How FAIR Data Principles Prepare Data for AI Agents, and Why They Need a G for Governance

FAIR was written for machines in 2016

The FAIR data principles come from a paper published in March 2016 in Scientific Data, The FAIR Guiding Principles for scientific data management and stewardship by Mark Wilkinson and co-authors. The four letters stand for Findable, Accessible, Interoperable and Reusable, and they break down into fifteen short sub-principles.

What set FAIR apart from earlier data management advice was its audience. The authors stated that the principles put specific emphasis on the ability of machines to find and use data automatically. Their worry was that research data had grown beyond what people could search and read by hand, so the data had to describe itself well enough for software to do the work.

For most of the following decade, that machine reader was a search index, a pipeline or a repository harvester. Today it is increasingly an AI agent, a program that receives a goal in natural language, decides which data it needs, retrieves it and acts on the result. An agent has no colleague to ask which table is the right one. Almost everything it knows about your data before it reads a single row comes from metadata, which is precisely the layer FAIR was designed to improve.

FAIR through the eyes of an AI agent

The easiest way to see why FAIR matters for AI is to read each principle as a question the agent must answer before it can do anything useful. The sub-principle numbers below are from the original paper.

Findable. The agent first has to discover that relevant data exists. Sub-principle F4 asks that data and metadata are registered or indexed in a searchable resource, and F2 asks for rich metadata to search on. In an enterprise that resource is the data catalog. Without one, an agent sees table names and nothing else, and picks whichever looks closest.

Accessible. Next the agent has to retrieve the data through a standard protocol. A1.2 adds that the protocol should allow authentication and authorization where necessary, and A2 asks that metadata stays reachable even when the data itself is not. Both matter for agents. The first means an agent acts under a known identity with explicit rights. The second means an agent can still learn that a restricted dataset exists, who owns it and how to request it, instead of concluding the data is missing.

Interoperable. Then the agent has to interpret what it retrieved. I1 and I2 ask for a shared language for knowledge representation and vocabularies that are themselves FAIR. For an enterprise agent that is the business glossary. When "active customer" means one thing in the CRM and another in finance, the agent cannot reconcile them on its own, and it will confidently answer with the wrong definition.

Reusable. Finally the agent has to judge whether it can trust the data and use it for this task. R1.1 asks for a clear usage license and R1.2 for detailed provenance. Inside a company the license becomes usage conditions and classification, and provenance becomes data lineage. An agent that can trace a revenue figure back to its source system can say where the number came from, which is what turns an answer into evidence.

FAIR-R, FAIR² and FAIR4AG² extend FAIR for AI

FAIR was written before today's large language models, so it is no surprise that researchers have started to extend it. Three proposals are worth knowing, because each one adds a different piece of what AI needs.

FAIR-R comes from Stefaan Verhulst, Andrew Zahuranec and Hannah Chafetz. Their paper Moving Toward the FAIR-R Principles argues that FAIR does not by itself address what AI needs, and adds a fifth principle, Readiness for AI. A summary by the Spanish open data portal lists what it emphasizes: appropriate labeling and enrichment, transparency about origin, lineage and processing, formats and metadata suited to AI, and enough coverage and quality to avoid bias.

FAIR² (FAIR squared) is the operational counterpart, a specification from the publisher Frontiers and its AI venture Senscience, introduced in March 2025. It aims to make research data AI-ready and aligned with responsible AI principles, and it is compatible with the MLCommons Croissant format for machine learning datasets. Verhulst himself describes the relationship simply. FAIR-R is a conceptual expansion and FAIR² adds the operational guidance.

FAIR4AG² is the most recent and, for anyone building agents, the most useful. Published in Frontiers in Artificial Intelligence in August 2026 by Qian and colleagues, it redefines FAIR around agent-actionability, the degree to which a resource can be discovered, accessed, executed, traced and interpreted by an autonomous agent without human intervention. The paper is written for agricultural research, yet its ideas travel well. It recommends exposing databases and models through standard APIs and then wrapping them in Model Context Protocol servers, so that agents call stable, authorized functions instead of scraping pages.

The part we find most telling is its scoring. Every principle is rated unmet, partial or met, and "partial" describes a resource that serves human users but falls short of autonomous agent use. That is the state of most enterprise data today. A person with context can work around a vague column name or a missing owner. An agent cannot.

Where FAIR stops, and the G for governance

FAIR makes data usable by machines. It says much less about whether a particular machine should use particular data for a particular purpose. For research data that gap was tolerable, because a human researcher still made the decision. With agents acting on their own, someone has to make that decision in advance and encode it where the agent will find it.

That is why we use the shorthand FAIR-G, FAIR plus governance. FAIR-G is our working shorthand rather than a published standard, a way of naming what an enterprise needs on top of the four letters before it lets agents loose on its data.

• Ownership. Every dataset, term and data product has a named owner who answers for it.
• Purpose and policy. Usage conditions say what the data may be used for, which is the enterprise version of the FAIR license.
• Classification. Sensitivity labels, such as personal or confidential data, travel with the data so an agent knows what it is handling.
• Approval. Access is requested and approved through a defined process, for agents the same way as for people.
• Audit trail. Every grant and every use leaves a record that an auditor or an AI governance team can read.

We are not alone in drawing this line. The FAIR4AG² authors describe their own framework as necessary but not sufficient and call for parallel governance mechanisms, among them machine-readable consent, attribution standards for automated reuse and policies for human oversight. And the research world has a precedent. In 2019 the Global Indigenous Data Alliance published the CARE principles for Indigenous data governance as a companion to FAIR, under the motto "Be FAIR and CARE". FAIR covers how data moves and connects, and CARE covers the rights that decide whether and on what terms it should. FAIR-G applies the same pairing to the enterprise.

Three AI use cases and the letter that breaks first

Abstract principles become concrete quickly once you watch an agent fail. Here are three common enterprise use cases and the FAIR-G letter that usually gives way first.

An assistant answering from documents. A retrieval-augmented assistant indexes policies, manuals and reports. The first failure is usually R, missing provenance. The assistant retrieves a superseded version of a policy because nothing marks which document is current and who approved it. The second is G, when a confidential file was indexed because nobody classified it.

An analytics agent writing SQL. A business user asks for last quarter's churn and the agent generates the query. Here I breaks first. The agent finds three columns called status and two tables that both look like customers, and with no glossary to say which definition of churn the company uses, it picks one. The answer is syntactically perfect and quietly wrong, which is the pattern we described in why text-to-SQL breaks in the enterprise.

An agent acting through MCP. An agent connected over Model Context Protocol does more than read. It may create assets, update records or trigger workflows. Now A and G carry the weight. The agent needs its own identity and scoped permissions under A1.2, and the organization needs an approval and an audit trail for what the agent was allowed to touch. An agent running under a shared service account is hard to tell apart in an access review, however FAIR the underlying data is.

Making FAIR-G operational

The good news is that FAIR-G does not ask for a new kind of system. Almost every letter is satisfied by metadata that a governed catalog already manages, and the original FAIR paper itself is mostly a list of metadata requirements. What changes with agents is that this metadata has to be complete, current and readable by software, because there is no longer a person in the loop to fill the gaps.

In Dawiso the mapping is direct. The data catalog indexes assets across platforms, which covers F. The business glossary gives terms a definition and an owner, which covers I. Lineage records where data comes from and where it goes, which covers the provenance half of R. Ownership, classification and approval workflows on data products supply the governance layer, so access is requested and approved through a recorded process rather than granted by hand.

The last step is delivery. The Dawiso Context Layer serves that governed metadata to AI agents through the Dawiso MCP server, so an MCP-compatible agent can look up what a term means, who owns a dataset and where a number came from at the moment it needs to. That is agent-actionability in the FAIR4AG² sense, applied to the data an enterprise already has.

If you want to check where you stand, start with the FAIR4AG² test and ask of each important dataset whether it is met for agents or only for humans. Our seven-point AI-ready data checklist turns the same idea into measurable checks, and the FAIR data principles glossary entry covers the four principles in more depth.