In brief
How do you know whether data is ready for AI?
Check whether the data exists for the specific use case, whether you are permitted and able to access it, whether its quality and labels are good enough, whether it represents the conditions the model will face, whether it will be available when predictions are needed, and who owns it. Answer these before building, using real samples rather than descriptions.
AI proposals often describe data in reassuring terms: we have years of transaction history, all our documents are in the system, the sensors log everything. When a project starts, those descriptions meet reality. Fields are missing, labels are inconsistent, access needs approvals nobody anticipated and the data describes a process that has since changed.
None of this is unusual. The problem is discovering it after budgets and expectations are set. A short readiness check at the start avoids most of these surprises.
The checklist
Availability and access
- Does the data needed for this specific use case exist, and where does it live?
- Are we legally and contractually permitted to use it for this purpose, including under data protection law and customer agreements?
- Can the project team practically access it, and how long will approvals take?
Quality and labels
- What proportion of key fields are missing, invalid or inconsistent?
- For supervised learning, do reliable labels or outcomes exist, and how were they recorded?
- Do definitions of important fields match across systems and teams?
Representativeness
- Does the data cover the situations, customer groups, products or sites where the model will be used?
- Has the underlying process, product range or policy changed in ways that make older data misleading?
- Are rare but important cases present in sufficient numbers?
Operational fit and ownership
- Will the same data be available, in the same form, at the moment a prediction or answer is needed?
- Who owns each data set and can approve changes to its definition or quality?
- What security, residency and retention rules apply?
Assess with samples, not descriptions
The most reliable way to answer the checklist is to obtain real samples early and profile them. Simple analysis of completeness, value distributions, duplicates and date ranges reveals most issues within days. For unstructured data such as documents or images, reviewing a random sample with a subject expert is equally revealing.
When the answers are uncomfortable
| Finding | Options |
|---|---|
| Data does not exist | Start collecting it now, narrow the use case, or use a different approach such as rules or expert systems |
| Access not permitted | Seek consent or a lawful basis, anonymise, or redesign to avoid the restricted data |
| Poor quality | Fix at source where possible, add validation, or scope the model to reliable segments |
| Missing labels | Label a focused sample with experts, use active learning, or start with unsupervised or LLM-based methods |
| Not representative | Collect targeted data, restrict deployment scope, or monitor closely with human review |
| Not available at prediction time | Redesign features around information genuinely available at that moment |
Sometimes the right answer is to invest in data foundations first and return to the AI use case later. That is not a setback. It is often the fastest route to AI that actually works.
Sources and further reading
About the author: Dr. AarzooCo-Founder and Chief Operating Officer, AARHIT Systems. Writes on delivery, quality assurance, data readiness, monitoring and human review. More articles by Dr. Aarzoo