Classify data first
Identify the data, people, recipients and training use before selecting a model.
Local is an option, not magic
Local models can reduce transfer and increase control, but still require access control, updates, logs, tests and deletion rules.
Design transparency
Where applicable, people should recognise AI interaction or generated content. Build labelling, human review and escalation into the product.
What to prepare for the first conversation
List data types, roles, storage locations, recipients and permitted actions. Will the model only suggest text or actually send emails and modify records? Define approval boundaries. Start with synthetic data, explicit quality criteria and a manual fallback. Evaluate the pilot before choosing local, controlled cloud or hybrid deployment.
Discuss a project ↗