RDA Global Community Priorities for Agentic AI in Research (November 2025)
Source: Global Community Priorities for Agentic AI in Research — Consultation Results Published: December 2025 Publisher: Research Data Alliance (RDA) in collaboration with Microsoft
Report overview and context
This report presents findings from a global community consultation launched by the RDA and Microsoft in November 2025 to identify researcher priorities for agentic AI across the research lifecycle. For this consultation, agentic AI was defined as “artificial intelligence systems capable of autonomous operation with minimal human oversight” — a deliberately expansive framing intended to capture both current tools and emerging multi-step agent systems.
The consultation comprised two components: four online information sessions (190 registrants, 70 participants from 33 countries across 6 continents), and an anonymous 15-minute survey that received 83 complete responses from 25 countries. Survey respondents evaluated 11 proposed agentic AI tools spanning the entire research lifecycle, each identified through Microsoft’s prior engagement with over 50 UK research institutions. Respondents were predominantly RDA members (59%), with researchers and scientists comprising the largest single stakeholder group (39.8%). European respondents formed the largest regional cohort (n=44 of 83).
Two methodological caveats matter for interpreting these results. First, the 11 proposed tools were not independently derived from the global community — they emerged from Microsoft’s UK-institutional engagement and were then validated globally, which may have primed respondents towards a particular framing of the solution space. Second, the European over-representation (53% of respondents) limits generalisability, particularly for Global South contexts.
The report also documents responsible AI use: Claude Sonnet 4.5 (Pro) was used for data analysis and writing assistance throughout, with a privacy-preserving configuration ensuring no training on input or output data. All AI-assisted content was reviewed and validated by the authors.
Finding 1: Literature discovery is the highest-priority agentic AI capability across researcher types
The Literature Librarian — a proposed tool to search literature using natural language queries integrated with library subscriptions — ranked first overall (weighted score 3.35/4.0) and appeared in the top three for European respondents, researchers and scientists, early career individuals, and librarians. Its near-87% positive rating was the highest of any proposed tool.
The conceptual framing is important: the Literature Librarian is not simply a better search engine. Its defining feature is library subscription integration — enabling AI-powered discovery of both open access and paywalled content via existing institutional licences. This is precisely the gap that researcher-facing tools such as Elicit, Research Rabbit, and Semantic Scholar do not close, as they operate on the open literature. The RDA community’s appetite for this tool reflects a practical pain point: paywall fragmentation across platforms, combined with the volume of published research, makes systematic literature discovery genuinely difficult.
Concerns raised by respondents were substantive: hallucination risks requiring verification of all AI-generated citations, reproducibility of natural language queries (the same question asked differently may return different results), licensing complications with scholarly publishers regarding AI access, and the risk of AI-generated “slop” contaminating search results. Several respondents noted that existing tools (ChatGPT, Perplexity, Deep Research) already provide partial functionality.
Sheffield relevance: Sheffield holds institutional licences for Web of Science, Scopus, and other major databases. A Literature Librarian-type tool integrated with those licences would be the single highest-impact agentic AI provision the institution could make for researchers. Tools in the literature-review capability area should be evaluated against this criterion: does the tool integrate with institutional subscription APIs, or is it limited to open access content? This is a useful discriminating dimension that the current schema does not explicitly capture — it could be added as a note in the licensing-model or existing-university-license field, or flagged in the tool body’s trust and governance section.
Finding 2: Data management automation is the second-highest community priority, driven by FAIR compliance burden
The Data Director (rank #2, weighted score 3.24/4.0) was proposed to automatically create metadata following FAIR principles, mint DOIs, and recommend appropriate data repositories. It ranked first for research performing organisations and libraries, and second for European respondents and researchers and scientists. The ~74% positive rating was the second-highest overall.
The appeal is well understood by the RDA community: metadata creation is work that researchers recognise as necessary but find burdensome and low-status. Automating it would reduce human error, lower administrative overhead, and potentially benefit researchers at smaller institutions without dedicated data stewardship support. Some respondents noted that experimental agentic tools for metadata creation and repository submission already exist.
Concerns centred on accuracy and accountability: AI-generated metadata may be plausible but incorrect; researchers need sufficient understanding of data publication requirements to verify AI outputs; minting DOIs has a cost and an institutional governance dimension; and some institutions already provide this service through data stewards and librarians. The risk of researchers bypassing institutional policies by following AI recommendations uncritically was explicitly raised.
Sheffield relevance: The Data Director concept maps directly onto the research-data-management capability area in the Sheffield catalogue, which is one of the less thoroughly assessed areas relative to its strategic importance. The RDA finding that data management automation is the second-highest global priority among researchers validates prioritising this capability area in a future sprint. Sheffield’s Research Data Management team and Library services are the existing provision the Data Director concept would augment — any tool assessment in this area should include explicit consultation with those services to avoid the policy-bypass risk that RDA respondents identified.
Finding 3: Funding discovery is a universal priority; grant application automation is contested
The Funding Finder (rank #3, weighted score 3.18/4.0) was the most consistently prioritised tool across all stakeholder groups — it appeared in the top three for every group analysed. The core appeal is automated scanning of funding databases matched to research ideas, with application template support and compliance checking against funder guidelines.
The concerns raised were more substantive than for the top two tools. Respondents argued that suggesting application improvements based on previous successful applications would entrench “buzzwords” and stifle innovative research. Several questioned whether agentic AI was necessary — well-organised databases or traditional search engines might suffice. Concerns about hallucinations in a high-stakes context (a missed funding opportunity or a rejected application based on incorrect guidance has real consequences) were prominent. One concern was that funding is relationship-based, not process-based, and AI cannot navigate that dimension.
The community consensus was that a Funding Finder should be a discovery and awareness tool, not an application writer.
Sheffield relevance: Research Professional (the existing institutional tool for funding discovery at Sheffield) was named as an existing relevant tool by respondents. This suggests the Funding Finder concept is solving a real gap beyond what Research Professional provides — specifically, natural language matching between research ideas and opportunities, rather than keyword-based alerts. A Funding Finder tool in the ideation-and-grant-writing capability area should be evaluated against Research Professional’s functionality to determine whether it adds incremental value or primarily repackages existing provision.
Finding 4: The community draws a sharp line between AI as assistant and AI as agent — especially for ethics and publication
Two proposed tools generated disproportionate scepticism: the Ethics Reviewer (rank #11, lowest overall) and the Publication Preparer (rank #10). Both performed consistently poorly across stakeholder groups and regions. The qualitative responses explain why.
For the Ethics Reviewer, respondents argued that the ethics review process is intentionally cognitively demanding — it requires researchers to think critically about their own work’s implications. Automating it would defeat its educational and reflexive purpose. Concerns about AI’s inability to handle jurisdictional nuance, cultural context, and institutional specificity were prominent. The community consensus was that any ethics AI tool should be strictly a compliance assistant, never a decision-maker.
For the Publication Preparer, the specific concern about content-rewriting capabilities dominated: respondents did not want AI to alter their voice, generate novelty statements, or recommend predatory journals. Formatting and grammar assistance was acceptable; content generation for publication was not. One reviewer stated they would “reject any novelty statement that appears to be written by AI.”
The Ethics Reviewer finding is significant in the context of the broader agentic AI framing. Despite high optimism about AI acceleration, respondents drew a consistent line at tasks requiring human moral judgment.
Sheffield relevance: These findings validate the schema’s human-oversight-required: true as a universal default and reinforce the importance of the output-explainability field for tools used in ethics-adjacent and publication contexts. More specifically, the community’s strong rejection of AI ethics review tools suggests that any Sheffield tool assessment for ethics-support software should be framed as process support (finding guidelines, checking completeness, flagging missing sections) rather than decision support. The distinction matters for how tools are described in the Key capabilities section of tool notes and for what catalogue-status they can credibly hold.
Finding 5: Ambivalence and conditional acceptance are the dominant community postures toward agentic AI
The consultation’s qualitative findings present a more complex picture than the quantitative rankings alone suggest. When asked about the future of agentic AI in research, information session participants viewed it as transformative and inevitable — but their primary concerns were transparency (understanding how models work), bias and hallucinations, loss of research skills through reduced hands-on data engagement, and unequal access.
Three concerns deserve specific attention for Sheffield planning:
Error propagation: Multiple respondents noted that as tasks are handed over to agentic systems, researchers may lose the ability to identify errors. One stated that “agentic AI systems must be completely dependable as researchers may lose the ability to identify errors as tasks are handed over to these systems.” This is not a concern about current tool quality — it is a structural concern about the epistemological effect of agentic AI on researchers’ critical capacities over time.
‘Black box’ risk: Respondents warned about creating a “black box” situation that compromises both the scientific process and open data principles. A system that produces correct outputs through opaque processes is insufficient for reproducible research.
Vendor lock-in and sustainability: Concerns about over-dependence on proprietary systems, cost escalation, and differential access between well-resourced and under-resourced institutions were raised. This is directly relevant to Sheffield’s institutional procurement decisions.
Sheffield relevance: The error propagation and black box concerns are the most practically significant for Sheffield’s catalogue. The concern about researchers losing error-detection capability as AI takes over tasks implies that tools should be introduced into workflows with explicit attention to maintaining researcher understanding of the underlying processes — not just whether the AI is accurate, but whether using the AI degrades the researcher’s own capability over time. This is a dimension of bias-risk-notes and human-oversight-required that the current tool notes do not capture well. A future revision of the draft-tool-note.md workflow might add a prompt asking: “Does routine use of this tool risk degrading the researcher’s ability to perform the underlying task manually?”
Implications for this project
| RDA finding | Catalogue implication |
|---|---|
| Literature Librarian is the #1 global priority, defined by library subscription integration | Evaluate literature-review tools explicitly against institutional subscription integration — add to trust and governance notes |
Data Director (#2) validates research-data-management as high-priority capability area | Prioritise RDM capability area in next sprint; consult Library and RDM services before finalising any tool notes |
| Funding Finder (#3) is universally valued as a discovery tool, not an application writer | Tools in ideation-and-grant-writing should be positioned as discovery aids; evaluate against Research Professional’s existing coverage |
| Ethics Reviewer ranked last — community rejects AI as moral decision-maker | Ethics-adjacent tool notes should use output-explainability: low and clearly scope AI role as compliance assistance only |
| Publication Preparer ranked second-last — community objects to AI content generation for publication | scholarly-writing tools where AI generates publishable content should carry explicit caveats in the tool note body |
| Error propagation concern: agentic AI may degrade researcher’s own critical capabilities | Consider adding a capability-degradation risk prompt to draft-tool-note.md workflow |
| Black box and reproducibility concerns are mainstream, not specialist | output-explainability should be treated as a primary procurement criterion, not an annotation |
| Vendor lock-in and sustainability concerns raised explicitly | Existing ongoing-maintenance and licensing-model fields should address lock-in risk; consider adding lock-in notes to tool body |
Recommended adjustments to the catalogue
Three proposals are suggested by these findings. None require immediate action in this commit.
1. Library subscription integration as a tool evaluation dimension
The highest-ranked agentic AI concept globally is defined by its ability to integrate with institutional library subscriptions. This dimension is not currently captured in the tool schema. It does not require a schema change: the existing-university-license field (currently boolean) could be supplemented with a note in the tool body explaining whether the tool integrates with Sheffield’s database subscriptions directly. For literature-review tools, this should become a standard element of the trust and governance notes.
2. Capability-degradation risk in the draft-tool-note workflow
The RDA consultation’s concern about researchers losing error-detection capability through AI delegation is a genuine pedagogical risk that the current tool notes do not address. Adding one prompt to _scripts/draft-tool-note.md — “Does routine use of this tool risk degrading the researcher’s ability to perform this task independently, and if so, how should this be mitigated?” — would bring this into scope without changing the schema.
3. Research data management sprint
The Data Director finding validates research-data-management as the second-highest priority capability area for researchers globally. It is currently one of the least thoroughly assessed areas in the Sheffield catalogue. A dedicated sprint covering automated metadata tools, repository recommendation engines, and data management planning assistants — developed in consultation with Sheffield Library and Research Data Management services — would address both the RDA-identified priority and a genuine institutional gap.
Drafted: 2026-04-01 · Branch: agent/2026-04-01-rda-ai-in-action-survey