Most writing about AI in business describes demonstrations. This is about what we actually ship: retrieval over a company's own documents and records, forecasting against its own history, and anomaly detection on its own operational data. The difficult parts are data quality, evaluation, and knowing when to keep a person in the loop.
Agents earn their place on work that is high-volume, text-shaped and cheap to get wrong per item: drafting a reply, extracting fields from a document, classifying and routing a request, summarising a long record before someone acts on it. They are a poor fit where an error is expensive and hard to reverse. We scope the first group aggressively and keep a human on the second.
Most failed AI projects are data projects in costume. Before any model work we establish which records exist, where they disagree with each other, what has been de-identified, and who is allowed to read it. A model trained over contradictory data produces confident nonsense at scale.
A model-backed feature ships with an evaluation set drawn from real cases, a confidence threshold below which it defers to a person, and a log of what it decided. If nobody can explain why the system produced an answer, it is not ready to sit inside a business process.
Where personal or clinical data is involved, de-identification happens before analysis, access is role-based, and the pipeline is auditable. This is designed to satisfy the controls your assessor will ask to see — certification itself belongs to your organisation and its auditor, not to us.
Tell us what you run today and we will tell you what we would change, and whether it is worth changing.