AI & automation
Before you automate a task with AI, watch someone do it
The useful part of AI automation starts with the work itself: the exceptions, the judgement calls and the cost of getting an answer wrong.
By CODXON · 6 min read · Published
Sit with the work before choosing the tool
A process diagram rarely tells the whole story. Watch somebody handle five or ten real cases and you will see the missing information, judgement calls and quiet workarounds that never made it into the documentation.
That is where a useful automation project begins. If the task repeats, the inputs are available and a person can explain what a good result looks like, there may be something worth improving. The choice of model comes later.
- Where does the information arrive?
- Which cases are genuinely routine?
- What makes somebody stop and ask a colleague?
- Where does the finished work need to go?
Be honest about the cost of a wrong answer
An imperfect first draft of a marketing description is inconvenient. An incorrect compliance decision can be serious. Those two jobs should not have the same controls.
For higher-risk work, the source should be visible, uncertain cases should stop, and a person should have a clear way to review or override the result. If nobody can explain what happens when the system is unsure, it is not ready for daily use.
Test one narrow workflow
A small pilot is usually more revealing than an impressive demonstration. Put it in front of the people who do the work and measure the entire task, including preparation, checking and corrections.
Time saved matters, but so do rework, response quality and whether people choose to use the tool after the novelty has worn off. If the pilot does not improve the working day, a bigger rollout will not rescue it.