Data Readiness Checklist
Bad data does not stop an AI project — it lets it start and then kills it. Work through each dimension before committing to any timeline or scope.
Data readiness checklist
Work through each dimension before starting any AI project. A gap in any one area can invalidate your data foundation before you've built anything.
Can the data actually be extracted, moved, and used by the system you are building?
-
Is there a documented path to extract this data from its source system?
Confirm the extraction method works end to end, including all system boundaries and transformations.
Critical -
Have all approval and access permissions been mapped and confirmed?
Approval chains, IT security reviews, and legal sign-offs can each add weeks to a timeline.
Critical -
Is there a stable, automated extraction method — not a manual export?
A manual spreadsheet export or one-time extract is not a reliable production data source.
Important -
Can all required data sources be accessed from the target environment?
Firewalls, on-premise systems, and cloud boundaries all require explicit connectivity planning.
Critical -
Are join keys available and validated across all source systems?
Cross-system joins without a reliable shared key often require significant data engineering work.
Important -
Has integration effort been estimated and included in the project timeline?
Integration typically accounts for 40–60% of delivery effort on AI projects.
Important
Is the data complete, accurate, consistent, and sufficient in volume for AI use?
-
Has a formal data profile been run against all sources in scope?
Null rates, format consistency, outliers, duplicates, and value distributions should all be documented.
Critical -
Are known data quality issues documented with remediation plans — not just flagged?
Parked issues become project risks. Each known problem needs an owner and a fix date.
Important -
Is there sufficient historical volume to train, fine-tune, or evaluate a model?
AI models require more historical examples than most teams assume; too little data means the model cannot generalise.
Critical -
Is the data accurate — does it reflect what it claims to represent?
Data that has drifted from what it was supposed to represent will produce outputs that reflect that drift.
Critical -
Are values consistent across the dataset — same things described the same way?
Inconsistent labelling or classification within the same field creates unpredictable model behaviour.
Important -
Has a bias review been conducted for any use case that affects people?
Historical data reflects historical decisions. If the use case affects people, systematic bias must be assessed.
Important
Is there a named individual accountable for each data source — not a team?
-
Is there a named individual — not a team — accountable for each data source?
Team ownership distributes accountability so broadly that nobody acts on it. Each source needs a named person.
Critical -
Does that person know the AI project will depend on their data?
Owners who do not know AI depends on their data will not flag changes, delays, or quality degradation.
Important -
Is the data documented in an actively maintained data dictionary?
An undocumented data source is a risk. If the owner leaves, institutional knowledge leaves with them.
Important -
Will the data owner remain in role for the duration of the project?
A data owner who changes mid-project creates a gap in accountability at the worst possible time.
Watch -
Is there a handover plan if the data owner changes during delivery?
Without a documented handover plan, a departure creates an unowned data source in production.
Watch
Is using this data in an AI model legally and ethically sound?
-
Has legal and data protection been involved at the start — not as a final sign-off?
Legal involvement at the end is sign-off. Legal involvement at the start is architecture guidance.
Critical -
Does the original consent or collection basis cover AI training and inference use?
Consent for reporting or analytics does not automatically extend to AI training or inference.
Critical -
Have data residency requirements been mapped before platform selection?
Selecting a cloud platform before mapping residency requirements can force costly architectural changes.
Important -
For regulated industries, have applicable frameworks been identified and confirmed?
GDPR, HIPAA, financial services, and public sector frameworks each impose specific constraints on AI use.
Important -
Is PII handling in training data, inference calls, and stored outputs compliant?
Passing personal data through a third-party API or storing AI-inferred data carries specific GDPR obligations.
Critical
Is the data fresh enough today, and is there a plan to keep it that way after go-live?
-
Is the training data recent enough to represent current conditions?
The age of training data relative to current conditions determines whether the model reflects reality.
Important -
Is there an automated refresh process — not a manual one — for ongoing data supply?
Manual refresh processes get deprioritised, forgotten, or orphaned when the responsible person moves on.
Important -
Is drift monitoring planned so degradation is detected before users notice it?
Model degradation is gradual and invisible without monitoring — users notice before teams do.
Important -
Has a retraining trigger and cadence been agreed with the data owner?
A retraining trigger without a cadence leaves the model drifting until someone remembers to update it.
Watch -
Are there named operational owners for refresh, monitoring, and retraining after go-live?
Operational ownership of refresh and monitoring must be defined before go-live, not after.
Important
About the Data Readiness Checklist
Who it is for
Use this checklist when you already have an AI use case in mind and need to know whether your data can support it. It suits data leads, architects and product owners who have to commit to a timeline.
Run it against one specific use case and the data it needs. Answering for the organisation in general gives a comfortable score and tells you little.
What it measures
The 27 questions cover five dimensions of the data that will feed the system.
- Accessibility. Whether the data can actually be reached, or is trapped in legacy systems and approval queues.
- Quality. Whether completeness, accuracy and consistency are known and documented.
- Ownership. Whether a named person is responsible for the data and its ongoing quality.
- Compliance. Whether personal data, sensitivity and regulatory constraints have been identified.
- Freshness. Whether the data can be kept current once the system is in production.
How to read the result
Tick each question you can answer yes to for the use case in front of you. The score is the share of questions ticked, per dimension and in total.
80% or more reads as ready to proceed. 55% to 79% means proceed with caution. 30% to 54% means significant gaps. Below 30% means not ready.
The weakest dimension sets your real position. AI amplifies whatever data quality exists, so one empty dimension is a reason to pause even when the total looks healthy.
Common questions
What does data readiness for AI mean?
It means the data a use case depends on is reachable, good enough, owned by someone, legal to use and able to stay current. A model can be swapped in a week. Data problems take months, which is why they are worth finding first.
How much data do I need?
It depends on the approach. Training a model needs labelled history. A generative AI use case built on retrieval needs far less, but the documents still have to be accessible, accurate and permitted for that use. The AI Approach Finder helps decide which case you are in.
Should I run this before or after the AI Readiness Checklist?
After. The AI Readiness Checklist tells you whether data is your weak point among five dimensions. This one tells you exactly where in the data the problem sits.
Is it free?
Yes. The tool is free and runs in your browser with no account. The downloadable report asks for an email address, and your scores are sent with it so I can follow up if you want to talk them through.