Data readiness before AI: the work nobody budgets for
AI projects rarely fail at the model. They fail because the data that would answer the question is duplicated, incomplete or locked in a system nobody can query.
The four conditions
A single identifier per entity. Consistent status values. A reliable event history. Access to the data in a form a program can read without manual export.
If any of these is missing, the first phase of the project is data work, and it should be scoped and funded as such.
Start with a sample you can label
Before choosing a model, assemble two hundred representative examples and label the correct output by hand. That set becomes your accuracy benchmark and your first honest signal about feasibility.
Plan for drift
Businesses change: new products, new policies, new customer behaviour. Schedule periodic re-evaluation against fresh labelled samples rather than assuming last quarter’s accuracy holds.
Common questions
- How much data do we need?
- For extraction and classification, a few hundred well-labelled examples often beat tens of thousands of messy ones. Quality of labelling matters more than volume.
- Can we skip the benchmark and tune later?
- You can, but then you cannot tell whether the system works or merely looks like it does.
Next step
If this question is live in your business right now, describe the situation and you will get a written view on the options, including whether you should build at all.
Discuss your projectRelated notes
All insightsPractical AI automation, and where it should stop
Where language models genuinely reduce work in a business workflow, where deterministic code is cheaper and safer, and how to tell the difference.
ReadAI and automationWhere AI should not be allowed to decide
The categories of decision that need human accountability, and how to design automation that assists without taking responsibility.
ReadAI and automationAutomating invoice and document workflows
A practical architecture for extracting, validating and posting document data, including the exception queue that makes it reliable.
Read