Documents, data and risk

How extraction with a schema, validation against what you expected, and a human exceptions queue turn the daily paper flood into records your systems can use, without a single silent error.

What has to be true before an AI can answer "which customers are slipping", how text-to-SQL works, why you cannot trust a chart you did not check, and which reports are worth automating.

The research-first method, the human edit that is not optional, the things you should never let a model write, and what search engines actually do with machine-written text in 2026.

Training clauses, zero-retention options, masking before the model, on-premise models, the contract terms that matter, and the questions to put to every vendor before a single record leaves your building.

A fair test you can apply to any consultant, including us: the questions that separate builders from sellers, the ownership terms that matter, and how to run a first project small enough to survive being wrong.

The one AI attack every owner should understand before letting AI read email, documents or web pages, and the four defenses that make it safe: separating instructions from data, allowed actions, approval gates and output checks.