Data readiness is rarely a question of volume. Teams with modest datasets run successful pilots regularly, while teams with very large datasets stall for months. What separates them is clarity: whether the fields mean what people think they mean, and whether the results can reach the workflow that needs them.
Access
- Can the project team obtain the data within weeks, not quarters?
- Is there a named data owner who has agreed to the use?
- Is the approval route documented, including any ethics or privacy review?
- Can data be extracted repeatedly, or only once by hand?
Meaning
- Is there a written definition for every field the pilot depends on?
- Do two people in the organisation define the key fields the same way?
- Do coded values have a maintained reference list?
- Has the meaning of any key field changed during the period covered?
Quality
- What proportion of records are missing each critical field?
- Are there duplicates, and is there a reliable way to identify them?
- Are timestamps consistent, and in which time zone?
- Are the rare cases the pilot cares about actually present in adequate numbers?
Representativeness
- Which groups, sites, or periods are under-covered?
- Was the data generated under conditions similar to those the tool will operate in?
- Would a model trained on this data systematically underperform for anyone?
Path to use
- Where would an output need to appear for someone to act on it?
- Can that system receive an output, or would results be copied manually?
- Who would be accountable for acting on an output, and do they know?
If the answers are not yet good
A weak readiness assessment is a useful result, not a blocked project. The usual response is to run a short, funded data-preparation phase with its own deliverables — definitions documented, an extract pipeline built, a quality baseline measured — and to start the AI pilot at the end of it. That sequencing is far cheaper than discovering the same problems halfway through a build.
Check access, meaning, quality, representativeness, and the path to use. Gaps in any of the five are worth fixing as an explicit phase before the pilot rather than during it.