The first process worth automating usually isn't customer service
Most companies want to start with customer messages because that's the visible pain. The best first candidate is usually something else. Here are the four questions we use.
Nine out of ten first conversations start with customer service. The reasoning is obvious: there are a lot of messages, they’re tiring, and everyone can see them.
But customer service is rarely the best first candidate, because the pain is spread thin. Thirty seconds per message is hard to point at afterwards and call a saving. What pays back is the work that’s concentrated and repeated.
The four questions we use
1. Does someone spend more than three hours a week on the same task?
Not “sometimes”. Every week. Three hours is a practical floor — below it, whatever you build won’t be remembered often enough to stick.
2. Is the method actually settled?
If three colleagues handle the same document three different ways, someone has to decide which way is right first. That’s a management decision, not an AI problem. AI accelerates the existing method, including a bad one.
3. How expensive is a mistake to undo?
Start where recovery is cheap. A report cleaned up wrong gets redone. A quote sent wrong gets explained to a client in person. Build confidence on the first kind before touching the second.
4. Is there someone willing to own it?
A workflow with no owner is unused within two weeks. This matters more than technical difficulty.
The three that usually qualify
In what we’ve built so far, the tasks that clear all four are generally:
- Quote and order document cleanup — supplier quotes and purchase orders arriving in every format, needing to end up in one comparable table.
- Contract term extraction — not replacing a lawyer, but pulling out payment terms, termination clauses, and liability caps so a person can scan them in a minute.
- Monthly report cleanup — a CSV export that needs columns renamed, figures reconciled, and charts produced.
All three share a shape: a file goes in, a file comes out, and the step in between is tedious by hand.
Why we suggest doing exactly one
The failures we’ve seen weren’t badly built. They shipped ten workflows at once, nobody knew which to start with, and none of them got used.
One Action people genuinely open every day beats ten sitting in a menu. Build the first, confirm the team actually uses it, then the second one means something.
