What does "AI-ready data" actually mean?

"AI-ready data" is now one of the most repeated phrases in life sciences data strategy, and one of the least precise. Ask five people what it means and you get five answers: some mean high data quality, some mean data in a lake, some mean data a model has been trained on. That ambiguity reflects the gap between data that looks usable and data an AI agent can actually reason over. This is where most agentic AI projects stall.

This became clear in a recent ONTOFORCE panel on the evolving AI landscape in life sciences, which brought together data and AI leaders, including experts from Sanofi and Roche. Across a wide-ranging discussion, one point recurred: the models are rarely the bottleneck. The data foundation beneath them is. This article works through what AI-ready data means in practice, why data quality alone does not get you there, how the FAIR principles are being pushed to their limit, and what to fix first.

Why isn't clean data the same as AI-ready data?

Data quality is the price of entry for AI-ready data, not the finish line. Clean, accurate, deduplicated data is necessary, but an AI agent needs more than correct values. It needs to know what each value means, how it relates to values in other systems, and where it came from. Quality without context still leaves an agent guessing.

Consider a field that contains a clean, valid string. A human analyst who built the system knows that string is a patient cohort identifier, that it links to a specific trial in another database, and what to do with it. An AI agent inherits none of that. The data can be flawless and still be unusable, because the meaning lived in the head of the person who wrote the query, not in the data itself.

This is the distinction between quality and context, and it is often underestimated. Enterprise data catalogs are typically built for humans browsing a user interface: verbose, descriptive, and readable by a person who already understands the domain. Fed to an agent, those same catalogs burn through the model's context window quickly and return little the agent can act on. Machine-readable context, not human-readable description, is what separates AI-ready data from merely clean data.

What makes data AI-ready in practice?

Data becomes AI-ready when it carries three things beyond quality: semantic meaning, connected relationships, and provenance. Together these let an AI system interpret data, traverse it, and judge whether to trust it, without a human supplying the missing context at every step.

Semantic meaning

The schema itself enriched with definitions, not just data types. Knowing a field is a string is useless; knowing that the string represents a patient cohort identifier that links to a specific clinical trial is what lets an agent reason across datasets. This is the role of a semantic layer, a layer of meaning over raw data that captures business definitions, relationships, and context so that systems, and AI, understand what data represents rather than just storing it.

Connected relationships

When data is modeled as entities and the links between them, rather than as isolated tables, an agent can traverse it. This is what a knowledge graph provides: a structure that an agent or a researcher can follow from a disease to a target to a compound, even when those facts originate in different systems. Reasoning across domains depends on those links existing in the data.

Provenance

Provenance is a record of where data came from and how it was transformed that travels with the data itself. Provenance is what lets an AI system, or a regulator reviewing its output, judge whether a given source is trustworthy. In a regulated industry, data an agent cannot account for is data an agent should not act on.

As Sabya Dasgupta, VP Global Head of Data and AI Platforms at Sanofi, put it during the panel, "context becomes the new compute." His point was economic as well as technical: without a semantic infrastructure sitting on top of the data layer, agentic queries become expensive and inaccurate, because the model has to work to reconstruct context that should have been supplied. With it, compute is spent on reasoning rather than on guessing.

How does AI-ready data relate to FAIR?

AI-ready data builds on the FAIR principles but pushes them further than FAIR was originally designed to go. FAIR data is data that is Findable, Accessible, Interoperable, and Reusable, a widely adopted standard in life sciences since 2016. FAIR established a solid foundation for data used by humans and traditional tooling. Agentic AI raises the bar on each principle.

For data consumed by autonomous agents rather than people, each FAIR principle takes on a stricter meaning. Findable comes to mean machine-discoverable, with enough metadata that an agent can assess relevance before retrieving a full dataset. Accessible comes to mean API-first and low-latency. Interoperable comes to mean aligned to shared ontologies, standardized vocabularies such as MedDRA for adverse events or SNOMED CT for clinical terms, not merely compatible in file format. Reusable comes to mean that context travels with the data, not just the values.

There is an honest debate in the field about whether FAIR alone is sufficient for the agentic era. A recurring theme in the panel was that FAIR moved the industry out of the spreadsheet era, but that the human-driven methods used to achieve it, governance stewards and large manual ontology projects, are difficult to scale to the volume of data products a large organization now manages. The direction of travel is toward generating semantic context and governance as data products are built, rather than curating it after the fact.

What should life sciences teams fix first?

Many life sciences organizations are already investing here. Data leaders at large pharma companies are building metadata enrichment, revamping knowledge graphs to hold multi-omic data, and standing up semantic layers over their existing systems; the panelists from Sanofi and Roche both described active work along these lines. The question for most teams is no longer whether to build a semantic foundation, but how to do it without stalling.

The most effective first step is to establish a semantic layer that connects existing systems, scoped to one real problem rather than the whole organization. The common failure mode is the opposite: drafting a multi-year, enterprise-wide harmonization roadmap before running a single agent workflow, by which time the landscape has moved. Starting narrow and expanding is what the panel's practitioners consistently recommended.

Several converging recommendations came out of the discussion. Start from a specific, high-value question and bring only the data required to answer it into a connected view. Build narrow, fit-for-purpose ontologies rather than one monolithic corporate ontology; small, well-scoped semantic structures can be composed into broader capability over time. And treat the semantic layer as connective tissue between systems, a translator that resolves terminology conflicts and supplies context, rather than as a rip-and-replace of the data infrastructure already in place.

As Victor Neduva, Head of Target Discovery and Disease Data at Roche, framed it, "start with a specific question, solve it with specific tools, and you get to something measurable in a reasonable amount of time." The problem context matters too: what the history of a project is, and what did not work before.

If you are weighing where your own data foundation stands, our semantic layer assessment checklist walks through the criteria that make data AI-ready, from schema context to provenance to ontology alignment. Assess your semantic layer with the checklist here.

This is the problem DISQOVER, ONTOFORCE's AI-ready semantic data platform for life sciences, is built to address. It connects fragmented research, clinical, and regulatory data into a unified, queryable view on a knowledge graph, and serves both the human researcher who needs to explore and trust data and the AI or platform consumer that needs data to be machine-ready and interoperable. Support for the Model Context Protocol (MCP), a standard way for AI models and agents to connect to external tools and data sources, sits under that AI-ready data capability. The value is the same in both directions: data that carries its own meaning and provenance is data an agent can reason over and a scientist can trust.

What should data teams take away from this?

For years the industry has repeated that data is the new fuel. The more useful framing now is that data is not useful without decision advantage, and decision advantage comes from the semantic infrastructure that makes data interpretable, connected, and trustworthy to an AI system. AI-ready data is less about accumulating more data or cleaning what exists, and more about giving data enough context to be reasoned over reliably. That is the work that turns an impressive demo into a workflow a regulated organization can actually deploy.

The full panel discussion goes deeper on agentic AI, the semantic foundation it depends on, and the trust and governance questions that decide what gets deployed. You can watch the complete conversation here.