Who this is for: Data owners, operations leaders and technology teams
A business can have a large amount of information and still be unable to answer a specific operating question reliably. Documents may disagree, records may use different definitions and important context may exist only in a colleague's experience.
Data readiness is therefore a review of fitness for a proposed task. This guide offers questions for that review. It does not require every organization to build a new data platform before experimenting with AI.
Identify the information the task needs
Describe the desired output and work backwards to the sources a person would use to produce it. Identify which records are authoritative, which are supporting context and which should be excluded. An internal knowledge assistant and a forecasting model may need very different preparation.
Review representative inputs with the people who understand the work. Ask what an apparently complete record leaves unsaid and where definitions change across teams. Document those differences before treating the sources as interchangeable.
Check quality and context
Inspect missing values, inconsistent formats, duplicate records and outdated material where relevant to the task. For documents, also consider version history, unclear ownership and information embedded in images or attachments.
Google's machine-learning education discusses preparing numerical data through appropriate cleaning and transformation. That is a useful reference for numerical ML work; it should not be treated as a complete preparation plan for every AI application.
- The data represents the cases the system will encounter.
- Definitions and units are understood.
- Relevant sources are current or clearly dated.
- Missing information and known limitations are recorded.
Agree ownership and access
Name the people responsible for the information and for approving its use in the proposed workflow. Decide which users should be able to retrieve which material. Combining sources should preserve the access boundaries that matter to the organization.
Review how information would move through the intended application and its providers. Resolve access and retention decisions with the responsible teams before processing sensitive material. Record the agreed boundaries so they can be tested.
Prepare evaluation examples and maintenance
Collect realistic examples of questions or inputs and the responses the team would accept. Include ambiguous cases, absent information and examples that should be declined or referred to a person. Keep the test material appropriately separated from model training where that distinction applies.
Choose an owner for changes to the source information. Decide how the system will incorporate an updated policy or corrected record and how the team will notice stale material. Readiness is a continuing responsibility after the first deployment.
A data readiness checklist for the first discussion
Use the checklist to identify gaps and owners. Resolve the gaps that matter for the bounded task before expanding the scope.
- The task and required information are defined.
- Source ownership and authority are understood.
- A representative quality review is complete.
- Use, access and retention decisions are agreed.
- Realistic evaluation examples include difficult cases.
- Updates, corrections and ongoing maintenance have an owner.
Frequently asked questions
Do we need a large dataset for every AI project?
No. Requirements depend on the task and implementation. Access to a smaller set of relevant, reliable information may be more useful than a large collection with unclear quality or ownership.
Is document collection the same as data readiness?
No. You also need to understand authority, versions, access, relevant context and how the proposed system will be evaluated and maintained.
Can we begin with one department?
A bounded workflow can help clarify the requirements. Check shared dependencies and access boundaries so that the department's pilot does not assume ownership of another team's information.
Sources and further reading
Technical references supporting the topics discussed above. The decision frameworks and recommendations are Codersbay editorial guidance.
- Working with numerical data (opens in a new tab)Google Machine Learning Crash Course
