Start with one workflow

Pick a workflow that already happens. It needs a person who owns the outcome, and you need to be able to describe it without a room full of diagrams. The shape is familiar even when the business is not: a request that becomes a quote, a document checked against a rule, a weekly report assembled from systems you already run. Those are shapes, not clients.

The first question is not which model to buy. It is which decision the workflow may make, and which decision must stop for a person. Money, access, and a record about a person are the usual stops. The engagement that starts there is an audit, then a pilot, then operations. That path is set out on the small business automation page.

What you keep

The data stays yours, in Australia, under your control. Retrieval over sources you approve is the usual way a model sees that data. The record should show which source was used. A model is fine tuned only when a general model misses your cases. Training a model from scratch is a different project, and it is not the starting point.

Hosted, private cloud, or on premises is a choice about where that data sits. It is not a reason to redesign the decision. The same boundary should hold in each place: an owner, a limit on actions, and a path that does not depend on the agent.

When to read the next part

Read the next part when the first workflow is actually running, not when a slide says it could. Part 2 is From a working workflow to a platform.