Description
Artificial Intelligence (AI) models and autonomous agents have become regular consumers of open data. It is no longer just the analyst who downloads a set of data: it is automated systems that consult catalogs, interpret columns and cross sources without human intervention. This change forces publishers to ask themselves a key question: is my data ready to be understood by a machine?
This post is derived from the practical exercise Fuel Price Analysis with GenAI as a co-pilot. Based on the conclusions drawn during its development, we review the three dimensions that determine whether a dataset is truly "consumable" by AI: access formats and mechanisms, semantic context, and privacy.
1. Access formats and mechanisms: how does AI best "consume"?
Structured and flat formats
AI models process data better when it arrives in clean and structured formats: CSV, JSON or Parquet. They are flat, predictable formats that are easy to interpret programmatically.
At the opposite extreme are over-formatted Excel files: merged cells, multilevel headers, colors with implicit meaning, or totals sandwiched between rows of data. All that visual format, useful to a human reader, becomes "noise" for the model, which can misinterpret the structure of the table and generate erroneous results. The recommendation is clear: if it is published in Excel, it should be as a flat table; and better yet, always offer a CSV or Parquet version.
Efficient and intelligent access mechanisms
Beyond the format, the channel matters. Well-documented REST APIs allow AI systems to perform dynamic and accurate queries using "Function Calling": the model translates a natural language question into a particular API call and gets only the data it needs, without downloading the entire dataset.
In this vein, emerging protocols such as MCP (Model Context Protocol) standardize the way models discover and use data sources, reducing integration work and opening the door for autonomous agents to interact with open data catalogs natively.
Balancing Bulk Download
However, access via API is not a substitute for mass downloading: both worlds coexist. For large volumes of information, such as High Value Datasets (HVDs), bulk download is still essential for two reasons:
-
Consumer efficiency: Training or analyzing on an entire dataset is more effective with a single download than with thousands of API calls.
-
Securing Publisher Infrastructure: Continuous access by autonomous agents can overwhelm servers. Offering downloadable bulk files reduces that burden and prevents publishing systems from being compromised by spikes in automated requests.
Incremental accesses
Incremental or delta discharge mechanisms are an essential complement to bulk. If the dataset includes reliable update dates (or endpoints that can be filtered by change date), an agent can perform an initial full download and then download only the new or changed data. The result: less traffic, less load on servers, and always-up-to-date data on the consumer side.
2. Semantic context: making "sense" of columns
AI doesn't guess: readable and actionable data dictionaries
A model can read a column called price, but it cannot know if it includes VAT, in which currency it is expressed or if it is an average or point value. Without that information, the AI will fill in the gaps with assumptions, and assumptions lead to errors.
That's why each dataset should be published accompanied by an explicit data dictionary: what each column means, what units it uses, what values are valid, and what conventions it applies (date formats, decimal separators, null treatment, etc.).
Practical Realization for Publishers
This is one of the most common problems among publishers, and it has specific solutions depending on the format:
-
JSON: Publish a JSON Schema that defines types, required fields, valid ranges, and descriptions for each property.
-
CSV: accompany the file with a metadata sidecar file (e.g. following the W3C's CSV on the Web standard or a simple dictionary in readable format) describing each column.
-
Parquet: take advantage of the fact that the format integrates the schema into the file itself, enriching it with descriptive metadata at the column level.
Normalization and consistency
The column name is the first semantic clue that the model receives. A name like average_price_petrol conveys meaning by itself; a cryptic code like X_VAL_2026 forces you to guess. Using semantic, readable, and consistent names across datasets—same convention, same language, same identifiers for the same concepts—facilitates both AI consumption and human reuse.
3. Privacy and security: the essential pre-filter
Anonymization at source
Before a piece of data interacts with any AI environment, it must have passed a strict privacy filter. The deletion and masking of Personal Data (PII) is not a recommendation: it is a pre-publication obligation. Names, addresses, personal identifiers or combinations of fields that allow individuals to be re-identified must be processed at source, never trusting that the end consumer will filter them.
Governance and Leak Prevention
There is also a critical risk on the side of internal use: uploading confidential or strategic information to open business models. Many of these services can reuse the data entered for the retraining of their models, so that a sensitive document could end up being part of the knowledge of a system accessible by third parties.
The answer lies in governance: defining clear policies on what information can be processed with AI and under what conditions, and relying on corporate environments and watertight and secure infrastructures, where data does not leave the organization's perimeter or is used to train external models.
Conclusion
Preparing open data for AI doesn't require starting from scratch: it requires rigorously applying good practices that most of us already know. Clean formats and access mechanisms that combine APIs, bulk download, and incremental updates; explicit semantic context that eliminates ambiguity; and a privacy and governance filter always applied at source.
It should be emphasised: none of these recommendations are new. All of them were already part of the repertoire of good practices of any analyst or data scientist. The difference is that they are no longer optional: they are a necessary condition if we want AI models and agents to work as co-pilots with a high degree of autonomy over our data.
Publishers who adopt these practices will not only make the work easier for AI models and autonomous agents, they will multiply the value and reach of their data. Because a piece of data that a machine can understand is a piece of data that can generate knowledge on a scale that was previously unattainable. The time to adapt our catalogues is now.
Content created by Alejandro Alija, expert in Digital Transformation and Innovation. The content and views expressed in this publication are the sole responsibility of the author.
Comments