucratic form, and discretion in public service. I'm uncertain about open access status for most of these. I'm also checking Young, Himmelreich, Honcharov, and Soundarajan's 2022 Government Information Quarterly piece on AI detecting administrative errors in unemployment insurance, and Alon-Barkat and Busuioc's 2023 JPART article on automation bias and selective adherence in human-AI public sector decisions, which I believe is open access under CC BY. The abstract continues describing the three-study design: study 1 (N=605) tests automation bias by comparing adherence to algorithmic vs. human-expert predictions, finding no evidence for it, while study 2 (N=904) moves on to test selective adherence against stereotypes. This paper is relevant for discussing automation's effects on citizen-facing decisions, though it's more about decision-making staff than administrative support roles. I'm noting Young, Bullock & Lecy (2019) on artificial discretion but excluding it as too old, and moving on to check the next candidate on AI and the future of work. Checking Mergel, Dickinson, Stenvall & Gasco's 2023 editorial introducing a special issue on implementing AI in the public sector — likely open access but I'm unsure of specific content details. I'm also flagging Tangi et al.'s 2022 AI Watch report from the European Commission as another possibility to evaluate. Next, I'm considering van Noordt & Misuraca's 2022 Government Information Quarterly article mapping AI use across EU public services, which appears open access given its JRC authorship — trying to recall whether their findings centered on service delivery improvements, internal management, or chatbot applications. This one seems like a strong candidate. I'm now checking other citations like Lindgren et al. 2019 on digitalization of public services, but that's too old, and Ranerup & Henriksen's work on automated decision-making in social services, which also predates the window I need. Continuing down the list, I'm checking Hjaltalin & Sigurdarson (2024) on AI strategic use in public sector values analysis, then Madan & Ashok (2023) on AI adoption/diffusion in public administration, a systematic review likely OA from University of Reading authors examining enablers and barriers. I'm also considering Medaglia, Gil-Garcia & Pardo (2023) in Social Science Computer Review mapping AI-in-government research into categories like governance and data, then Wirtz, Weyerer & Kehl (2022) on AI governance risk frameworks, and Selten, Robeer & Grimmelikhuijsen (2023) — trying to pin down OA status for each. Selten et al. is likely CC BY open access from Utrecht University, examining a survey experiment with Dutch police officers showing officers trust AI recommendations that match their own professional judgment, with explainability having little effect on trust or compliance. the hope that fair AI systems can correct human biases—relevant to frontline workers but focused on police, not admin support. I'm scanning through more candidates: a 2021 study on AI voice robots in Chinese public sector (unsure about open access), a 2019 paper on AI chatbots for citizen-government communication (too old), and a 2020 experimental study on public trust in AI chatbots. Continuing to check dates—a 2020 GIQ paper and an 2018 PAR piece are both too old. I also recall "Smart technology and the emergence of algorithmic bureaucracy" by Vogl et al. (2020) on AI in UK local authorities, which is open access but still published in 2020, so it doesn't meet the after-2020 criterion. Moving on, I'm ruling out the Deloitte 2017 report as non-academic, but noting Bright et al. (2024) on generative AI in the public sector as a good candidate, along with a possible 2025 paper by Hashem and colleagues on generative AI in the UK public sector that I'm not fully certain about. I'm also considering Humlum & Vestergaard's NBER/BFI working paper on large language models and labor market effects, which uses Danish administrative data on thousands of workers across exposed occupations. I'm checking Brynjolfsson, Li, & Raymond's "Generative AI at Work" study on customer support agents, which found meaningful productivity gains from AI assistance, especially for novice and lower-skilled workers. I'm screening a few more candidates and ruling them out — Dell'Acqua's consulting study, Noy & Zhang's writing-task experiment, and OECD's non-academic report don't fit the government focus. I'm also checking Canadian public-sector sources like Treasury Board and the Brookfield Institute, but these are policy reports rather than peer-reviewed academic articles. The terminology in the prompt — "Administrative Support," "Services Delivery Division" — hints this could be the Government of Canada's AS classification system, though I shouldn't assume the country and will keep my framing general enough to apply broadly, perhaps touching on union consultation or collective agreement obligations. I'm also scanning for relevant academic literature on RPA in the public sector, like work from Lindgren, Toll, and Melin on automation in Swedish local government, though none of these seem directly on point for this specific scenario. Still checking other generative AI in government candidates like Cantens (2024) in AI & Society and Salah et al. (2023), along with Wirtz, Langer & Fenner's 2021 research agenda piece on AI in the public sector. Moving on, Ingrams, Kaufmann & Jacobs (2022) in Policy & Internet looks like a strong CC BY open access candidate examining citizen trust in AI versus human decision-making in government, likely involving a survey experiment that varies task complexity. For the remaining entries — Miller & Keiser (2021) in JPART on representative bureaucracy and automated decision making, Schiff, Schiff & Pierson (2022) in Public Administration on public value failure in AI adoption, and Criado, Valero & Villodre (2020) — I'm uncertain about their open access status and need to flag them for further verification. I'm also checking Wenzelburger et al. (2024) in Public Administration on algorithms and context, and Grimmelikhuijsen (2023) in Public Administration Review, which I believe is CC BY licensed and reports Dutch experiments showing algorithmic transparency affects perceived trustworthiness of automated decisions. Looking closer at Grimmelikhuijsen's findings: explainability (more than mere accessibility) of algorithms boosted perceived trustworthiness, and decisions made by humans rather than algorithms were generally viewed more favorably. The study tested procedural fairness theory across two experimental scenarios—a low-discretion visa denial and a high-discretion welfare fraud suspicion—to see how accessibility and explainability of street-level algorithmic decisions affected citizen trust. This is useful background on citizen trust in AI-driven decisions, though I note it centers on citizens' perspective rather than directly on administrative jobs. I need to refocus more specifically on what the user asked: AI's impact on government administrative service jobs and functions. I'm now identifying strong candidates — Straub et al.'s 2024 arXiv paper measuring AI's potential to automate UK government transactions is highly relevant and publicly accessible, and Bright et al.'s 2024 survey on generative AI's spread in the public sector looks promising too. Bright et al.'s survey covers nearly a thousand public service professionals across education, health, social work, and emergency services, examining their awareness, usage, and attitudes toward generative AI, including optimism about reduced bureaucratic burden alongside concerns about trust and guidance gaps. I'm trying to recall the specific statistics and findings from the abstract to confirm its relevance. NHS workers estimated bureaucracy time could fall from 50% to 30% with proper generative AI use, saving roughly one day a week. Trust in outputs was high (61%) and fear of replacement low (16%), yet 76% felt the UK was missing opportunities and only 32% reported clear workplace guidance—showing adoption is happening without organized oversight. This points to an urgent need for clearer public sector guidelines. Next I want to verify the ILO working paper by Gmyrek, Berg & Bescond (2023) on generative AI's effects on job quantity and quality, particularly their finding on clerical work exposure. Given the user's framing around government and administrative service jobs, this ILO paper fits well since clerical occupations overlap heavily with government administrative roles—though I'm weighing whether a more government-specific study might be preferable, and trying to recall if any open-access, post-2020 government workforce impact studies come to mind. I recall the specific breakdown: 31% of Canadian employees in high-exposure, low-complementarity occupations, 29% in high-exposure, high-complementarity roles, and 40% in low-exposure occupations, though this is a StatsCan research publication rather than a peer-reviewed academic paper, and it's not specific to government employment. Other candidates I'm turning over include Wirtz et al. (2020) on AI governance frameworks for public administration, which doesn't fit since it predates recent developments, and Johnson, Coggburn & Llorens (2022) in Public Personnel Management on AI and public HR management -- though I'm uncertain about its open-access status. I'm also weighing less likely fits: something on generative AI and the public sector workforce, robotic process automation studies on public employees, and van Noordt & Misuraca (2022) on the EU landscape, which seems like a solid lead, while Wilson & van der Velden's sustainable AI piece doesn't seem relevant to my focus here. I'm considering a few more candidates — Selten & Klievink (2024) on organizing public sector AI adoption in Dutch organizations, and Maragno et al. (2023) on factors shaping AI implementation in public sector organizations — though I'm uncertain whether either is actually open access. Henman (2020) on AI in public services probably isn't relevant enough either. I'm trying to recall Busuioc's 2021 piece and whether it was actually published online earlier. Alon-Barkat & Busuioc (2023) seems solid and open access, and I'm fairly confident Giest & Klievink (2024) in Public Management Review covers how AI is reshaping bureaucrats' roles across organizational contexts. I don't think that one applies either. I also consider other possible citations like OECD governance papers (Berryhill et al. 2019 on AI in the public sector), but that seems too early to be the right match, so I keep searching for a better fit. I'll note that arXiv preprints and working papers from places like the Alan Turing Institute or ILO are legitimate academic research even if not peer-reviewed, so I should flag publication type transparently in the table rather than overstate them as peer-reviewed journal articles. I'm also uncertain whether certain 2024 papers (Straub et al., Bright et al.) were later published in peer-reviewed venues, so I should be cautious about asserting that without confirmation. Now I'm weighing the five best sources for admin support strategic planning: the automation potential of government transactions paper, the ILO clerical exposure study highlighting augmentation over automation and gender/job quality implications, and the bottom-up usage study showing informal adoption among public servants with little guidance. 4. AI deployment patterns in government (van Noordt & Misuraca, 2022) — covering chatbots, process automation, and where these tools concentrate in service delivery vs internal management. 5. Human-AI interaction risks, weighing between Alon-Barkat & Busuioc's work on automation bias and selective adherence versus Grimmelikhuijsen's work on algorithmic transparency and trust in citizen-facing decisions like visa denials or welfare fraud detection. I'm also considering Neumann et al. on AI adoption readiness in Swiss public organizations, and Selten et al. on street-level bureaucrats trusting AI that confirms their existing judgment — weighing which set best fits a focus on impacts to administrative jobs and functions, while also checking which sources are peer-reviewed versus working papers. I'm also considering articles specifically on automation's effects on public sector employees' work, weighing candidates like Bullock, Huang & Kim on machine intelligence and bureaucracy, Busch & Henriksen on digital discretion, and de Boer & Raaphorst on automation and discretion — checking each for open access and peer-review status. Checking "Ranerup & Henriksen" on Trelleborg's RPA for social assistance in Social Science Computer Review — but the online publication date seems to be mid-2020, which makes it ambiguous for my cutoff and I should set it aside. I'm also trying to recall whether Ida Lindgren published a 2024 piece in Government Information Quarterly on the "ironies of automation" in public service contexts. I'm less confident citing specific findings here though. There's also Toll, Lindgren, Melin & Madsen (2020), and Madsen, Lindgren & Melin (2022) on "the accidental caseworker," examining how digital self-service shifts administrative burden from caseworkers onto citizens — likely OA via Government Information Quarterly. Though it's digital self-service rather than AI directly, it's still automation-adjacent since front-line staff shift to assisting citizens with digital channels. Refocusing on the core topic of AI and automation's impact on administrative service jobs, I'm building an ideal source mix: Straub et al. 2024 on automation potential of government transactions, Gmyrek et al. 2023 ILO findings on clerical job exposure and augmentation, and Bright et al. 2024 on generative AI already in use by public servants. I'm also weighing van Noordt & Misuraca 2022 on where AI is applied in EU government (service delivery and internal management), and Alon-Barkat & Busuioc 2023 on human-AI decision-making biases among public sector staff, against Madsen et al. 2022's "accidental caseworker" concept, which fits citizen-facing service shifts well but isn't strictly AI-focused — though it still aligns with the broader goal of expanding automation and AI tools. I'm not fully certain on the scenario details, but I recall the study involved three survey experiments in the Netherlands with sample sizes roughly 605, 904, and 1,345, totaling around 2,850 participants, including civil servants in at least one study. Checking publication years, Alon-Barkat & Busuioc appeared online in JPART in early 2022, with print publication in 2023. For van Noordt & Misuraca, I believe the GIQ article is open access under CC BY given JRC/EU-affiliated authors, and the other sources — Gmyrek et al., Bright et al., and Straub et al. — all fall in the 2023-2024 range, satisfying the post-2020 requirement. I'm weighing whether the ILO paper fits the topic since it's a general labour-market study rather than government-specific, but it does cover clerical/administrative occupations using ISCO-08 task exposure scoring, so I can justify including it with a note on its global, cross-sector scope. This finding about clerical roles being most exposed, with women disproportionately affected, is directly relevant to workforce planning for admin support teams, and it's a highly credible, well-cited source worth including. I also recall ILO released a 2025 update (Gmyrek et al., with NASK Poland) reaffirming that about one in four jobs globally are exposed, with clerical work still topping the list — worth noting as a more recent corroborating source alongside the original 2023 paper. Now I'm also thinking about the Straub et al. paper on AI for bureaucratic productivity, measuring the potential to automate 143 million UK government transactions, and recalling the full author list for proper citation. The key findings stick with me: about 84% of these complex transactions are highly automatable, and saving just one minute per transaction would equate to roughly 1,200 person-years of work annually. Given high turnover across services, the paper argues for automating common process steps rather than specific services, and they built a model to estimate transaction volumes without exhaustive measurement. For practical implications, this suggests freed-up capacity should be redeployed toward complex cases and citizen support rather than treated as headcount cuts, and I should map my own service's process steps to prioritize high-volume procedures like intake, verification, and eligibility checks for reusable automation. Now I'm looking at Bright et al. (2024), a paper on how generative AI is already widespread in the public sector. This survey of 938 UK public service professionals found 45% aware of GenAI use in their area and 22% actively using it, with NHS staff believing bureaucratic workload could drop significantly, strong trust in outputs, low fear of replacement, but weak workplace guidance and a sense of missed opportunity—adoption is happening bottom-up in a disorganized way. This points to an urgent need for clear usage policies around privacy and accuracy, targeting documentation burden as an early win, training staff in verification given their high trust in outputs, and engaging staff as co-designers given their low fear of replacement. I should also flag that this sample is frontline professionals rather than admin staff specifically, so I'd want to validate against my own organization's data before generalizing. Now I'm turning to Gmyrek, Berg & Bescond's 2023 work on generative AI. Automation potential is concentrated in clerical roles, which disproportionately affects women since they're over-represented in clerical work, especially in high-income countries. I'm trying to recall the exact figures—something like globally 3.7% of female employment is potentially exposed to automation effects versus 1.4% of male employment, with the gap widening in high-income countries to roughly 7.8% for women versus 2.9% for men. The ILO's August 2023 findings noted that clerical work had the highest technological exposure, with about a quarter of tasks highly exposed and over half medium-exposed, while managers, professionals, and technicians showed much lower shares of highly exposed tasks. The safest framing is that women's employment is more than double as likely to be affected by automation due to their concentration in clerical roles. I'm also weighing whether to mention the high-income (5.5%) versus low-income (0.4%) exposure gap, though it's secondary. The key implications: treat administrative support as the priority job family for reskilling and job redesign, favor augmentation over replacement by shifting roles toward client service and exception handling, apply a gender-equity lens given the female-dominated admin workforce, and engage unions early while monitoring job quality. I'm also noting van Noordt & Misuraca's 2022 study reviewing 250 EU government AI cases, which found AI is mostly applied to service delivery and internal management rather than policy-making, with different technologies (chatbots for citizen interaction, predictive analytics for inspections), though most cases remain pilots with limited evidence of real impact. Other common categories include predictive analytics and simulation, computer vision and identity recognition, expert/rule-based decision systems, and NLP or text mining tools. I'm fairly confident chatbots and virtual assistants rank as the most frequent application for citizen-facing public service delivery, though I'm not fully certain this specific paper states it that explicitly. This suggests different AI types map to different governance functions — citizen-facing tools like chatbots handle public enquiries, while analytics and automation tools support internal administrative processes like records management and correspondence. Turning to another relevant study, Alon-Barkat & Busuioc (2023) examined human-AI interactions in public sector decisions, specifically testing for "automation bias" and "selective adherence" to algorithmic advice through survey experiments with Dutch civil servants, conducted in the aftermath of the Dutch childcare benefits scandal. Their findings showed no clear automation bias, but did reveal selective adherence—people followed advice more readily when it matched stereotypes, though notably this effect disappeared among civil servants, likely due to heightened scrutiny after the scandal. This suggests that "human in the loop" safeguards alone may not prevent AI from reinforcing biases against vulnerable groups. For presentation, I'm now thinking through formatting practicalities—keeping the comparison table compact enough to fit one page, probably landscape orientation with tight margins and 8-9pt font, five rows by three columns with brief bullets per cell. Since I can't generate an actual .docx file directly, I'm considering the best workaround: producing an HTML-based table with Word-compatible styling (including landscape page setup) that the user can save with a .doc extension and open directly in Word, which handles HTML rendering well for this purpose. I also need to be upfront with the user that I can't browse live online — my research is drawn from my training knowledge, so they should verify the sources before finalizing anything. For formatting, I'll present the deliverable as a markdown table with columns for Study Information, Key Findings, and Implications for the Government, giving the user something polished they can copy straight into a Word document. Given the length tradeoff, I'll go with the rendered markdown table as the primary deliverable, plus quick formatting tips for turning it into a landscape Word page (narrow margins, small font, repeated header row), and offer the HTML/.doc version as a follow-up if they want it rather than duplicating all the content upfront. I'm drafting a title line for the scan and weighing whether to add a short "cross-cutting takeaways" note beneath the table — useful for strategic planning, but risky for the one-page limit, so I'll keep it minimal or optional rather than core to the deliverable. Now let me verify the details for each article are accurate before finalizing. For Straub et al. (2024), I'm confirming the title about AI's potential to automate UK government transactions, the Alan Turing Institute affiliation, the focus on roughly 400 citizen-facing services, and the core goal of measuring the scale of bureaucratic decisions ripe for automation. I'm fairly confident about the headline figures: roughly one billion citizen-facing transactions yearly across 400 services, with about 143 million being complex repetitive transactions, 84% of which are highly automatable, and saving even one minute per transaction equating to roughly 1,200 person-years of work saved annually. I double-check this math—143 million minutes works out to about 2.38 million hours, which divided by a ~1,900-2,000 hour work-year lines up well with the 1,200 person-years claim. Now I'm moving to the second article, Bright et al. (2024), on generative AI's spread in the public sector. I'm recalling the survey details: 938 UK public service professionals across education, health, social work, and emergency services were surveyed. I'm trying to verify the specific percentages—61% trust in generative AI outputs paired with only 16% fear of replacement, plus 76% feeling the UK is missing out on AI opportunities for public services and just 32% feeling there's clear workplace guidance. I'm also recalling a detail about NHS workers estimating that time spent on bureaucracy could fall from 50% to 30% with proper generative AI use. Now I'm shifting to the third article, Gmyrek, Berg & Bescond's 2023 ILO working paper on generative AI's effects on job quantity and quality—a global analysis using GPT-4 task-level exposure mapped to ISCO-08 occupations alongside employment data by country income group, with clerical work exposure around 24%. I'm then moving to the fourth article, van Noordt & Misuraca's 2022 piece on AI landscaping across EU public sector institutions, published in Government Information Quarterly. Looking at the 250 AI use cases from the JRC AI Watch inventory, I note the findings show AI is mostly used for public service delivery and internal management rather than policy decision-making, with different AI types suited to different governance functions. I'm trying to recall whether this article is open access, thinking through Colin van Noordt's TalTech PhD funding and believing it was published under CC BY. Now I'm moving to the Alon-Barkat & Busuioc 2023 paper on automation bias and selective adherence to algorithmic advice in public sector decision making, published in JPART and confirmed open access under CC BY 4.0 via Oxford Academic. The study setting is the Netherlands, using three survey experiments with civil servants. I'm also double-checking that each of the five selected articles genuinely addresses the impact on administrative service jobs and functions, confirming Straub covers transaction processing automation and Gmyrek covers clerical job impacts as I continue reviewing the rest. Good coverage overall across scale of opportunity, job exposure, staff behavior, deployment patterns, and oversight risks. Now I want to tailor the implications column specifically to Admin Support positions in a citizen-facing Services Delivery Division, keeping it concise while drafting the table with columns for study info and key findings. I'll structure Study Information to include title, authors, date/source, setting, goal, and a verifiable link for each entry so the user can check sources directly. For van Noordt & Misuraca I'm fairly confident it's DOI 10.1016/j.giq.2022.101714, matching Government Information Quarterly 39(3), article 101714. For Alon-Barkat & Busuioc I recall the DOI being 10.1093/jopart/muac007. I'm weighing whether to include DOIs and arXiv IDs at all given verification risk, but I'm moderately-high confidence on 2401.01291 for Bright et al.'s generative AI paper and 2403.14712 for Straub et al.'s bureaucratic productivity paper, so I'll include them with appropriate caveats. I'll mark the access info with venue, "open access," and identifiers where I'm fairly confident, like the muac007 DOI for the Alon-Barkat & Busuioc JPART piece. Rather than a separate references list, I'll keep citation details compact within each Study Information cell to stay within the one-page limit. For Row 1, I'm drafting the Straub et al. (2024) entry covering the Alan Turing Institute's AI automation study of UK government transactions — setting up the study information, UK central government context across ~400 services, and the goal of measuring automation potential across roughly a billion citizen transactions. Of those, 143 million are complex and repetitive, with 84% rated highly automatable, and even saving a minute per transaction would free up around 1,200 person-years annually. The implications point toward focusing automation on high-volume admin tasks like intake and eligibility checks, building reusable tools across programs rather than service-specific ones, and planning where the freed-up capacity should go before rolling automation out. Now I'm starting to pull together the next row's source information. Bright et al. (2024) surveyed 938 UK public service professionals and found generative AI is already widespread: 45% were aware of its use in their area and 22% actively use it, with NHS staff estimating bureaucratic time could drop from 50% to 30%. Trust in outputs was high (61%) and fear of job loss low (16%), though only 32% reported having clear workplace guidance on its use. Given this bottom-up, disorganized adoption, I think organizations should issue clear rules now on approved tools and data privacy, target early wins like drafting correspondence and summarizing notes, and train staff to verify AI outputs while keeping accountability human. I'm now moving to the next source, Gmyrek, Berg & Bescond (2023). This ILO working paper uses GPT-4-assisted scoring to estimate GenAI exposure across occupations globally, finding clerical support roles are by far the most exposed (24% highly exposed tasks, 58% medium) compared to just 1-4% in other occupational groups, with the likely effect being augmentation rather than replacement. Since clerical work skews heavily female, affected job shares for women run more than double those for men, and the actual impact on job quality depends heavily on implementation and social dialogue. This points to admin support as the priority workforce planning area, with recommendations to redesign roles around tasks rather than positions, fund reskilling into adjacent career paths, apply a gender/equity lens, consult unions early, and monitor job quality metrics post-implementation. Now I'm moving to the next study — van Noordt & Misuraca's 2022 landscaping analysis of AI use across European Union public administrations, covering 250 cases examining impacts on service delivery, internal management, and other core government functions. The key finding is that AI adoption concentrates on improving public service delivery and internal management, with much less use in actual policy-making, and different AI types map to different functions — chatbots for citizen interaction, analytics for internal processes. I'm noting that evidence on real-world impact remains limited, which aligns with the study's call for deeper investigation into AI's actual role and effects. This suggests citizen-facing services and back-office administration are the primary entry points for AI, pointing toward admin support functions like handling enquiries and correspondence. Moving to the next study, Alon-Barkat and Busuioc's 2023 paper on human-AI interaction in public sector decision-making examines "automation bias" and "selective adherence" to algorithmic advice, based on three survey experiments in the Netherlands including one with civil servants, testing whether officials over-rely on AI recommendations or apply them selectively. The findings show no evidence of automation bias — officials don't follow AI advice more than they would otherwise. Instead there's "selective adherence": people follow AI advice more when it confirms stereotypes about certain groups, creating risk for vulnerable clients, though notably civil servants surveyed after a major Dutch algorithm scandal showed no such selective adherence, suggesting awareness can reduce bias. This implies that "human-in-the-loop" isn't an automatic safeguard — staff need bias-awareness training and clear override protocols, their roles shift toward validating AI outputs requiring new competencies like critical review and data literacy, and outcomes should be monitored by client group to catch disparate impacts. Estimating line counts per column at 8pt font in a three-column landscape layout, each row likely needs around 9-10 lines to fit everything comfortably, so I should plan row heights accordingly to avoid overflow. This is tight given the available page height after margins and header, so I'll need to trim bullet content, cap bullets per cell, and adjust column widths—giving Study Info roughly 32% and the other two columns 34% each to balance the tallest cell. For the first row's Study Info cell, I'm drafting a tight summary: the paper title on AI for bureaucratic productivity, authors from the Alan Turing Institute, publication date and arXiv link, and the UK government setting covering roughly 400 citizen-facing services. The Key Findings and Implications cells capture that ~1B citizen transactions happen yearly with 143M being highly automatable (84% rated so), that saving even a minute per transaction equals huge person-year savings, and that the practical takeaway is to focus automation on shared procedures across programs rather than individual services, prioritizing high-volume intake/verification work and deciding upfront how freed capacity gets redeployed. Now I'm looking at a second study on GenAI adoption in UK public services, finding that while awareness is moderate (45%) and active use lower (22%), trust is high and fear of job loss is low — but guidance remains unclear, so adoption is happening organically rather than through coordinated top-down rollout. I'm drafting implications for our own context: treat informal GenAI use as already present and get policies on approved tools and client data in place now, target quick wins like drafting, summarizing, and note-taking, pair high trust with verification training, and involve Admin Support staff directly in piloting since the original study didn't focus specifically on administrative roles. Moving to a third study, I'm noting it's an ILO global working paper from 2023 examining how generative AI could affect both the quantity and quality of jobs worldwide. Key findings show clerical work is by far the most exposed occupation group, with most of the effect likely being task augmentation rather than full job replacement, though this automation risk falls disproportionately on women since their jobs are exposed at more than double the rate of men's -- I want to phrase that precisely as "share of women's jobs exposed to automation >2x men's" rather than the looser wording. I'm starting to draft the implications section, beginning with the fact that Admin Support is the most exposed function. Next I'm noting the practical recommendations: redesigning roles task-by-task toward higher-value work, investing in reskilling, applying equity analysis with union consultation, and monitoring job quality post-rollout. Moving to the next study, van Noordt & Misuraca's 2022 landscaping of AI use in EU governments, which examined 250 cases across public administrations to assess effects on service delivery, internal management, and policy-making -- finding AI is used mostly for service delivery and internal management, rarely for policy decisions, with different AI types serving different functions. For the implications, I'm noting that front-counter and back-office admin functions are where AI lands first, so matching tools to function and building evaluation metrics like service quality and wait times into pilots is key. Then moving to the fifth study, Alon-Barkat & Busuioc's 2023 work on automation bias and selective adherence to algorithmic advice among Dutch civil servants. Findings show no blanket automation bias, but "selective adherence" where advice matching group stereotypes was followed more, posing risk to vulnerable clients — though awareness from a prior scandal reduced this bias. This means human review isn't an automatic safeguard, so bias-awareness training and clear override rules are needed, admin roles should shift toward critical review skills, and outcomes should be monitored across client groups. For the study info section, I'm deciding to use clear labeled fields like Title, Authors, Published, Setting, and Goal in bold so comparisons across studies are easier to scan. I'm also adding a descriptive header above the table framing the research scan around AI and automation implications for administrative support. After the table, I'll include a few cross-cutting takeaways and a short note on sourcing — noting these five sources are all freely accessible and published 2022–2024, with a mix of two peer-reviewed journal articles, two preprints, and one ILO working paper. I'm weighing whether to offer peer-reviewed alternatives in case the department wants strictly peer-reviewed sources — candidates like Grimmelikhuijsen (2023, PAR), Neumann et al. (2024, PMR), and Madsen et al. (2022, GIQ) seem open-access and reasonably solid, though I want to flag my confidence level on each rather than overstate certainty. I also need to be upfront that I compiled this from memory rather than a live search, so the user should verify each link before circulating. Then I'll add quick Word formatting tips — landscape orientation, narrow margins, small font, autofit table — to keep everything on one page, and maybe mention an HTML export option as a convenience. I'm also weighing whether to add a short cross-cutting synthesis row, since the request is really about guiding strategic planning. I'll frame it as optional so it doesn't threaten the one-page constraint. I'll finalize the three takeaways: augment rather than replace roles, prioritize governance since staff are already using GenAI informally, and emphasize piloting with equity monitoring for both staff and citizens. Now I'm re-verifying the factual claims from Straub et al. against the abstract—checking the transaction volume figures and the breakdown of complex versus repetitive decisions to make sure I haven't overstated anything. Checking Bright et al. figures: 938 participants across education, health, social work, and emergency services, with 45%/22% splits, NHS shifting from 50% to 30% (saving about a day a week), 61% trust versus 16% fearing replacement, and only 32% reporting clear guidance, often implemented in a disorganized way. For Gmyrek et al., I'm verifying clerical automation rates (24% high, 58% medium) versus other occupations (1–4% high), the augmentation-over-automation framing, the finding that women are affected over twice as often, and the conclusion that job quality outcomes hinge on implementation and social dialogue. I'm also trying to recall the abstract's exact framing around work intensity, autonomy, and how outcomes depend on managing the diffusion of generative AI, along with the methodology using GPT-4 to score task-level exposure and estimate employment effects by country income group, where clerical work stands out as uniquely exposed. entially exposed to automation, climbing to 5.5 percent in high-income countries, with women disproportionately affected at more than double the rate of men. Augmentation effects are larger still, touching 10.4 percent of employment in low-income countries and 13.4 percent in high-income ones, though infrastructure gaps could deepen the digital divide by limiting adoption in poorer regions. I keep coming back to the idea that these numbers matter less for their precision than for what the overall distribution reveals about the likely direction of change, which points toward a need for proactive policies on job quality and fair transitions. Turning to van Noordt and Misuraca's study of roughly 250 AI cases across the EU, I'm recalling that AI was found mainly supporting public service delivery, then internal management, with only limited direct involvement in policy decision-making -- and that different AI technologies map onto different governance functions, which I'll illustrate with examples like chatbots likely being common in service delivery. Checking my full set against the criteria: all five sources are published 2022-2024, satisfying the post-2020 requirement, and I'm reasonably confident all are publicly accessible, especially the two journal articles. Now I'm thinking about framing the final output as a polished deliverable written for an administrative operations lead, matching the professional tone they set up in their request. I should note that I can't attach an actual .docx file, but I can present the table in a way that pastes directly into Word on a single landscape page. I also want to be transparent that these sources come from my own knowledge rather than a live search, recommending the user verify the links themselves, and I'll keep that disclosure brief in the notes after the table. I'm also double-checking the exact titles and capitalization of each article for accuracy. I'm less confident about the exact author ordering for the remaining citations, so to minimize error risk I'll use "et al." after the first couple of authors I'm confident about, like Straub, Hashem, Bright for one paper and Bright, Enock, Esnaashari for another. For the shorter author lists where I'm fully confident — ILO (3 authors), van Noordt (2), and Alon-Barkat (2) — I'll spell out full names rather than using "et al." since those are easy to verify. For ILO's scope, I'll clarify it as occupation-level rather than sector-specific, noting it covers clerical support roles that include government admin staff. For the publication dates, I'll note the preprint date for the arXiv paper, the journal and issue for the Government Information Quarterly piece, and an approximate online publication date for the JPART article, then move on to including arXiv identifiers for the preprints. To avoid risking a wrong DOI, I'm weighing whether to include the arXiv ID and DOI for the GIQ paper (fairly confident) versus the JPART DOI "muac007" (moderately confident but plausible given the article's early 2022 acceptance). Given the risk of error, I'm now considering just citing "Free via arXiv.org" or similar access notes instead of specific identifiers for the less certain ones. I'll include the arXiv IDs for the Turing papers since the numbering patterns (2401.01291 for early January 2024, 2403.14712 for mid-March 2024) seem consistent with their expected posting dates, and I'm fairly confident in them. I'll skip DOIs since volume/page info suffices for journal citations, and now I'm checking the setting details for the Alon-Barkat study. For the Netherlands study, I'm noting the three survey experiments used public samples and civil servants, with participants likely drawn from a Dutch online panel for studies 1 and 2 — though I'm not certain of the exact scenario used and will leave that detail vague. I do want to confirm Study 3's timing relative to the childcare benefits scandal, and now I'm thinking through how to frame the "implications for government" column neutrally. Imperative phrasing works well since the user works in government, so I'll use action-oriented statements. I'm now structuring the markdown table with columns for study information, key findings, and government implications, using bullet points within cells for clarity. Generating markdown table... would free up roughly 1,200 person-years annually, with the biggest gains coming from high-volume processing tasks like intake, data entry, and eligibility checks. I should map our own transactions and process steps to rank automation candidates, build reusable tools for shared steps, and decide upfront where the freed-up capacity goes—complex cases, citizen support, or QA. A second paper I'm noting is on generative AI already being widespread in the public sector, from Bright, Enock, Esnaashari et al. at the Alan Turing Institute, published January 2024. Looking at their survey of 938 UK public service professionals—covering health, education, social work, and emergency services—I see that 45% are aware of GenAI use in their area and 22% actively use it, with NHS staff estimating bureaucracy time could drop from 50% to 30%. Trust in outputs is high (61%) and fear of replacement is low (16%), but only 32% report having clear guidance, meaning uptake is informal and disorganized. This tells me I should assume informal use already exists among our staff and set rules on approved tools and data handling now, start with easy wins like drafting letters and summarizing files, and involve Admin Support staff directly in designing pilots since this survey wasn't admin-specific. Next I'm turning to the ILO's global analysis of GenAI's effects on job quantity and quality, which used GPT-4-assisted scoring of task exposure across occupations linked to employment data. Clerical support emerges as by far the most exposed occupational group—24% of tasks highly exposed and 58% moderately exposed, compared to just 1-4% for other groups—with the likely effect being augmentation rather than full job loss, though automation risk falls disproportionately on women's jobs since they're concentrated in clerical roles. This confirms that Admin Support should be treated as our top workforce-planning priority given how exposed that job family is. I'm also noting practical recommendations here: redesigning admin roles task-by-task toward client service, exception handling, QA, and data stewardship; funding reskilling pathways into program, data, or digital roles; applying a gender/equity lens and involving unions early; and tracking workload and job quality after any rollout. Then I move to a 2022 peer-reviewed study examining 250 cases of AI use across EU public administrations, looking at how AI affects service delivery, internal management, and policy-making—finding that AI is used mainly to improve service delivery, then internal management, and rarely for policy decisions, with different AI types suited to different functions like chatbots for citizen-facing services. This suggests admin support functions—enquiries, correspondence, scheduling, records—will be among the earliest to change, so I should match tools to functions, adapt proven use cases from peer governments, and build evaluation of service quality, wait times, and staff workload into every pilot. Finally I'm bringing in a 2023 study on automation bias and selective adherence, testing whether Dutch civil servants over-rely on or selectively follow algorithmic advice. (4) Pilot, measure and monitor the impacts on both staff—including gender equity—and citizens, covering bias and service quality. I should also check what sources these five findings draw from before finalizing notes. All five are free to access online and published between 2022 and 2024, though only two are peer-reviewed journal articles—the rest are preprints or an ILO working paper, which I should flag if the team needs strictly peer-reviewed material. Since I compiled this from memory rather than a live search, I want to note that someone should verify each link and confirm open access before relying on it. I'm also working out formatting for fitting the table onto one page in Word—landscape orientation, narrow margins, smaller font, and adjusted column widths—and I'm noting two peer-reviewed alternatives (Grimmelikhuijsen 2023 and Madsen, Lindgren & Melin 2022) that could swap in if stricter sourcing is required. Double-checking those citations now: both appear correctly attributed with proper journal details, volume/issue numbers, and open-access status confirmed for each. I'll drop the swap suggestion to keep things short and instead offer to swap preprints for peer-reviewed alternatives if needed, maybe naming Grimmelikhuijsen 2023 as an example. I should also mention I can provide an HTML version that opens in Word with formatting intact, folding that into the closing line. Now checking the fit for one page — estimating roughly 8-9 lines per row across study info, findings, and implications columns at this font size, which should work within the space. Rows 4 and 5 add roughly 9 and 10 lines each. Adding up the header, title, and takeaways sections, the total comes to about 53 lines, which at 8pt single spacing works out to roughly 7.1 inches before accounting for cell padding. This is pushing right up against the usable height for a landscape Letter page with 0.5" margins, so I need to tighten things up — shrinking margins to 0.4", keeping the font at 8pt, minimizing cell padding, and possibly condensing the authors/published info into a single combined line to save vertical space per row. I'm also considering adjusting column width proportions (roughly 32%/34%/34%) since the tallest cell in each row drives the row height, and I'm tightening the bullet language itself — trimming each finding down to its essential numbers and takeaway so each cell stays compact. For Row 1 Implications, I'll note that the biggest gains come from high-volume processing work like intake, data entry, verification, and eligibility checks, so it's worth mapping our transactions to rank candidates, building reusable tools for shared steps, and deciding in advance where the freed-up capacity should go. For Row 2, I'm summarizing the GenAI awareness and trust figures, then noting that since informal use already exists, we should set clear rules on approved tools and client-facing boundaries. For early wins, I'm thinking about drafting letters, summarizing files, and note-taking as low-risk starting points, paired with training staff to verify AI outputs given the high trust levels, and co-designing pilots directly with Admin Support staff rather than relying on a non-admin-specific survey sample. For Row 3, clerical work stands out as the most exposed category — 24% of tasks highly exposed and 58% moderately, compared to just 1-4% in other groups — with the main effect being augmentation rather than full job loss. Since automation risk concentrates in clerical roles, women's jobs are affected more than twice as much as men's, so outcomes will depend heavily on policies around job quality, fair transitions, and social dialogue. I want to double-check the precise wording of that gender disparity claim before citing it. For implications, Admin Support becomes the top workforce-planning priority: redesigning roles task-by-task (client service, exceptions, QA, data stewardship), funding reskilling pathways into programme, data, and digital roles, applying a gender/equity lens, involving unions and staff early, and tracking workload and job quality post-rollout. For Row 4, AI is mainly used for service delivery and internal management, rarely in policy decisions, with the type of AI varying by function (chatbots for enquiries, analytics/automation internally) — though actual impacts remain poorly understood and need deeper evaluation. This points to citizen-facing and back-office admin work as the first landing zones for AI, covering enquiries, correspondence, and scheduling. For Row 5, findings show no "automation bias" (algorithmic advice isn't followed more than equivalent human-expert advice), but there is "selective adherence" — advice followed more when it matches group stereotypes, risking harm to vulnerable clients — though civil servants surveyed after a major Dutch algorithm scandal showed no such bias, suggesting awareness may help curb it. This means a human-in-the-loop isn't an automatic safeguard, so bias training, override/escalation rules, critical-review and data-literacy skills, and outcome monitoring across client groups are all needed to keep decisions explainable and accountable. Now I'm compiling the study details: the Alan Turing Institute's work on AI's potential to automate 143 million UK government transactions (Straub et al., Mar 2024), their survey on widespread generative AI use in the UK public sector (Bright et al., Jan 2024), and the ILO's global analysis of generative AI's effects on job quantity and quality (Gmyrek, Berg & Bescond, Aug 2023). Next I'm noting van Noordt & Misuraca's 2022 landscaping study of AI across EU public administrations, examining its impact on service delivery, internal management, and policy-making, followed by research on human-AI interaction effects like automation bias and selective adherence to algorithmic advice in public sector decisions. I'm also thinking about how to wrap up with a brief "bottom line" section summarizing cross-cutting takeaways for planning purposes. I should also note I can't actually generate a .docx file, so I'll mention that briefly in the intro line and follow it with the page-setup tips to keep everything on one page. I also want to flag that these sources come from my background knowledge rather than a live search, so the person should verify access and figures before sharing. I'm debating whether to add "(open access)" or "(free)" tags next to each entry to make the access criterion visible right in the table itself—I think it's worth the extra words to show compliance clearly. I'll drop the "ISCO-08" jargon to keep things concise, just referring to "occupational tasks" instead, and confirm I'll use `
` for line breaks with bold/italics inside cells where needed. For the deliverable title, I'm settling on "Research Scan: AI & Automation in Government – Implications for Administrative Support Positions & Functions" with a subtitle noting it's for the Services Delivery Division and the 5-Year Strategic Plan, rendered as a markdown header so it converts cleanly to a Word heading. I'm now drafting the cross-cutting takeaways section, starting with a bullet on augmentation versus replacement as the first common thread for the strategic plan. Next I'm adding bullets on governance — flagging that staff already use GenAI without guidance, so policy and oversight are urgent — and on building evaluation into pilots to track impacts on staff and citizens. Then I'm adding a notes section explaining the source criteria: all five are free, published 2022–2024, with a mix of peer-reviewed articles, preprints, and an ILO working paper, flagging that I can suggest peer-reviewed swaps if the planning team requires it. I should also remind the person to verify each source themselves since I compiled this from knowledge rather than a live search, and give them quick Word formatting steps to fit the table on one page — landscape orientation, narrow margins, 8pt font, AutoFit, and roughly 30/35/35 column widths. I'll note upfront that I can't attach a .docx directly, and add a nice touch with a "Prepared by" line under the title for their Word document. Now I want to double-check the labeling on the third column and verify my phrasing of the "~143M complex but repetitive" transaction figure, since I'm not fully certain how Straub et al. define "complex repetitive" — I should hedge this by rephrasing it as transactions requiring staff decision-making, like processing applications, rather than stating their exact definition with full confidence. Rows 4 and 5 check out. I'm also verifying author names—Enock, Esnaashari, and Straub as lead author all seem correct, and the ILO working paper citation with Gmyrek, Berg, and Bescond checks out too. For the remaining citations, van Noordt & Misuraca's GIQ article and Alon-Barkat & Busuioc's JPART piece both look right. I'm thinking through formatting now—the table needs clear headers like "Study Information," "Key Findings," and "Implications for the Government," and I could suggest adding banded rows as a Word formatting tip since markdown itself can't do shading. I want to double check the recency criterion too: the earliest sources are from 2022, so framing it as "five open-access studies (2022–2024)" works. One thing worth flagging is that the ILO paper is economy-wide rather than government-specific, so I should note it's the most direct evidence on clerical/administrative roles even though it's not government-focused specifically. I'll format the table in markdown since it should render into HTML cleanly enough to paste into Word while preserving structure and formatting. Since I can't actually attach a .docx, I'll explain that upfront and give page-setup guidance for keeping it on one page, then draft the final response with a title header for the research scan deliverable. Now I'm filling in the first row with the Alan Turing Institute study on AI automating UK government transactions, including its authors, publication details, and key findings on citizen-facing transaction volumes. The findings show that roughly 1 billion citizen transactions occur yearly, with 143 million classified as "complex repetitive" bureaucratic decisions, 84% of which are rated highly automatable, and I'm capturing implications for Admin Support: mapping transaction types, building reusable shared tools, and planning where freed capacity should go. I'm starting a second row for a related study on generative AI adoption in the public sector. This Turing Institute survey of 938 UK public service professionals finds 45% aware of GenAI use in their area but only 22% actively using it, with NHS staff estimating bureaucracy time could drop from 50% to 30%, high trust (61%) but low organized guidance (32%) — suggesting we should set rules on approved tools now since informal use likely already exists, with early wins around drafting correspondence. Next I'm looking at the ILO working paper on generative AI's global effects on job quantity and quality, which used GPT-4-assisted scoring to estimate occupational exposure — clerical support emerges as the most exposed occupational group. Recommendations should include an equity lens and early union/staff involvement, with workload and job quality tracked post-rollout. I'm now moving to a 2022 study by van Noordt and Misuraca examining 250 cases of AI adoption across EU public administrations. It finds AI mostly improves service delivery and internal management, rarely informs policy, with different tool types (chatbots vs. analytics) serving different functions — so citizen-facing and back-office tasks change first, and pilots should build in evaluation of service quality and workload. Now I'm looking at a study on automation bias: Alon-Barkat and Busuioc's 2023 Dutch experiments find civil servants don't blindly over-trust algorithmic advice, but do show "selective adherence" — following AI recommendations more when they align with group stereotypes, which risks harming vulnerable clients. For the strategic plan, this points to three practical lessons: human-in-the-loop isn't a guaranteed safeguard, so bias-awareness training and clear escalation rules matter; staff roles are shifting toward critically reviewing AI outputs, requiring data-literacy skills; and organizations need to monitor outcomes across client groups to keep decisions explainable and accountable. I'm pulling together common threads for the plan — augmentation over replacement in Admin Support roles, and the urgent need for governance given staff already using GenAI without formal policy. I should note my source criteria (five free, 2022-2024 sources, mix of peer-reviewed and preprints, with one ILO working paper), flag that the planning team should verify access and details themselves since I compiled this from literature knowledge rather than live search, and add a quick tip for fitting the table onto one page in Word using landscape orientation, narrow margins, and smaller font. I'm also weighing whether to suggest dropping the "Common threads" box if it pushes the document past one page, and double-checking the claim about citizen-facing and back-office admin work being where AI lands first in terms of enquiries, correspondence, scheduling, and records. I think softening "will change earliest" to "likely to change first" fits better as a prediction, and the other implication rows—biggest gains in high-volume processing, early wins like drafting correspondence and summarizing files, and funding reskilling pathways—all seem like reasonable inferences from the study. I also want to double-check formatting consistency: using "open access" rather than "free" across entries, and avoiding "|" inside table cells since it breaks markdown—I'll switch to " – " as the separator for things like publication dates and preprint numbers. Citing the van Noordt & Misuraca figure of 250 cases looks solid. I don't need a relevance or evidence-strength rating since it wasn't requested, and I should stay transparent that this reflects synthesized research rather than live online lookups. I'll finalize attribution formatting and tighten a few bullets for clarity—like rephrasing the automation-exposure point to more clearly state that women's jobs face more than twice the exposure share compared to men's. Finishing the remaining goal rows for rows 4 and 5, confirming the header row structure, and double-checking markdown formatting—single-line rows, `
` for breaks, no stray pipes, proper bold/italic nesting. I think the table is complete and ready. I can't attach a .docx file, but the table below pastes directly into Word. Page-setup steps to keep it on one page follow the table. --- ### Research Scan: AI & Automation in Government – Implications for Administrative Support Positions & Functions *Services Delivery Division · Input to the 5-Year AI & Automation Strategic Plan · Prepared by: Administrative Operations Lead · [Date]* | Study Information | Key Findings | Implications for the Government | |---|---|---| | **1. AI for bureaucratic productivity: Measuring the potential of AI to help automate 143 million UK government transactions**
**Authors:** Straub, V.J., Hashem, Y., Bright, J., et al. (Alan Turing Institute)
**Published:** Mar 2024 – arXiv preprint 2403.14712 (open access)
**Setting:** UK central government; ~400 citizen-facing services
**Goal:** Measure the volume of citizen-facing bureaucratic transactions and their potential for AI automation | • ~1 billion citizen transactions/yr; ~143M are "complex repetitive" transactions involving bureaucratic decisions
• ~84% of these rated highly automatable
• Saving just 1 min per transaction ≈ 1,200 person-years of work/yr
• Services change often → automate shared procedures, not individual services | • Biggest gains likely in high-volume processing (intake, data entry, verification, eligibility checks), which is core Admin Support work
• Map our transactions and process steps to rank automation candidates
• Build reusable tools for steps shared across programs
• Decide up front where freed capacity goes (complex cases, client support, QA) | | **2. Generative AI is already widespread in the public sector**
**Authors:** Bright, J., Enock, F.E., Esnaashari, S., et al. (Alan Turing Institute)
**Published:** Jan 2024 – arXiv preprint 2401.01291 (open access)
**Setting:** UK survey of 938 public service professionals (health, education, social work, emergency services)
**Goal:** Gauge current use, trust and perceived potential of generative AI (GenAI) | • 45% aware of GenAI use in their area; 22% actively use it
• NHS staff estimate time on bureaucracy could fall from 50% to 30% (~1 day/week)
• High trust in outputs (61%); low fear of replacement (16%)
• Only 32% report clear guidance; uptake is bottom-up and "disorganised" | • Assume informal use already exists → set rules on approved tools and client data now
• Early wins: drafting correspondence, summarizing files, note-taking
• High trust → train staff to verify outputs; accountability stays with staff
• Co-design pilots with Admin Support staff; survey our own staff (study sample was not admin-specific) | | **3. Generative AI and jobs: A global analysis of potential effects on job quantity and quality**
**Authors:** Gmyrek, P., Berg, J. & Bescond, D. (International Labour Organization)
**Published:** Aug 2023 – ILO Working Paper 96 (open access)
**Setting:** Global, economy-wide; GPT-4-assisted scoring of task exposure across occupations, linked to employment data
**Goal:** Estimate GenAI exposure by occupation and effects on job quantity and quality | • Clerical support is the most exposed occupational group: 24% of tasks highly exposed + 58% medium (other groups: 1–4% highly exposed)
• Main effect likely augmentation (some tasks automated, freeing time for others), not whole-job loss
• Automation exposure is concentrated in clerical roles → the share of women's jobs affected is more than double men's
• Outcomes depend on policies for job quality, fair transitions and social dialogue | • Admin Support is the most exposed job family → make it a workforce-planning priority
• Redesign roles task by task toward client service, exception handling, QA and data stewardship
• Fund reskilling and career pathways (e.g., into program, data, digital roles)
• Apply a gender/equity analysis; involve unions and staff early
• Track workload and job quality after rollout | | **4. Artificial intelligence for the public sector: Results of landscaping the use of AI in government across the European Union**
**Authors:** van Noordt, C. & Misuraca, G.
**Published:** 2022 – *Government Information Quarterly*, 39(3), 101714 (peer-reviewed, open access)
**Setting:** European Union; 250 cases of AI use in public administrations
**Goal:** Examine how AI use affects core government functions: service delivery, internal management, policy-making | • AI used mainly to improve public service delivery, then internal management; rarely for policy decisions
• Different AI types serve different functions (e.g., chatbots for citizen enquiries; analytics/automation for internal processes)
• Actual impacts are still poorly understood → deeper evaluation needed | • Citizen-facing and back-office admin are where AI lands first → enquiries, correspondence, scheduling and records likely to change earliest
• Match tools to functions; adapt proven use cases from peer governments
• Build evaluation (service quality, wait times, staff workload) into every pilot | | **5. Human–AI interactions in public sector decision making: "Automation bias" and "selective adherence" to algorithmic advice**
**Authors:** Alon-Barkat, S. & Busuioc, M.
**Published:** 2023 (online 2022) – *Journal of Public Administration Research and Theory*,

Multi-skils