How to create a data dictionary with AI assistance

A data dictionary created with AI assistance describes what columns mean before anybody starts calculating. It records what each field represents, which values it accepts and how incomplete situations should be interpreted.

AI can organize this document and identify ambiguities, but somebody familiar with the source must confirm the meaning. Our exercise uses a fictional support-request table without any real customer data.

Start with the unit represented by each record

Write what one row represents: an open request, an interaction or a customer. Three rows belonging to one customer can mean three requests rather than three different people.

That detail changes counts, rates and subsequent questions. Before sharing a sample, check which fields the exercise needs and remove identifiers that do not help explain its format.

In this example, every row is a request identified by its own code. The opening date belongs to that request, while the category describes the subject recorded during triage.

Build an initial column description

Use the following table as a first draft. Its values are invented to show how a definition can include format, units and a missing-value rule.

Field	Meaning	Proposed rule
request_id	Request identifier	Required, unique text
opened_on	Opening date	Date formatted as YYYY-MM-DD
duration_min	Support duration	Number of minutes; empty means unrecorded
category	Request subject	List of approved categories

The W3C’s data best practices address metadata that helps people understand and reuse datasets. A dictionary applies that idea to individual columns alongside your team’s specific decisions.

Do not define an approved category list by copying every text value found in the file. A common spelling mistake remains a quality issue rather than automatically becoming an accepted category.

Ask AI for questions before filling gaps

Provide a small sample and explain how the data were collected. Request preliminary definitions that visibly distinguish confirmed information from assumptions.

Help create a dictionary for this fictional support table.
Each row represents a request rather than a customer.
For every column, propose meaning, type, unit, permitted values and missing-value policy.
Separate supplied information from details needing confirmation.
Do not invent official lists, deadlines or completion rules.
List questions for the person responsible for collection.

If AI concludes that a missing duration means support has not started, request evidence. Absence might also result from an export problem, optional collection or an event still in progress.

Resolve ambiguities with the source owner

Ask who fills in the field, when they do it and why the information is collected. The answer may reveal that “date” means opening in one report and completion in another.

Record whether a field is required for every row or only in specific situations. This prevents a legitimate exception from being classified as a quality failure.

When two teams use the same name with different meanings, retain distinct names or explain the scope. Unifying labels should not erase an important difference in the process.

Validate the document against examples

Choose several rows and check whether somebody else can interpret them using only the dictionary. Include a missing value, an unusual record and a category requiring clarification.

Your completion criterion is straightforward: each column has a verifiable meaning, and remaining questions are identified. An elegant description that cannot distinguish absence from zero is still incomplete.

Ask the reviewer to explain one calculation they could make from those definitions. If you disagree about the denominator or period, return to the affected fields before beginning analysis.

Avoid three common shortcuts

Do not confuse a technical type with meaning: “number” does not specify money, minutes or quantity. Do not copy rules from another dataset just because its columns have similar names.

Also avoid editing the dictionary without recording the corresponding collection change. Keep an owner, version and review date so older analyses can be interpreted in their original context.

Before calculating, follow the CSV preparation guide and check whether the file follows the approved definitions.

Frequently asked questions

What is a data dictionary?

It explains field meanings, formats, units and interpretation rules.

Can AI determine every column’s meaning by itself?

It can suggest hypotheses and questions. Someone familiar with collection must confirm definitions.

How can I validate the dictionary?

Ask another person to interpret examples, including missing and unusual values, using only approved definitions.