FAIR is AI-Ready: How FAIR Research Outputs Enable Agentic AI in Research
Type: Sheffield analysis (original) Date: 2026-04-17 Related: rda-agentic-ai-consultation-nov2025, rda-ai-in-action-feb2025, research-data-management
Overview and context
This analysis argues that FAIR principles — Findable, Accessible, Interoperable, Reusable — are not merely good open science practice but the technical precondition for AI agents to operate effectively on research outputs. It maps each FAIR sub-principle onto specific agentic AI capabilities, drawing on the tools and workflows already in scope for the Sheffield AI Catalogue.
The framing is strategic as well as technical. FAIR compliance creates the audit trail that governance frameworks and REF 2029 impact claims will require from AI-assisted research. Institutions that build FAIR-by-default workflows now will have the evidential infrastructure to support those claims; those that do not will be retrospectively scrambling to reconstruct lineage from poorly documented notebooks and undated model outputs.
This analysis connects two threads already present in the Sheffield intelligence layer: the RDA’s November 2025 finding that data management automation driven by FAIR compliance burden is the second-highest global priority for agentic AI (see rda-agentic-ai-consultation-nov2025, Finding 2), and the forthcoming UKRI harmonised research data policy mandating FAIR practices across all research councils (see research-data-management).
The core argument
The FAIR principles were explicitly designed to emphasise machine-actionability — the capacity of computational systems to find, access, interoperate, and reuse data with none or minimal human intervention — because humans increasingly rely on computational support to deal with data as a result of the increase in volume, complexity, and creation speed of data.
That last clause is doing a lot of work. Written in 2016, the principles anticipated exactly the kind of agentic AI workflows that are now practically realisable. FAIR outputs are not just good open science practice — they are the substrate on which agentic AI agents operate effectively. A research corpus that is FAIR is, almost by definition, one that an AI agent can navigate autonomously.
F — Findable: Agents need to locate things
The first step in reusing data is to find it. Metadata and data should be easy to find for both humans and computers. Machine-readable metadata are essential for automatic discovery of datasets and services.
Agentic systems — whether Claude Code, an AutoGen workflow, or a multi-agent research pipeline — must first find relevant data, code, or prior results before acting on them. FAIR’s findability requirements map directly onto this:
| FAIR sub-principle | Agentic AI enabler |
|---|---|
| F1 — Globally unique persistent identifiers (DOIs, ORCIDs, ARKs) | Agents can unambiguously resolve references; no disambiguation required |
| F2 — Rich metadata | LLMs can use structured metadata as context to assess relevance without processing entire datasets |
| F3 — Metadata explicitly includes the data identifier | Agents following a metadata record can reliably locate the underlying object |
| F4 — Indexed in a searchable resource | Agents can query registries (DataCite, Zenodo, arXiv) programmatically via APIs |
Sheffield relevance: A Jupyter Notebook deposited on Zenodo with a DOI, tagged with rich metadata (methodology, discipline, software dependencies), becomes something an AI agent can locate and cite in a literature synthesis workflow. Without F1–F4, the agent either hallucinates references or cannot navigate the corpus at all. This directly bears on the RDA community’s highest-priority agentic tool — the Literature Librarian (see rda-agentic-ai-consultation-nov2025, Finding 1): a Literature Librarian agent can only work reliably on findable literature.
A — Accessible: Agents need to retrieve things
Once the required data is located, it must be retrievable via mechanisms that agents can use without human mediation.
Key sub-principles:
- A1 — Retrieval via standardised communications protocol (HTTP/S, REST APIs)
- A1.1 — Protocol is open, free, universally implementable
- A1.2 — Protocol allows authentication and authorisation where necessary
- A2 — Metadata remain accessible even when data are no longer available
Agentic tools interact with research infrastructure through APIs and standardised protocols — not through browser GUIs. A dataset retrievable via an open REST API is directly consumable by a Claude Code agent or a LangGraph workflow. One locked behind institutional VPNs with no programmatic access is effectively invisible to automated pipelines.
The A1.2 nuance is important for universities: authentication is permitted, but it must be machine-implementable. OAuth tokens, API keys, and institutional SSO flows that can be scripted are compatible with agentic access. Manual login pages are not.
Sheffield relevance: A1 with A1.2 enables a middle path — data can be FAIR (accessible to authorised agents) without being fully open. This is the technical basis for federated research agent networks where institutions share compute outputs but not raw sensitive data, an architecture that Sheffield’s data sovereignty priorities make relevant. Any tool assessment in the RDM capability area should note whether the tool exposes a scriptable API (A1-compliant) or requires manual browser interaction.
I — Interoperable: Agents need to combine things
Data usually need to be integrated with other data. In addition, the data need to interoperate with applications or workflows for analysis, storage, and processing.
| FAIR sub-principle | Agentic AI enabler |
|---|---|
| I1 — Formal, shared, broadly applicable language for knowledge representation | Structured formats (JSON-LD, RDF, controlled vocabularies) that LLMs and downstream tools can parse without ambiguity |
| I2 — Vocabularies that follow FAIR principles | Ontologies (SNOMED, Gene Ontology, etc.) that agents can use to reason across disciplinary boundaries |
| I3 — Qualified references to other (meta)data | Explicit links between datasets, enabling agents to traverse provenance chains and research graphs |
This is arguably the most consequential cluster for agentic research. An agent conducting a cross-disciplinary synthesis — linking a materials science dataset to a climate model output, for example — can only do so reliably if both datasets use shared vocabularies and explicit cross-references. Without I1–I3, the agent must perform heroic interpretation work, which introduces error and limits trust.
In the context of Jupyter Notebooks and git repositories:
- A notebook that outputs data in well-documented CSV/JSON with a published schema is I1-compliant
- A git repository with a
CITATION.cfffile and explicit dependency pinning (requirements.txt,environment.yml) gives an agent enough context to reconstruct the computational environment - A notebook referencing other datasets via DOIs (not local filenames) is I3-compliant — an agent can follow those links
Sheffield relevance: This is where FAIR and the Claude Code PR-based oversight model (as used in this repository) intersect tightly. Code review — the human-in-the-loop — can specifically check for I-principle compliance: are outputs in standard formats? are cross-references explicit and resolvable? This review criterion can be operationalised as a checklist item in agent branch PR templates.
R — Reusable: Agents need to act on things
The ultimate goal of FAIR is to optimise the reuse of data. To achieve this, metadata and data should be well-described so that they can be replicated and combined in different settings.
| FAIR sub-principle | Agentic AI enabler |
|---|---|
| R1 — Richly described with accurate, relevant attributes | Agents can assess fitness-for-purpose without human curation |
| R1.1 — Clear and accessible data usage licence | Agents can be programmed to respect licence constraints (e.g., refuse to incorporate non-commercial data into commercial pipelines) |
| R1.2 — Detailed provenance | Agents can trace the lineage of results — essential for attribution, audit trails, and REF impact claims |
| R1.3 — Meets domain-relevant community standards | Ensures outputs are in forms that domain-specific agents and tools expect |
R1.2 (provenance) is strategically the most important for Sheffield’s context. An agentic system that generates code, analyses data, and produces outputs needs to maintain a provenance chain that humans can audit. Git commit histories, Jupyter notebook cell-level execution records, and structured logging in Claude Code all serve as R1.2 artefacts. The Aletheia paper’s Autonomy Levels taxonomy — documenting the degree of AI contribution at each stage — is a domain-level operationalisation of R1.2 for AI-assisted mathematics (see aletheia-autonomous-maths-2026).
The licence question (R1.1) becomes critical as agentic systems start composing outputs from multiple sources autonomously. A research agent that inadvertently combines a CC-BY-NC dataset with outputs intended for a commercial spinout creates a legal problem. Explicit machine-readable licence metadata is the only scalable solution.
Sheffield relevance: For REF 2029, impact claims derived from AI-assisted analysis will require precisely the kind of provenance documentation that FAIR mandates. The PR-based oversight model used for this repository’s agent branches is a working implementation of R1.2 at the code level: every AI contribution is documented, reviewable, and attributable.
Synthesis: FAIR as interface specification for agentic research
Think of FAIR not just as a quality standard but as an interface specification between human-produced research assets and AI agents:
Human researcher → [FAIR outputs] → AI agent → [FAIR outputs] → Human researcher
The loop is only closed if outputs remain FAIR throughout — including the AI-generated artefacts. This means:
- Claude Code outputs should be deposited with DOIs, rich metadata, and explicit provenance (who/what generated them, under which model version)
- Jupyter Notebooks should be version-controlled (git), use persistent data references (not local paths), and include environment specifications
- Git repositories should use
CITATION.cff, include data availability statements, and tag releases with persistent identifiers
The governance angle
FAIR compliance creates a natural audit trail for AI-assisted research — which is exactly what governance frameworks need. The PR-based oversight model operationalises R1.2 at the code level: every AI contribution is documented, reviewable, and attributable.
For REF 2029, this matters acutely. Impact claims derived from AI-assisted analysis will require precisely the kind of provenance documentation that FAIR mandates. Institutions that build FAIR-by-default workflows now will have the evidential infrastructure to support those claims.
The strategic framing for Sheffield leadership: FAIR compliance is not additional overhead on top of AI adoption — it is the precondition for AI adoption that is auditable, reproducible, and REF-defensible. The two agendas are the same agenda.
Quick reference: FAIR principles × agentic tools
| FAIR principle | Git / GitHub | Jupyter Notebooks | Claude Code | Research data repositories |
|---|---|---|---|---|
| F1 — Persistent IDs | Release tags, commit SHAs | Zenodo DOI on deposit | Model version + run ID | DOI (Zenodo, Figshare) |
| F2 — Rich metadata | CITATION.cff, README | Notebook metadata, kernel info | System prompt + context logs | Metadata schemas |
| A1 — Standard protocol | HTTPS/SSH APIs | nbformat, REST APIs | API access | OAI-PMH, REST |
| I1 — Formal language | Standard code formats | Structured cell outputs | Structured outputs (JSON) | RDF, JSON-LD |
| R1.2 — Provenance | Full commit history | Cell execution order | PR review trail | Versioning, lineage fields |
| R1.1 — Licence | LICENSE file | Notebook licence metadata | Output licence declaration | SPDX identifiers |
Recommendations for Sheffield
-
Treat FAIR compliance as an evaluation criterion in tool assessments. When assessing tools in the RDM capability area (and where relevant in others), note explicitly whether the tool supports F1 (DOI/identifier integration), A1 (scriptable API access), and R1.2 (provenance/audit logging). These are not bonus features — they determine whether the tool is compatible with agentic research workflows.
-
Prioritise the
research-data-managementcapability area in the next sprint. The RDA November 2025 consultation identifies data management automation driven by FAIR compliance burden as the second-highest global priority for agentic AI. The Data Director concept — automated metadata creation, DOI minting, repository recommendation — operationalises R1.2 at scale. Sheffield’s RDM team and Library services should be consulted early to avoid the policy-bypass risk that RDA respondents identified. -
Adopt and document the PR-based oversight model as Sheffield’s operational expression of R1.2 for AI-generated research artefacts. The agent branch workflow already used in this repository (agent/YYYY-MM-DD-description → human review → merge) should be positioned explicitly in Sheffield’s GenAI Principles for Research as the workflow that satisfies provenance requirements.
-
Reference this analysis in Sheffield’s GenAI Principles for Research documentation. The FAIR × agentic AI framing provides a technically grounded basis for why FAIR data practices should be positioned as an enabler of (rather than a constraint on) AI adoption. This reframe may be useful in researcher-facing communications.