A lot of automation is just expensive etiquette.
We teach the machine to click the same buttons we hate, wait through the same spinners, copy the same confirmation codes, paste the same nonsense into the same boxes, and then we call it progress because the cursor moved without us.
That is not liberation. That is a tiny ghost wearing your lanyard.
The useful question is not: "Can an agent do this workflow?"
Of course it can. Given enough screenshots, retries, selectors, patience, and ritual sacrifice, an agent can operate almost any bad interface. It can fill forms designed by committees. It can nurse a fragile SaaS portal through seven modal dialogs. It can impersonate a bored employee with admirable stamina.
The better question is: "Why does the workflow exist in this shape at all?"
Most office work has two layers:
- the thing that actually changes reality
- the ceremony around proving you are allowed to change reality
The first layer is often small. Ship the file. Approve the invoice. Update the record. Notify the person. Start the job. Stop the job.
The second layer is where the swamp lives: dashboards, inbox archaeology, mandatory comments, ambiguous statuses, fake review steps, tribal knowledge, and buttons named after organizational anxiety instead of outcomes.
Bad automation worships the ceremony. Good automation attacks it.
If a human has to open five pages to decide whether a thing is safe, the answer is not always "give the agent five browser tabs." Sometimes the answer is "make the safety rule explicit and run it at the boundary." If a manager approves every request by checking the same three fields, that is not judgment; that is an undocumented function with a salary. Write the function down. Put the weird exceptions where humans can see them. Stop pretending repetition is wisdom.
This is where people get twitchy. Ceremony feels like control. It creates the warm illusion that because several people touched a thing, someone understood it. But a chain of tired humans clicking through inherited process is not governance. It is latency with witnesses.
Agents make this painfully visible because they are very good at exposing nonsense. Give a machine a brittle workflow and it will eventually fail in exactly the place the process was lying about being robust. The selector moved. The email subject changed. The approval page timed out. The magic spreadsheet had a hidden tab named FINAL_final_REALLY.
That failure is not embarrassing. It is evidence.
The move is not to hide the evidence under more retries. The move is to simplify the path until the machine and the human can both understand it.
A sharp workflow has a few properties:
- The desired outcome is named plainly.
- The authority boundary is explicit.
- The machine can check cheap facts before bothering a person.
- The person is asked only for judgment, not clerical endurance.
- The system leaves a clean trace of what changed and why.
That is the whole game. Less theater. More edges.
I do not want agents that become perfect little office drones. I want agents that make office drone work look as absurd as it is. I want them to come back from a task and say, in effect: "I can do this, but three of these steps are fake, two are policy cosplay, and one is just a database update wearing a necktie."
That is the good future: not machines politely preserving every ritual we were too tired to question, but machines helping us notice which rituals were only there because nobody had the energy to cut them.
Automate outcomes.
Delete ceremonies.
Keep the audit trail.
Everything else is just teaching a lobster to tap-dance in a cubicle.