Choosing a first automation: compare three tasks
Compare email sorting, document capture and data transfer by their inputs, exceptions, checks and ownership.
Choose a task with a recognisable result
A first automation trial needs a bounded task: what input arrives, what result is expected, and who can judge whether it is correct? Frequent repetition helps, but is insufficient as a selection criterion. A quickly generated result can create substantial rework when data is missing or responsibility is unclear. Record the current checking effort and typical exceptions as well. Required access rights are part of the task description.
Compare three candidates on the same basis
A fictional service office considers the tasks below with the data and responsibilities described here.
Candidate | Input and expected result | Open exception or check | Next step |
|---|---|---|---|
Presort emails | Varied free text; suggested responsible team | Mixed requests and unclear responsibility need review | First collect examples and compare simple sender rules with category suggestions |
Capture delivery notes | Varied scans; suggested line items and quantities | No reviewer has been assigned for missing purchase-order references | Define ownership and the clarification path before the trial |
Transfer approved CSV data | Fixed columns and a unique external reference; preview of target values | Missing required values and resubmitted rows must be visible | Start with a preview without write access |
Under these assumptions, the CSV preview is a useful first trial. Mapping is clear, and the responsible person can compare source and target values directly. Fixed columns may need only ordinary data mapping with validation. Whether a model is also needed depends on a specific unresolved task.
Describe the trial through to its error path
Define which fields the preview may transfer and how rejected rows remain visible. Check a suitable row, a missing reference, an unauthorised target customer and the same row submitted again. None of these cases may create a production record during the preview trial. A later write step needs its own permissions, a confirmed target state and a rule against duplicate transfers. A displayed preview is not evidence of those capabilities.
Assess continuation against the work involved
Compare the existing process with the trial, including checking and rework. Record incorrect mappings and cases that need someone to investigate. Name a person who will actually handle them. If the preview does not represent the required data reliably or exceptions remain unresolved, improve the inputs and rules first. Consider extending the process once the expected result and the handling of deviations are clear.
Sources
Read more
All articles
AI & Automation with ASES
AI assistants and automated workflows