Enterprise Data Readiness for AI

Verify source ownership, data quality and permissions before selecting a model.

01 / cards

Six readiness layers

Availability

Are required records accessible?

Quality

Are values complete, correct and deduplicated?

Meaning

Do critical terms have shared definitions?

Freshness

How quickly do changes and revocations propagate?

Permissions

Can source-level access be preserved?

Integration

Which owners, APIs, failure modes and schedules exist?

02 / table

Source-of-truth questions

Question	Why it matters
Who owns each fact?	Resolve contradictions instead of averaging them
Are identifiers stable?	Avoid incorrect cross-system joins
Can users trace an answer?	Source version, citation and audit
Do deletions sync?	Permission and retention safety

03 / table

Failure patterns

Pattern	Example
Stale retrieval	Last month’s policy appears current
Permission leak	An unauthorized user sees a private indexed file
Ambiguous join	Customer IDs or units mismatch
Silent sync failure	Index looks available but stopped refreshing

04 / note

Worked example

For CRM + docs + tickets, map source ownership, contradictions, permission scopes and update schedules before selecting retrieval or generation.

05 / list

Readiness questions

Could a normal employee answer using the available records?
Who can access each underlying source?
What happens when an item changes or disappears?
Is the defect in data, retrieval, or generation?

GOING DEEPER

Four source failures a better model cannot fix

Failure pattern and test

Failure	Concrete test
Stale retrieval	An answer cites last month’s superseded policy: compare source timestamp and index freshness
Permission leakage	User sees confidential indexed material: enforce source-user authorization, including revoked access
Ambiguous join	Customer identifiers or measurement units disagree: reconcile owners and definitions before analysis
Silent sync failure	Pipeline stopped updating but search looks healthy: alert on ingestion lag and missing updates

Source-of-truth and lifecycle: Map authoritative owner for each fact and a durable ID for joins. Trace an answer to exact source/version. Make deletions, revocations, conflicts and reindexing observable; verify that access changes propagate to retrieval.

CRM + docs + tickets example: Before choosing the model, define which system answers which question, test contradictory records, verify per-user source access, measure update cadence and inspect citations. This is fundamentally a data-readiness decision.