AI governance
Your data dictionary stops at the column
Data dictionaries exist because one word can mean four things. They cover columns in a database. The policies, contracts and procedures an AI assistant actually reads carry the same ambiguity, with nothing pinning it down.
Cognatum Team · Nov 16, 2026 · 6 min read
Somebody asks the assistant a simple question. What counts as an active account? The answer comes back in one clean sentence, drawn from a document a colleague wrote last year.
Cognatum governs the entry
source · version · approver · permissions
The word active appears in four places across the company. The billing team means one thing by it. The service team means another. The contract means a third.
The answer picked one of them. It did not say which one, and nobody asked.
There is already a tool built for exactly this problem. It just does not reach this far.
What a data dictionary does
A data dictionary is a central record of what each piece of data means. The reference definition, adapted from the IBM Dictionary of Computing, describes a repository covering meaning, relationships to other data, origin, usage and format.
The field has long distinguished between two kinds. A passive dictionary is maintained by hand and drifts out of step with the system it documents. An active one is wired into the database and updates when the structure does.
One word, one meaning
The purpose is narrow and useful. A shared dictionary gives everyone working with the same data a common vocabulary, so the meaning, relevance and quality of each element are the same for every user.
Federal science guidance puts the same weight on upkeep. When structures change, the dictionary changes with them, and a dictionary left out of step with the data it describes is treated as a sign of weak stewardship.
The column nobody expects
Open Science Framework guidance on building one asks for something subtle. Record the synonyms for a variable, and make them reflect the meaning as you use it, not the meaning the same word might carry in a different context.
That instruction exists because the authors knew the problem. A word that is precise inside one team is ambiguous the moment it leaves. So you write down which sense you meant, and you sign the document that says so.
The discipline stops at the field
Everything above describes structured data. Variable names, data types, allowed values, units, minimum and maximum. A dictionary pins a term to a column in a table, and the pinning holds because the column is a fixed thing.
Prose was never in scope
Most of what a company knows is not in a column. It is in policies, procedures, contracts, decision memos, service notes and onboarding guides. Those documents use the contested words freely and define almost none of them.
No standard to fall back on
There is no universal standard for how much detail such a document must carry. That is survivable when the dictionary covers a database somebody owns. It is a different matter when the vocabulary in question is the whole organization's.
When the definition itself moves
Research data managers go further and keep a crosswalk, a table mapping how variables changed from one version of a dataset to the next. The point is that a definition has a history, and the history has to be readable.
What an assistant reads
A retrieval assistant does not query the dictionary. It reads the documents. The definitions it works from are the ones embedded in sentences, written by different authors, at different times, for different readers.
Nobody is being careless here. The assistant is doing the job it was given, which is to find the relevant passage and summarize it. The passage simply does not announce which of the four meanings its author had in mind.
Two sources, two meanings
Researchers surveying how language models handle conflicting knowledge name this case directly. Inter-context conflict is what happens when two retrieved sources disagree. The model does not flag the disagreement. It produces one answer.
So the ambiguity a dictionary was invented to remove is reintroduced at the last step, by the tool everyone now treats as the fastest route to an answer.
No record of which one
Provenance, in the standard security definition, is the chronology of origin, development, ownership, location and changes to a thing and its associated data. An answer that names no source and no version carries none of that.
Regulators asked for this once already
Banking supervisors made the argument after the financial crisis that began in 2007, when many banks, including the largest, could not aggregate their risk exposures and identify concentrations fully, quickly and accurately.
The published principles set out to strengthen risk data aggregation capabilities and internal risk reporting practices. That is a real discipline with real supervisory weight, and its subject is data moving through systems.
The AI rules follow the same line
European AI legislation on data and data governance attaches its requirements to the training, validation and testing datasets behind a high risk system. The obligation is real, and it points at the model's inputs.
The documents sit in between
Neither regime reaches the company's own written knowledge at the moment an assistant answers from it. One governs the warehouse. The other governs the training set. The policy on the shared drive falls between them.
What closes the gap
An answer needs the furniture a column already has. One approved definition of the term. A named person who approved it. A date. The source it came from, at the version that was live when the answer was given.
None of this requires moving the documents. They can stay in the departmental drives and repositories where they already live. What is missing is not a new destination. It is a governed layer sitting above the ones you have.
Conflicts surfaced, not settled
Where two documents disagree, the useful behavior is to show both and say where they part, rather than quietly choosing. Two divisions can legitimately hold two policies. That is a conflict to link, not an error to delete.
A person signs, and gets told
The finding, gathering and surfacing can be automated. The decision about what is true is handed to a named human, who can rework the wording before approving it. When a source later changes, the system notifies rather than rewriting.
Cognatum is built around that loop, with a step for wording that must not change. Your company's knowledge isn't missing. It's unusable. Cognatum changes that.
The short version
A data dictionary was the right idea applied to the narrow half of the problem. The words that cause the most trouble are not in your tables. They are in your documents, and now something is answering out of them.
The fix is not more documents. It is one approved meaning per term, a person who stands behind it, and a record of which version an answer used. That is ordinary governance, moved one step closer to where people ask.
Sources
- Data dictionary (en.wikipedia.org)
- Data Dictionaries, Data Management (usgs.gov)
- What is a Data Dictionary? (datamanagement.hms.harvard.edu)
- How to Make a Data Dictionary (help.osf.io)
- Knowledge Conflicts for LLMs: A Survey (arxiv.org)
- Provenance, Computer Security Resource Center glossary (csrc.nist.gov)
- Article 10: Data and Data Governance (artificialintelligenceact.eu)
- Principles for effective risk data aggregation and risk reporting (bis.org)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
What is a data dictionary?
A data dictionary is a central record of what each data element means, covering meaning, relationships to other data, origin, usage and format. It typically lists variable names, human readable names, data types, units, allowed values and a definition for every field, so that everyone working with the data shares one vocabulary.
Does a data dictionary cover documents as well as databases?
In practice, no. Dictionaries are built around fields in a database or a dataset, and the discipline assumes a fixed column to pin a term to. Policies, procedures, contracts and memos use the same contested terms in prose, usually without defining them, and they sit outside the dictionary's scope.
What happens when an AI assistant finds two documents that define a term differently?
Research on knowledge conflicts calls this inter-context conflict, where two retrieved sources disagree. A typical assistant does not raise the disagreement. It produces a single answer built on one of the sources, and the reader has no way to tell which definition was applied.
Do AI regulations require definitions for the documents an assistant reads?
Not at present. European requirements on data and data governance attach to the training, validation and testing datasets behind a high risk system, which is a different artifact from a company's own policies and procedures. Software can supply evidence of how knowledge was governed, but it does not by itself confer compliance.
How do you govern terminology without migrating every document?
By putting a governed layer above the repositories you already have rather than moving the content into a new one. Terms carry an approved definition, a named approver and a date, conflicts between sources are surfaced instead of silently resolved, and when an underlying source changes the system notifies the people who depend on it.