What Makes Data AI-Ready?
Key takeaways
- Four conditions make data AI-ready: reachable, correct, permitted and scoped. The first three are usually met.
- Readiness checklists treat each term as having one meaning that cleaning will reveal.
- Deduplication assumes two records describing the same thing should merge. Lineage assumes a field keeps its meaning along the path.
- Both assumptions hold inside a domain and break across domains, which is where an agent operates.
- Four systems in a hospital group hold four precise definitions of patient, each correct for its purpose.
- Test readiness with four questions about one contested term, and the gap shows up within an hour.
What makes data AI-ready? Most readiness checklists answer with volume, cleanliness and access. Those matter and they are solved problems in a well run estate. They also do not predict whether an agent will answer correctly.
Data is AI-ready when an agent can tell which definition applies to the question it was asked.
What makes data AI-ready?
Four conditions, in order of how often they get met.
- Reachable. The agent can query it. Usually true.
- Correct. The values are complete and current. Often true.
- Permitted. The agent is entitled to read it. Usually true.
- Scoped. Each term carries its definition and the boundary that definition holds within. Rarely true.
The first three decide whether an agent can produce an answer. The fourth decides whether the answer addresses the question.
What do the usual readiness checklists cover?
Infrastructure and hygiene. Consolidated storage, documented lineage, deduplicated records, labelled training sets, retention rules, access policies.
Every item is worth doing. A team that completes the list has a well run estate and a real asset.
What do those checklists miss?
They treat each term as having one meaning that cleaning will reveal. Deduplication assumes two records describing the same thing should merge. Lineage assumes a field keeps its meaning along the path.
Both assumptions hold inside a domain. Both break across domains, which is exactly where an agent operates.
What does scoped meaning look like in practice?
Take patient inside a hospital group. Four systems, four precise definitions.
| Function | A patient is | Why |
|---|---|---|
| Clinical | A person with an open episode of care | Treatment and handover |
| Registration | A person holding a medical record number | Identity across visits |
| Billing | A guarantor account | Claims and payment |
| Research | A consented study subject | Protocol and ethics |
Ask how many patients the group treated last month and a clinician asks you which sense you mean. An agent produces a number.
How do you test whether data is AI-ready?
Pick a contested term and ask four questions about it.
- How many definitions of this term are in use right now.
- For each one, which systems, processes and reports does it govern.
- Who may change each definition.
- How do the variants combine into the figure that gets reported.
If all four answers exist in a system an agent can read, that term is AI-ready. If they live in people's heads, it is not, whatever the data quality score says.
What does the test tell you?
Usually that the estate is in good shape and the meaning layer is empty. That is the common result and it is a useful one, because it moves the work from a vague readiness programme to a specific list of terms.
It also reframes the budget conversation. Another cleaning cycle will not change the result of this test.
How do the patient definitions reconcile?
- Scope. Clinical patient governs care reporting. Registration patient governs identity. Billing patient governs claims. Research patient governs study enrolment.
- Relationship. Registration is the superset. Clinical and billing are overlapping views of it. Research is a consented subset.
- Reported figure. Monthly treated volume uses the clinical definition, with the day case rule named.
With that written down an agent can answer precisely. Data governance for AI covers the underlying argument, enterprise metadata management covers where the records live, and data observability covers the quality side.
So what should you take away?
Your data may be clean, governed, documented and still unready. Readiness for an agent is a question about meaning.
Run the four questions against one contested term this week. The gap will be obvious within an hour.
TopQuadrant builds the systems that hold definitions, scope and ownership together. Test one term before you fund another cleaning programme.
