Everyone wants to start with the impressive automation. The one that pays back fastest is almost always boring.
The ranking
For each candidate:
Value = (times per month x minutes each) x error weight ÷ build effort
Where error weight is 1 if a mistake is annoying, 3 if it costs a deal, and 10 if it reaches a customer in a way you would have to apologise for.
Fill this in for ten candidates and the order usually stops being controversial.
The five that usually win
1. Lead routing. Happens on every lead, takes seconds of human attention each time, and delays contact by hours when it waits for a person. Cheapest thing on the list to build and it moves speed to lead immediately.
2. Enrichment before contact. Removes the research step that delays every first call and improves every conversation that follows. High frequency, easy to verify, and it makes the next three automations better.
3. The after contact write back. Outcome, summary and next step onto the record, every time. Stops customers repeating themselves, which is the failure they actually remember.
4. Follow up scheduling. The second and third touches that currently depend on memory. Low effort, and it recovers deals that were already paid for.
5. The recurring report. Somebody spends part of Monday assembling it. Automating it returns those hours and, more usefully, makes the number current rather than three days old.
What to leave for later
Anything irreversible without a checkpoint. Automate up to it, put a human approval there, automate after it.
Anything where the process is still in debate. You would be encoding one side of an argument.
Anything touching data you do not trust. Fix the data or route the bad records into a flagged path first.
The impressive one. Usually the agent that does everything. It generally depends on two or three of the boring five working first, and building it early means debugging several things at once.
Start narrow on purpose
Take the top candidate, build the smallest version that produces the result, run it for a week, then extend. The guidance from people who build these systems is the same: simplest thing that works first, autonomy added only where it earns its cost.
Before you build any of them
Write down how you will know it worked. Not the execution log, the business result: open the record, and it is either there or it is not. Teams that decide this before building find their failures in days. Teams that decide it after find them in months.