Look for bounded work, not general intelligence
The strongest early automation candidates have a clear input, a recognisable output and an existing review step. Examples include classifying enquiries, extracting defined fields from documents, drafting a response from approved knowledge, or summarising a case for a specialist.
Design the control system around the model
An AI feature needs more than a prompt. It needs source controls, permission boundaries, output validation, escalation rules and a record of what happened. These surrounding controls determine whether the workflow is dependable enough for production use.
- Restrict the knowledge and actions available to the model.
- Show source material where verification matters.
- Define confidence or exception rules.
- Log decisions without exposing unnecessary sensitive data.
Pilot against a real baseline
Measure the current time, cost and error pattern before introducing automation. During a pilot, compare results with the existing process and record the kinds of work that still require human judgement. Scale only when the combined human-and-system workflow is demonstrably better.
Write down the current workflow, the people involved, the most costly exception and one measure that would show improvement. That creates a stronger starting point than a feature list.