What can AI automation actually replace?
Judgement that repeats. If a person makes the same decision hundreds of times a week from information a system can already see, that is automatable now. If the decision is rare, high-stakes, or depends on context nobody has written down, it is not — and pretending otherwise is how automation projects fail.
The test: does the decision repeat?
Count how many times a week somebody makes the same call, and ask what they look at to make it. If the answer is "four hundred times" and "these three screens", you are looking at the best automation candidate in the building. If it is "twice a quarter" and "experience", leave it alone.
Frequency is what pays for the build, and legibility is what makes the build possible. A decision that repeats but depends on something nobody has ever written down is not ready yet — the first piece of work is writing it down.
What automates well today
The work that sits between systems rather than inside anyone's head:
- Catalog and content operations
- Descriptions, variants, tagging, translations, keeping the same product accurate in five places at once.
- Intake, routing and triage
- Reading what arrived, deciding what it is, putting it where it goes, and telling the right person it is there.
- Reporting and reconciliation
- Pulling the same numbers from the same places on the same day, and flagging only what moved.
- Production pipelines
- The mechanical steps between a file existing and a file being usable — renaming, versioning, transcoding, checking.
What does not automate well
Anything where being wrong is expensive and rare: pricing exceptions, hiring, anything legal, anything a customer will remember for years. Not because a model cannot produce an answer, but because you cannot buy enough confidence in the answer to stop checking it — and a step you still check is a step you have not removed.
Also anything undefined. "Automate our operations" is not a workflow; it is a department. Systems get built around specific, boring, nameable sequences.
Why the plumbing matters more than the model
The model is the part that changes every few months. Everything around it — how work arrives, how it is checked, what happens when a step fails, where it hands back to a person, what gets logged — is the part you actually live with, and it is where a project succeeds or quietly stops being used.
Build so the model can be swapped without rebuilding the system around it. Anything else is a bet on a vendor's roadmap, made with your operations.
How do you pick the first workflow?
Pick the one that is annoying, frequent, and low-stakes when it is wrong. You are not trying to capture the biggest prize first; you are trying to learn how your own business behaves when a system starts doing part of its work, while the cost of being wrong is a shrug.
Then measure the thing you meant to fix — hours, error rate, time-to-response — before and after. An automation nobody can prove worked will be switched off the first time it is inconvenient.