How long does the assessment take?

About 10 minutes. Thirty multiple-choice questions across five sub-dimensions of data readiness for GenAI.

What are the five sub-dimensions?

Catalog & Lineage, Data Quality, Access & Permissions, Retrieval & Vector, Compliance & Governance. Each gets 6 questions and a per-sub-dimension score.

How does this relate to the GenAI Maturity Framework?

This is the operational counterpart to the Data & Infrastructure dimension of the framework. Data readiness is one of the most common blockers to scaling GenAI beyond pilots.

What if we do not use RAG?

The catalog / quality / access / compliance dimensions still apply. If a specific retrieval-focused question doesn't apply, choose the lowest score to represent the absence — the tool interprets that correctly.

Are my answers stored anywhere?

No. Nothing is submitted or logged. The quiz runs entirely in your browser; answers persist to localStorage on your device only.

GenAI MaturityGAI

Framework Try Assessments Use Cases Industries Implementation Learning Paths Blog Resources

Home
›
Tools
›
Data Readiness

Data Readiness Checker for GenAI

A free 30-question assessment of whether your data platform is ready to support enterprise GenAI at production quality — across catalog & lineage, data quality, access & permissions, retrieval & vector, and compliance & governance. About 10 minutes. Nothing submitted.

30 questions · max score 90 · operational counterpart to the Data & Infrastructure framework dimension

Progress

0 / 30

Catalog & Lineage

Is your data findable and traceable — with clear ownership and downstream lineage?

1. How does someone in your org find a dataset relevant to a new use case?
Slack search and tribal knowledge.A wiki page or Confluence, occasionally updated.A data catalogue tool covering major datasets.A catalogue with every production dataset, ownership, freshness SLA, and downstream consumers.
2. What percentage of production datasets have a documented owner?
Unknown / near zero.Some, but ownership is inconsistent across teams.Most; a few orphans exist.100% with named accountable role and rotation policy.
3. Can you trace a downstream metric back to its source tables and transformations?
Not without significant reverse engineering.Manually possible for key metrics.Automated lineage in place for most pipelines.Column-level lineage, machine-queryable, updated on every schema change.
4. How is training-corpus provenance tracked for GenAI use cases?
It is not.Tracked per-project in slack / docs.Tracked in the catalogue with source, license, and version.Automated corpus-provenance manifest generated per training / fine-tune run and preserved for audit.
5. How stable are dataset schemas?
Change frequently without notice; downstream breaks are common.Schema changes are usually announced in advance.Schema changes are governed by contracts (breaking vs non-breaking) and reviewed.Data contracts enforced in CI; producers can't ship breaking changes without consumer opt-in.
6. How is data-source deprecation handled?
Sources decay quietly; consumers discover it when queries return stale data.Deprecation announced with a deadline.Deprecation with migration path and dashboard tracking cutover.Deprecation automated with usage-graph analysis, migration paths, and forced-cutover after grace period.

Data Quality

Do you know when data breaks, and can you trust what you retrieve?

1. How do you detect data-quality issues in critical pipelines?
Downstream users complain.Manual dashboards, checked ad-hoc.Automated data-quality tests (Great Expectations / Soda / dbt tests) on critical tables.CI-gated quality tests + streaming anomaly detection + auto-quarantine of bad partitions.
2. How are freshness SLAs defined and monitored?
Freshness is implicit and undocumented.Defined for a few key tables; monitoring is manual.Defined per-table with automated alerts on SLA breach.SLA + SLO framework with per-consumer subscription; breaches trigger paging and incident channels.
3. How do you handle duplicate / conflicting records across sources?
Not systematically — every consumer solves it locally.Dedup rules exist for key entities but drift over time.Master data management for critical entities (customer, product) with governance.MDM plus provenance chain — every entity has a canonical source and merge rules with tests.
4. How is metric definition drift prevented?
It happens regularly; metrics quietly differ across tools.A metric dictionary exists but isn't enforced.A semantic layer / metrics store computes shared metrics for all consumers.Enforced semantic layer with versioned metric contracts and change reviews.
5. For RAG systems, how do you measure retrieval quality?
We don't.Anecdotal — users report bad retrievals occasionally.Recall@k / MRR measured periodically on a labelled query set.Live streaming eval of retrieval quality with alerting; golden query set grows from real incidents.
6. How is training / eval data curated for quality?
Whatever's available — no curation.Filtered by rule (e.g., min length, dedup).Human review of sampled records + quality classifier for large corpora.Structured curation pipeline with per-record quality scoring, provenance, and rejection rationale preserved.

Access & Permissions

Can the right people (and agents) reach the right data without a ticket?

1. How does a new team get access to relevant datasets?
File a ticket, wait days-to-weeks.Ticketed but with an SLA (24-48h).Self-service portal for most datasets with policy-based auto-approval.Attribute-based access — new team member inherits access based on role + project immediately on join.
2. How is access to sensitive data (PII, financials, HR) controlled?
Ad-hoc — access is granted broadly to speed work.Sensitive data flagged; access requires manual approval.Sensitive data isolated in a governance zone with role-based access.ABAC / column-level policies enforced at query time; access is time-bound and audited.
3. How do LLM agents authenticate when reading data on a user's behalf?
Shared service account for all agents.Per-agent service accounts.Per-user token exchange for LLM calls; per-request attribution.Fully user-scoped OAuth / MCP-based flows with least-privilege scopes and audit trail.
4. How are access reviews conducted?
They aren't.Annual, spreadsheet-based.Quarterly, automated pull from IdP with sign-off workflow.Continuous — unused access decays automatically; anomalous access flagged in real time.
5. How is production data used for testing / development?
Copied to dev environments as-is.Copied with basic PII redaction.Synthetic datasets generated from production statistics; PII prohibited in dev.Data-clean-room pattern — dev workloads query production via policy-enforced views only.
6. How is third-party (vendor, partner) data access governed?
Ad-hoc — we generally share what they ask for.Contract-driven, but implementation varies per vendor.Standard data-sharing agreements + monitored access boundary.Data clean rooms / privacy-enhancing tech for regulated categories; usage auditable end-to-end.

Retrieval & Vector

Is retrieval infrastructure (vector store, hybrid search) in a shape that supports production RAG?

1. Where do your RAG systems store embeddings?
No RAG systems in production.Local files / per-app databases; no shared substrate.A shared vector store used by multiple teams.Shared vector store treated as first-class infrastructure — SLOs, observability, tenant isolation.
2. How are chunks and embeddings kept fresh as source documents change?
Full re-indexing on a schedule; sometimes stale.Full re-index nightly.Incremental updates on source change, with lag under 1 hour for critical corpora.Event-driven — source change triggers re-embed + re-index within minutes; lineage preserved.
3. What retrieval strategy do your RAG systems use?
No RAG.Pure vector similarity.Hybrid search (vector + BM25 or full-text) with basic re-ranking.Hybrid retrieval + cross-encoder re-ranking + query decomposition for multi-hop questions.
4. How do access controls carry from source data into the vector store?
They don't — vector store has broader access than source.Manual policy per corpus.Metadata-filtered retrieval enforces per-user access.End-to-end propagation — vector store inherits source ACLs at ingest and enforces at query time.
5. How is prompt caching used in retrieval-heavy workloads?
Not used; every request re-sends full context.Ad-hoc — some teams use it, most don't.Standard practice for stable prefixes (system prompt, tool schemas).Cache-aware prompt design across the org; cost dashboards show cached-input ratio per system.
6. How do you handle document versioning in retrieval?
Latest version overwrites; no history.Version-aware ingest but retrieval always returns latest.Time-travel retrieval available for specific use cases (compliance, audit).Full temporal RAG — queries can specify "as of" and retrieve the corpus state at that time.

Compliance & Governance

Do PII redaction, retention, cross-border, and DSR flows meet regulatory expectations?

1. Where does PII redaction happen in your GenAI pipelines?
It doesn't, or per-app inconsistently.Redaction rules exist per app.PII detection + redaction at the platform gateway.Platform-layer PII detection + configurable redaction policies per data class per tenant.
2. How are data-subject deletion requests handled for GenAI-adjacent data (prompts, completions, embeddings)?
They aren't.Manual process; unclear whether embeddings are covered.Automated propagation covers primary stores and embeddings.End-to-end DSR pipeline covers prompts, completions, embeddings, backups; evidence produced automatically.
3. How is cross-border data flow controlled for GenAI systems?
Not tracked; regional constraints often violated.Documented per region; enforcement varies.Regional tenancies with automated routing based on data classification.Full data-residency framework with automated blocking of prohibited flows and audit trail.
4. How is retention defined for prompt / completion logs?
Not defined.Defined per-app.Standard policy across GenAI systems; enforced automatically.Retention policy varies by data class per record; enforced with automated purge + audit.
5. For any high-risk AI system, do you have training / eval data documentation ready for regulatory audit?
No — we would struggle to assemble it.Partial documentation exists.Documentation exists per system, maintained per release.Auto-generated on every release from lineage + evaluation harness; ready for auditor on demand.
6. How is copyright / licensing status verified for training / retrieval corpora?
Not verified.Verified per project on ingest.Verified per source at catalogue time with license metadata.License metadata enforced in query time — non-permitted content excluded automatically; usage audit-ready.

FAQs

How long does the assessment take?
About 10 minutes. Thirty multiple-choice questions across five sub-dimensions of data readiness for GenAI.
What are the five sub-dimensions?
Catalog & Lineage, Data Quality, Access & Permissions, Retrieval & Vector, Compliance & Governance. Each gets 6 questions and a per-sub-dimension score.
How does this relate to the GenAI Maturity Framework?
This is the operational counterpart to the Data & Infrastructure dimension of the framework. Data readiness is one of the most common blockers to scaling GenAI beyond pilots.
What if we do not use RAG?
The catalog / quality / access / compliance dimensions still apply. If a specific retrieval-focused question doesn't apply, choose the lowest score to represent the absence — the tool interprets that correctly.
Are my answers stored anywhere?
No. Nothing is submitted or logged. The quiz runs entirely in your browser; answers persist to localStorage on your device only.

Related on this site

Framework dimensions

Data & Infrastructure

Free tools

LLM Model Comparison

Whitepapers

Data Readiness for GenAI

Comparisons

RAG vs Fine-tuning vs Prompt Engineering: Decision Framework

Next steps

LLM Evaluation & Assurance— Metrics, benchmarks, red-team
LLM Evaluation Playbook— Six-phase implementation guide