What Data Do You Need Before Starting an AI Project?

Most AI projects that stall do so because of data, not models. The good news is that you rarely need a perfect data estate to start — you need the right data for the approach you choose, and an honest view of the gaps before the budget is committed.

What each approach needs

• Document Q&A and assistants (RAG) — the documents, policies, or knowledge-base articles that hold the answers; a way to keep them current; and the permissions that decide who may see what.
• Predictions and scoring (machine learning) — historical records that include the outcome you want to predict, such as which invoices were paid late or which customers churned, with enough examples of each outcome, including the rare ones.
• Fine-tuning — a curated set of example inputs paired with the ideal output, reviewed for consistency. Quality matters far more than quantity.
• Agents — API or integration access to the systems the agent reads and changes, test environments to try actions safely, and clear rules for what needs approval.
• Document extraction — representative samples of the real documents, including poor scans and unusual layouts, with the correct values marked.

Data readiness checklist

• Ownership and legal basis — you own the data or have the right to use it this way, and personal data has a lawful basis under GDPR or your local equivalent.
• Access — the data can be exported or queried by API, not only viewed in a screen.
• Quality — duplicates, gaps, and conflicting versions are known and manageable.
• Coverage — the data includes the edge cases and rare outcomes the system must handle, not just the common ones.
• Labels — someone with domain knowledge has defined what "correct" looks like, and different people label the same case the same way.
• Freshness — there is a reliable way to pick up new and changed data after launch.
• Sensitive data — personal, health, or financial fields can be redacted or kept within approved environments.
• An evaluation set — a few hundred real examples with known-good answers to test against before and after launch.

Common gaps and how to close them

• No labels — label a representative sample first, with clear written guidelines, rather than the whole dataset.
• Data spread across systems — build the integration or export before the AI work, and treat it as its own milestone.
• Poor quality — run a short cleanup and profiling phase; it is cheaper than training on bad data.
• No way to measure success — work with your domain experts to build the evaluation set before choosing a model.
• Not enough examples of a rare outcome — start with a rules-based or human-in-the-loop version and collect examples as it runs.

How we assess data readiness

Before quoting an AI build, we review sample data with your team against the checklist above and tell you which approach the data supports today, what it would take to support more, and what that preparation costs. Our free AI Readiness Assessment gives a quick first view of where an organisation stands across data, process, and governance.