Is your data good enough for AI?

Check data against a particular task: quality, origin, labels, access, responsibility and permitted use. Overall volume or a good result in a short test does not establish readiness for every AI project.

Which six areas should you check?

Acora’s article describes six data characteristics. The following questions help check them for your task:

Quality and currency: which errors, missing values and stale records affect the task? Record how they are checked and updated.
Origin: where did information come from, who changes it and what transformations has it undergone? Keep a traceable connection to the source.
Labels and context: are fields, units, labels and metadata understandable? The same name does not always mean the same thing.
Access and protection: who can read, change and share? Check permissions and appropriate measures rather than treat a password as complete protection.
Governance: who owns accuracy, changes and problem correction? Shared rules can also operate across distributed systems.
Permitted use: on what basis do you process personal data for this purpose? The European Data Protection Board explains that consent is one possible GDPR legal basis, not a universal requirement for every processing activity. The legal assessment depends on the case.

AI adoption figures do not measure your dataset’s quality. In CDO Insights 2026, Informatica reports a survey of 600 data leaders: 69% generative AI adoption against 48% in the previous comparison, and 57% naming reliability as a key barrier to moving AI from pilot to production. These are different survey questions, not an assessment of every company.

Access control is also separate from clean records. Wiz describes an incident publicly disclosed in 2023 involving 38 TB of additional private data shared by a Microsoft AI team through an overly permissive token. This illustrates sharing permissions, not the measured risk of your project.

Is volume more important than suitability?

Define the task first. Using a ready-made model on documents, analysing sales and training your own model do not impose identical requirements for volume, coverage and quality. A small organised sample can help expose gaps; it does not itself establish reliability across all real cases.

Choose a bounded use and examine examples, including difficult cases, rather than wait for undefined “enough data”. Record the desired outcome, how errors will be found and who judges acceptable quality. Do not assume clean data always compensates for missing coverage.

Why organise before choosing a tool?

Clarify where data originates, where it goes and who must act. SixFive recommends mapping the process before choosing the tool. This is practical advice, not evidence that every business must buy one central system.

A predictable place to find information can be a well-managed spreadsheet, several connected systems or another existing environment. Ownership, clear fields and a checkable flow matter. AI does not remove the need for a person to know what an acceptable result looks like.

What should you clarify before buying a tool?

Start with three questions:

Can you find the right information for the task and distinguish it from outdated copies?
Is there an owner for accuracy and a procedure for correcting errors?
Are the source, purpose and basis or rights for the particular use recorded?

A negative answer identifies work to do. Positive answers begin the review; they do not certify technical or legal readiness. Check particular examples next and establish the task’s limits.

Questions

Must we buy a data management system first?

Not necessarily. Check what can be organised in the existing environment. Buying does not replace ownership, clear rules or checking sources. The decision depends on the particular flow and needs.

Where should a small business start?

Check one important dataset: source, owner, changes, access and some typical and difficult examples. Record gaps for this task before extending the review.

How long does preparation take?

Atlan gives estimates for organisational AI readiness: 3–6 months with established practices and 12–18 months where fundamental governance and quality improvements are needed. This is a vendor estimate for broader readiness, not a deadline for your dataset, a particular project or the AI Opportunity Map.

Sources

Separate available data from gaps for the particular task.

The AI Opportunity Map orders tasks by the hours they cost you today; for the first three it gives a proposed tool, requirements, an owner and a build-effort range in hours, without a build price. It connects uses to available data, established limits and questions for further checking. The Map does not certify model quality, security or legal compliance.

Ask about the AI Opportunity Map

or write to contact@ergostrat.com