Automating a mess produces a faster mess
The most common failure in AI projects is not technical. It is automating a process nobody had mapped, which converts an inconsistent manual workflow into an inconsistent automated one that now runs at volume.
Map the process first, on paper, with the person who currently does it. Half the time that exercise finds steps that should simply be deleted, which is a better outcome than automating them.
A three-part test
Ask three questions of any candidate task. Is it repetitive enough that a person does it many times a week? Is it rules-based enough that two people would do it the same way? And what is the cost of getting it wrong, confidently, without anyone noticing?
Tasks that are repetitive, rules-based and low-consequence are ideal: data entry, routing, enrichment, summarizing a call into a CRM record, drafting a first-pass response for review.
Tasks that are high-consequence need a human in the loop regardless of how repetitive they are. Anything that makes a commitment to a customer, quotes a price, or states a fact about your product belongs in that category.
Guardrails are the deliverable
A workflow that works ninety-five percent of the time is not ninety-five percent as good as one that works always — it depends entirely on what the other five percent does.
Design the failure mode explicitly. What happens when the model is uncertain? Where does the item go for review? Who is alerted, and how quickly? Is there an audit trail showing what was sent and on what basis?
In regulated contexts this is not optional. For clinics in particular, nothing generated by a system publishes or sends without a named human approving it.
Measure what you claimed you would save
Baseline the manual process before you replace it: how long it takes, how often it is wrong, how much rework it generates. Without that, any claim about time saved is an impression.
Then track error and rework rate alongside throughput. Automation that doubles output while tripling corrections has not helped anyone, and it is surprisingly easy to ship without noticing.