A process is genuinely AI-automatable if it is already precisely described, has enough volume to justify the cost, tolerates a detectable and repairable error, and accepts human validation at first. Otherwise, a simple rule does the job better.
The AI-project trap
“We should use AI for this” is not a use case. It is a technology looking for a problem. The projects that fail share the same profile: a vague process, a never-quantified ROI, and forgotten human validation.
The five-question filter
Before considering AI, a candidate process must pass this filter:
- Is the process already precisely described? If three people describe three different processes, automation will codify chaos.
- Does the volume justify the cost? A two-hour-per-month manual task does not deserve a pipeline. Two hours per day, yes.
- Is an error detectable and repairable? A process where an error slips through to the client is not automatable — it becomes automatable after adding a check.
- Can it be humanly validated at first? The best systems start as “proposal with review” and gain autonomy as measurement proves it.
- Does AI bring something a simple rule does not? Many “AI cases” are calculation, format validation or plain orchestration. A rule costs less, fails less, and can be explained.
Candidates that pass the filter
Those where data is textual or unstructured, where human judgement is the bottleneck, and where the tolerable error rate is documented: triaging incoming requests, pre-filling documents, structured extraction, first-pass quality control.
Candidates to reject
Processes where an error is irreversible, where volume is low, where rules change monthly, or where nobody can describe the reference process.
The first deliverable is not a prototype
It is a measurement: current time, current error rate, volume. Without those three numbers, no automation project can be judged. With them, the decision becomes arithmetic.