AcademyUnderstand

AI-Ready Data

How to assess data readiness for a specific AI use case, design retrieval that is accurate and permission-aware, and govern the data products and contracts that agents rely on.

U201Practitioner12 slidesAbout 10 minutes

Start the course Download the PDF
1 / 12
PDF
AxiareteAcademy

U201Understand · Practitioner

AI-Ready Data

Preparing enterprise data for retrieval, agents and models: fitness, context, lineage, permission and access.

FormatKnowledge pack · 12 slides
ForData and AI leaders
LevelPractitioner
Axiarete Academy · U20102 / 12

The premise

Data is ready for a use case, not in general

The same customer table can be adequate for a dashboard and inadequate for an agent that issues refunds. Readiness depends on what the AI will do, how much error is tolerable, and who may see the result.

DecisionThe action the output will drive.
ToleranceThe cost of a wrong answer.
AudienceWho is entitled to see what.
AI-Ready DataUnderstand
Axiarete Academy · U20103 / 12

Why it matters

Lack of AI-ready data puts AI projects at risk

60%

of AI projects unsupported by AI-ready data, Gartner predicts, will be abandoned through 2026.

Gartner prediction, February 2025

63%

of organizations either did not have, or were unsure whether they had, the right data management practices for AI.

Gartner survey, reported February 2025

Use case

In Gartner’s definition, AI-ready data “must be representative of the use case, of every pattern, errors, outliers and unexpected emergence that is needed to train or run the AI model for the specific use.”

Gartner, February 2025

AI-Ready DataUnderstand
Axiarete Academy · U20104 / 12

Assessment

Five dimensions of readiness

FitnessAccurate, complete, timely and representative enough for the decision it supports.
ContextDefinitions, units and business meaning a model can use.
LineageWhere the data came from and how it was transformed.
PermissionWho may use it, for which purpose, under which obligations.
AccessProgrammatic access at suitable latency, cost and reliability.
How to use itScore each dimension against the use case. The weakest dimension sets the limit on what the AI may do.
AI-Ready DataUnderstand
Axiarete Academy · U20105 / 12

Two kinds of data

Structured and unstructured data fail differently

Structured dataUnstructured content
Typical failureInconsistent definitions, missing keys, stale extractsDuplicates, outdated versions, drafts and AI output mixed with approved policy
Hidden riskJoins that silently drop or duplicate recordsSensitive data embedded in documents and tickets
Missing metadataBusiness definitions and unitsOwner, version, effective date and permissions
RemedySemantic layer, data contracts, quality checksCuration, versioning, permission-aware indexing
AI-Ready DataUnderstand
Axiarete Academy · U20106 / 12

Retrieval

Retrieval is a pipeline, and every stage can fail

IngestParse sources with structure intact; capture owner, version, effective date and permissions.
IndexCombine vector and keyword indexes; add a graph index for corpus-wide questions.
RetrieveHybrid search and re-ranking, pre-filtered to what both the user and the agent may see.
GenerateAnswer only from retrieved sources, cite them, or decline to answer.
EvaluateTest retrieval recall, then groundedness and completeness, on real questions.
The ruleFilter by permission at retrieval. Never rely on a model to withhold what it has already been given.
AI-Ready DataUnderstand
Axiarete Academy · U20107 / 12

Failure modes

Why retrieval answers go wrong

Stale documents outrank current ones
Index effective dates; exclude superseded versions
Chunking breaks tables and procedures
Split on headings; keep tables and steps whole
Duplicates crowd out the right answer
Deduplicate and mark the authoritative version
Policies from other regions are mixed in
Filter by jurisdiction and business unit
Users see content they are not entitled to
Enforce current source permissions; fix oversharing
Answer quality is not measured
Rerun a known-answer evaluation set after every change
AI-Ready DataUnderstand
Axiarete Academy · U20108 / 12

Data products

Data products give agents a contract they can rely on

Publish the terms below as a machine-readable data contract, such as the Open Data Contract Standard, so pipelines and agents can check schema, quality and service levels before use.

OwnerA domain team accountable for quality and change.
Schema and meaningFields, definitions and units, published and versioned.
Quality objectivesFreshness, completeness and accuracy targets, monitored.
Access policyWho may use it, and for which purposes.
LineageUpstream sources and transformations, captured automatically.
Change processVersioning and deprecation consumers can plan for.
AI-Ready DataUnderstand
Axiarete Academy · U20109 / 12

Governance

Govern the data that AI uses

ClassifyLabel sensitive and regulated data before it is indexed.
Limit purposeRecord the purposes for which each data set may be used.
Propagate deletionDeletion and retention must reach indexes, embeddings, caches, logs and fine-tuning data.
LogKeep prompts, sources and outputs for audit; restrict and expire logs as the sources require.
Control providersConfirm whether providers train on your data, how long they retain it, and where.
MinimizeRetrieve only the fields a task needs.
Review accessInclude agent identities and indexing connectors in every access review.
MonitorAlert on unusual retrieval volumes and patterns.
AI-Ready DataUnderstand
Axiarete Academy · U20110 / 12

Method

Assess one use case in a two-week time box

1DefineThe decision, the action and the tolerance for error.
2ListThe data the use case needs, field by field.
3ProfileQuality against the thresholds the decision requires.
4CheckPermissions, purpose compatibility, lawful basis and regulatory constraints.
5TestAn evaluation set built from real questions and cases.
6BaselineRun it, record failure modes, fix the largest first.
AI-Ready DataUnderstand
Axiarete Academy · U20111 / 12

Measurement

Readiness measures for each use case

FreshnessLag between a change at source and its availability, against the requirement.
Critical completenessShare of records with every field the use case needs.
Definition coverageShare of fields with an agreed business definition.
Permission coverageShare of indexed documents whose permissions match the source.
GroundednessShare of answers fully supported by the retrieved sources.
Exposure incidentsCount of answers that exposed content the user was not entitled to see.
AI-Ready DataUnderstand
Axiarete Academy · U20112 / 12

Summary

Key takeaways

  1. Assess readiness against a specific use case, its decision and its tolerance for error.
  2. Check five dimensions: fitness, context, lineage, permission and access.
  3. Treat retrieval as a pipeline and evaluate every stage with real questions.
  4. Enforce permissions at retrieval, and propagate deletion to indexes, embeddings and caches.
  5. Publish data products with contracts for schema, quality, access and change.
Next course · T201Application Rationalization: From Evidence to Switch-OffTransform & Optimize · Practitioner
AI-Ready DataUnderstand

Further reading

Continue learning

U101 · UnderstandTechnology Portfolio IntelligenceT201 · Transform & OptimizeApplication Rationalization: From Evidence to Switch-OffAxiarete AcademyAll courses and tracks

Bring this course to your leadership team

We run private working sessions that apply these frameworks to your own estate, data and pending decisions.

Request a Private BriefingExplore the Academy