Is your data actually ready to ground production AI?
Most stalled AI programmes are content and entitlement problems wearing a model costume. Score six dimensions that decide whether grounded answers hold up in production.
Score each dimension
Answer as your organisation operates today, not as it is described in a strategy deck. Scores update the result live.
Is the content AI will cite accurate, current and owned by someone accountable?
1 = stale wikis, no owners · 5 = owned content with review cycles and freshness SLAs
Can AI systems reach systems of record through supported APIs rather than exports?
1 = manual extracts and batch files · 5 = supported real-time APIs across core systems
Is there a shared retrieval layer — chunking, embeddings, ranking — rather than per-project ad hoc pipelines?
1 = each team builds its own · 5 = shared, evaluated retrieval service
Are completeness, accuracy and duplication measured with thresholds and alerting?
1 = quality issues found by users · 5 = profiled data with quality SLOs and alerts
Does retrieval respect per-user permissions so answers never leak restricted content?
1 = single index, no permission filtering · 5 = permission-aware retrieval enforced at query time
Can you see what was retrieved and cited for any answer, and trace it back to a source?
1 = no citation tracing · 5 = full retrieval and citation traces per answer
Good enough for production in defined domains. Invest in quality SLOs and citation observability as usage grows.
Weighted average 3.00 of 5. Scores stay in the page link and the PDF — nothing is stored and no email is required.
Your prioritised next 90 days
Generated from your three weakest weighted dimensions.
- 01Entitlements — scored 3/5
Enforce permission-aware retrieval at query time using source-system entitlements, and test with restricted-content probes before launch.
- 02Knowledge quality — scored 3/5
Identify the top 100 documents an assistant will cite, assign owners, set review cycles and retire contradictory duplicates before indexing.
- 03Access & integration — scored 3/5
Expose the three systems of record the use case depends on through supported APIs with test environments and rate limits documented.
Entitlements, knowledge quality and system access carry the highest weight because they are the failure modes that stop pilots from reaching customers.
Common questions
- Why do AI programmes stall on data rather than models?
- Models are commoditised; grounded, current, permissioned enterprise content is not. Most failed pilots trace back to stale knowledge, missing entitlements or no reliable retrieval layer.
- Do we need a data warehouse before starting?
- No. Retrieval-based use cases need governed content and APIs more than a central warehouse. Score access and retrieval honestly and start where the content is already trustworthy.
- What is the quickest score to improve?
- Knowledge freshness. Assigning owners and review cycles to the top 100 documents an assistant will cite usually lifts answer quality faster than any model change.