Where to look first
The best candidate is almost never in the area that asks loudest for artificial intelligence, but in the one that repeats work the most. Start with processes that run many times a month, follow a written procedure and today depend on someone copying data from one place to another. Volume, repetition and clear rules: if any of the three is missing, the project arrives too early.
The signs that a process is a candidate
There are four signs you can see without instrumenting anything. The first is that the process already has defined steps: if the team can explain it on a whiteboard, it can be automated. The second is that data already exists in digital form, even if it is messy. The third is that the result can be verified, because without a way to check the outcome nobody will trust the system. And the fourth, the most important one, that someone on the team is willing to own the process while it is redesigned.
The ones to leave for later
Processes that change shape every week, those that depend on negotiation between people and those that happen very rarely tend to be expensive. It also pays to wait when the process is badly designed to begin with: automating a mess produces mess faster. In those cases the previous work is to tidy the process, which sometimes cuts the time in half without writing a single line of code.
How to prioritise without endless spreadsheets
We place each candidate on two axes: how much time it frees every month and how much it costs to leave it running on its own. The ones that free a lot of time and are cheap go first; the expensive ones with low impact move to the waiting list. That list is not a plan, it is a conversation: the final order is decided by the business, not by the technology.
