An AI project can have a clear business case and a promising use case, yet struggle when the data behind it is incomplete, inconsistent, difficult to access or poorly governed. Before an organisation starts an AI pilot, it needs to understand whether its existing data can support the intended application.
This is the purpose of an AI data readiness review. It helps business and technology teams identify data gaps, understand access and governance requirements, and decide what needs attention before development begins.
The review does not require every enterprise dataset to be cleaned or migrated. It starts with one proposed use case and examines only the data needed to support it. This guide provides a practical checklist covering data quality, ownership, integration, architecture, privacy, machine learning datasets and GenAI knowledge sources.
Key Takeaways
- Data readiness should be evaluated against a specific AI use case, not the organisation’s total data volume.
- Data quality, governance, integration and access controls are essential considerations before an AI pilot.
- Machine learning and GenAI applications have different data preparation requirements.
- A readiness review should identify evidence-based gaps, owners and priorities.
- Organisations can begin with one use case rather than attempting an enterprise-wide data clean-up.
Why Should You Review Data Readiness Before an AI Project?
Data problems often become visible only after a team starts building or testing an AI application. Missing records can affect model predictions, disconnected systems can prevent access to relevant information, and outdated documents can lead a GenAI application to retrieve incorrect answers.
These issues can increase development effort and make pilot results difficult to trust.
Gartner reported that 63% of organisations either lacked or were unsure whether they had the right data management practices for AI. It also forecast that organisations would abandon 60% of AI projects lacking AI-ready data through 2026. These findings highlight why data preparation deserves attention before a project moves into development.
Common warning signs include:
- Different systems contain conflicting versions of the same business information.
- Important records have missing fields, duplicate entries or inconsistent formats.
- Teams cannot identify who owns or approves access to particular datasets.
- Business documents contain outdated policies or unverified information.
- Data access depends on manual requests and individual employees.
- There is no regular process for monitoring data quality or freshness.
A readiness review helps teams distinguish between issues that can be addressed within a pilot and problems that require wider attention.
What Should an AI Data Readiness Checklist Include?
A useful checklist examines eight areas: accuracy, completeness, consistency, relevance, accessibility, governance, security and timeliness.
| Readiness Area | What to Review | Example Warning Sign |
|---|---|---|
| Accuracy | Correctness of records and values | Conflicting customer or product information |
| Completeness | Availability of required fields and records | Missing historical transactions |
| Consistency | Standard formats and definitions | Different date, currency or naming formats |
| Relevance | Connection between data and the use case | Large volumes of unrelated information |
| Accessibility | Availability to approved systems and teams | Manual access requests delay work |
| Governance | Ownership, lineage and policies | No clear data owner |
| Security | Privacy, permissions and protection | Sensitive records exposed to unauthorised users |
| Timeliness | Refresh frequency and data freshness | Outdated records used for current decisions |
The checklist should be supported by evidence. For example, a team can measure duplicate records, check data refresh intervals, review access permissions and document ownership gaps instead of relying only on subjective maturity scores.
How Do You Check Data Quality Before an AI Pilot?
Data quality should be assessed against the information an application actually needs. A dataset that is suitable for monthly reporting may not be suitable for a predictive model that depends on detailed historical records.
Start by identifying the fields, time periods and source systems required by the proposed use case. Then examine the following:
Accuracy: Check whether values reflect the underlying business events. Compare records against trusted source systems where possible.
Completeness: Identify missing fields, incomplete records and gaps in historical coverage. Determine whether the missing information affects the intended outcome.
Consistency: Review naming conventions, date formats, units, identifiers and business definitions across systems.
Duplicates: Identify repeated records that could distort counts, training examples or retrieval results.
Freshness: Confirm how frequently source data changes and whether the refresh schedule matches the application’s needs.
For a demand forecasting pilot, for example, the team may need historical sales, inventory, pricing and promotion data. Missing promotion records or inconsistent product identifiers could affect the quality of the resulting predictions.
Document each issue, its likely impact and the person responsible for resolving it. This gives the project team a practical list of data requirements before development starts.
What Data Preparation Is Needed for Machine Learning?
Machine learning applications use data to identify patterns and produce predictions or classifications. Their preparation requirements depend on the prediction target and the learning method.
A readiness review should establish whether the organisation has enough relevant historical information, whether the records can be labelled where necessary, and whether the dataset can be reproduced for testing.
Use this checklist:
- Define the prediction target: Identify the outcome, such as customer churn, demand, fraud or equipment failure.
- Identify relevant sources: Locate the CRM, ERP, transaction, product, support or sensor data that may contain useful signals.
- Clean the records: Address missing values, duplicates, inconsistent formats and outliers.
- Prepare features: Identify the variables that may help explain or predict the target.
- Label data where required: Establish reliable labels for supervised-learning tasks.
- Create separate datasets: Maintain training, validation and test sets to evaluate performance properly.
- Review representation: Check for bias, class imbalance and gaps across customer groups, regions or time periods.
- Version the data: Record the dataset and transformations used for each experiment.
A particularly important consideration is reproducibility. If a team cannot identify which data was used to train and evaluate a model, it becomes harder to investigate performance changes or repeat an experiment reliably.
What Should You Review for GenAI and RAG Data Readiness?
GenAI applications often rely on enterprise knowledge rather than only structured training datasets. A retrieval-augmented generation (RAG) application, for example, retrieves relevant information from approved sources before using a language model to generate a response.
The quality of that information directly affects the usefulness of the output.
Before a GenAI pilot, review these areas:
- Source quality: Identify the policies, manuals, product documents, contracts, internal knowledge bases and other sources the application will use.
- Document relevance: Remove duplicate, obsolete or unrelated material.
- Document structure: Preserve headings, tables and relationships when preparing content for retrieval.
- Metadata: Record details such as document owner, source, version, date, region and language.
- Permissions: Ensure retrieval respects the access rights of the person using the application.
- Content freshness: Establish how updated documents will replace older versions.
- Answer evaluation: Test whether retrieved information supports accurate, relevant and properly attributed responses.
A vector database may be useful when an application needs semantic search across large volumes of unstructured content. It is not automatically required for every GenAI use case. The choice should depend on content volume, query patterns, latency, access requirements and how frequently information changes.
For example, an internal policy assistant needs access to current, approved policy documents. If outdated versions remain in its knowledge sources, the application may retrieve conflicting guidance even when the underlying language model performs as expected.
How Do Data Integration and Architecture Affect Readiness?
Enterprise information is often distributed across systems such as CRM, ERP, finance platforms, support applications, warehouses and operational databases. An AI use case may need information from several of these sources.
A readiness review should establish whether the required information can be connected reliably and made available to approved applications.
Consider these questions:
- Which source systems contain the required information?
- Are common identifiers available to connect related records?
- Can data be accessed through APIs, pipelines or approved interfaces?
- Are data transformations documented?
- Can the existing infrastructure support the expected processing and storage requirements?
- Is there a way to monitor pipeline failures and data freshness?
The architecture should reflect the use case rather than introduce unnecessary technology. A warehouse, lakehouse, cloud data platform or API-based integration may be appropriate depending on the information involved and how the application will consume it.
The goal of this review is to identify integration dependencies early, before they become delays during a pilot.
How Should Privacy, Security and Governance Be Checked?
Data readiness is not only a technical question. Organisations must also understand who owns the information, who can access it and what restrictions apply to its use.
This becomes especially important when an application processes personal, financial, healthcare or other sensitive information.
Review the following controls:
- Ownership: Assign clear data owners and stewards.
- Classification: Identify sensitive, confidential and publicly available information.
- Access management: Apply role-based or least-privilege access.
- Data protection: Use masking, encryption or other safeguards where appropriate.
- Lineage: Record where data originates and how it is transformed.
- Auditability: Maintain relevant records of data access and usage.
- Retention: Define how long information should be stored and when it should be deleted.
- Human oversight: Establish review processes for high-impact decisions.
Organisations operating in India should consider applicable requirements under the Digital Personal Data Protection Act and relevant sector-specific rules. Companies handling information across multiple markets may also need to account for regulations such as GDPR.
For GenAI applications, permissions should be enforced during retrieval as well as at the original data source. Uploading a document to a knowledge base should not automatically make it accessible to every user.
How Can You Assess Data Readiness and Prioritise Gaps?
Once the review is complete, organise the findings into a simple readiness assessment. The purpose is not to produce a score for its own sake. It is to establish which gaps could affect the proposed AI pilot and what needs to happen before work progresses.
| Area | Early-Stage Condition | More Mature Condition |
|---|---|---|
| Data Quality | Manual corrections and inconsistent records | Defined quality checks and regular monitoring |
| Governance | Unclear ownership and policies | Named owners, documented policies and lineage |
| Integration | Disconnected source systems | Reliable connections to required sources |
| Architecture | Limited capacity or undocumented dependencies | Suitable storage, processing and access interfaces |
| Privacy and Security | Broad permissions and unclear controls | Defined access rules, protection and audit trails |
| ML Data | Unlabelled or unversioned datasets | Documented, reproducible datasets |
| GenAI Knowledge | Outdated or unstructured documents | Curated, versioned and permission-aware sources |
After identifying the gaps, group them into three priorities:
- Resolve before the pilot: Issues that could make the application unsafe, unreliable or impossible to test.
- Address during the pilot: Improvements that can be completed as the use case is validated.
- Plan for later: Wider data improvements that are useful but not necessary for the initial use case.
This approach prevents organisations from treating every data issue as an immediate blocker.
What Does a Practical 90-Day Data Readiness Review Look Like?
A focused 90-day review can help an organisation move from an initial use-case discussion to a documented data plan. The timeline is indicative; the actual effort depends on data complexity, system access and governance requirements.
Days 1–15: Define the use case
- Select one business problem with a measurable outcome.
- Identify the intended users and application requirements.
- List the data sources needed to support the use case.
Days 16–30: Map and inspect the data
- Identify source systems and data owners.
- Review quality, completeness and consistency.
- Document access restrictions and integration dependencies.
Days 31–45: Assess governance and architecture
- Review permissions, privacy and retention requirements.
- Identify the storage, processing and access interfaces needed.
- Document the main technical and governance gaps.
Days 46–60: Prioritise the requirements
- Separate critical issues from longer-term improvements.
- Assign owners and target dates to priority actions.
- Define the minimum data foundation required for the pilot.
Days 61–90: Validate the plan
- Prepare a representative dataset or approved knowledge collection.
- Test whether the required information can be accessed reliably.
- Document unresolved risks and dependencies.
- Agree on the next steps and success criteria for the pilot.
The outcome should be a practical readiness report, not simply a statement that the organisation is ready or not ready.
What Mistakes Should Organisations Avoid During Data Readiness Reviews?
Several common mistakes can make an otherwise useful review less effective:
- Preparing every dataset first: Start with the information required for one use case rather than attempting an organisation-wide clean-up.
- Focusing only on data volume: More records do not necessarily mean better information.
- Ignoring ownership: Data without clear accountability is difficult to maintain and govern.
- Treating GenAI as basic document search: Retrieval quality also depends on metadata, permissions, document structure and freshness.
- Overlooking privacy: Sensitive information needs appropriate controls before it is exposed to an AI application.
- Ignoring ongoing monitoring: Data quality and freshness can change after a pilot begins.
- Measuring technical results alone: Evaluate whether the data and application support the intended business outcome.
A readiness review should be revisited when the use case, source systems, data access requirements or application scope changes.
Conclusion: Make Data Readiness the First Checkpoint
A successful AI initiative needs more than a promising use case. It needs reliable information, clear ownership, appropriate access and a data foundation that supports the intended application.
An AI data readiness review helps organisations identify those requirements before development begins. By starting with one use case, documenting the gaps and prioritising the most important improvements, teams can avoid unnecessary preparation and make better-informed decisions about their AI roadmap.
If your organisation is evaluating an AI initiative, the next step is to understand how its data, architecture and business requirements fit together. GoodWorkLabs can help translate that assessment into a practical path from an initial use case to an AI pilot and subsequent implementation.