01
Rules-based automation or AI assistance
A predictable task with clear inputs may suit ordinary automation. A task involving summarising, classifying or drafting information may benefit from AI assistance, with appropriate review. Choosing the simpler suitable approach can make the process easier to explain, support and test. Rules-based automation follows an explicit condition and action, such as routing an approved form to a named team. Generative AI produces a response from a prompt and context, so its output requires a different test approach. A workflow can combine both: AI proposes a classification while deterministic rules govern what may happen next.
02
Data, access and integrations
Identify the systems a workflow reads or changes and the permissions needed for each action. Existing sharing can influence what authorised AI tools can discover. Credentials, connectors, retention and external processing should be reviewed for the chosen products and information involved. An integration can read from one system, write to another or trigger an action. Those permissions should be listed separately. A drafting tool does not automatically need permission to send a message, and a reporting tool does not automatically need permission to edit source records. This separation makes a pilot easier to control.
03
Quality checks and human decisions
Define examples of acceptable output and the situations that require a person to intervene. Test missing information, ambiguous requests and failed connections as well as normal cases. Approval steps are especially useful before a process changes important records, communicates externally or commits an action. Use a representative test set and record errors, not just successful demonstrations. For example, measure whether the proposed summary omits key facts or whether a routed request reaches the correct owner. Time saved matters only when the output also meets the task’s quality requirements.