Skip to content

Library · Data Practice OS

What AI-ready means

Reference guide · Published · Source reviewed

A human analyst copes with ambiguous column names, stale extracts and a quiet word with the person who built the report. A model does not. It will use whatever it is given, confidently. AI-ready data is data a model or agent can use safely without a human explaining it.

That is a higher bar than "it's in the warehouse". It applies to data products, not to platforms.

The seven properties

# Property What it means How to test it
1 Described Tables and columns have business descriptions; metrics are defined in the Semantic layer with synonyms Can an LLM answer "what does reg_stat_cd = 'S' mean?" from metadata alone?
2 Owned A named team and product owner, reachable, accountable for quality Is there an owner in the catalog, and do they answer?
3 Contracted Schema, semantics and quality rules are a Data contract, enforced in CI Does a breaking change fail a build before it reaches consumers?
4 Trustworthy Quality checks run every load; Data observability alerts on freshness and volume Would the platform tell you before a user does?
5 Classified and governed Every column carries a Data classification; masking and row policies apply by tag Can you prove an agent running as user X cannot see what X cannot see?
6 Traceable Column-level Lineage from source to answer Can you explain an AI answer back to source records?
7 Reachable A stable interface: governed SQL, a semantic API, a search index or an MCP tool, not a file share Can an agent reach it without a bespoke credential?

Properties 1, 5 and 6 matter most for AI. They are also the ones legacy estates most often lack.

Quick triage of a dataset

Score each property 0 (absent), 1 (partial) or 2 (in place, automated). This is a five-minute triage: it shows where the gaps are and which product to work on first. It is not a readiness decision, and it has no pass mark. A zero on property 5 or 6 means AI use stays in a sandbox with a Human in the loop until the gap is fixed.

Readiness belongs to a named AI use, not to a dataset. Before a pilot or production release, run the AI readiness assessment: 54 scored controls and five gates (rights and handling, access isolation, context sufficiency, fitness for purpose, operational control), scored in AI readiness scorecard (workbook v2). Its gates decide readiness; no score can override them.

Worked example: vehicle registrations (illustrative). Described 1 (columns documented, no metric definitions). Owned 2. Contracted 0. Trustworthy 1 (row counts only). Classified 1 (PII tagged by hand, not by policy). Traceable 0. Reachable 2. The triage points at the fastest wins: define the five core metrics in the semantic layer, publish a dbt model contract, and move classification to tag-based policy. All three are platform-agnostic work. The zero on traceability keeps any AI use in a sandbox until it is fixed.

What AI-ready is not

  • Not a platform feature. Cortex Analyst and Genie are only as good as the semantic model under them. See The semantic layer.
  • Not "all the data in one place". Centralising undocumented data creates a bigger undocumented pile. Start with the handful of products the first AI use cases need.
  • Not only structured data. Policies, procedures, inspection reports and correspondence are often the most valuable AI input. They need the same ownership, classification and freshness as tables. See RAG and enterprise search.
  • Not perfection. A product with known gaps and an evaluation set beats a "complete" model nobody has tested.
  • Not a high score. A high maturity score never makes up for a failed readiness gate.

Where this shows up next