Every operations manager has a process document. Very few processes run the way the document says.
The real workflow lives in the people who run it: the spreadsheet that bridges two systems, the colleague who knows which supplier to call, the exception that happens every Tuesday. AI built on the document inherits none of that.
Leaders and workers see different results
Camunda's September 2026 survey asked leaders and employees whether AI was making them more productive. 89% of leaders said yes. 64% of employees agreed.
The same survey explains part of the gap. 69% of employees said they were not fully consulted about how AI would be used in their day-to-day work. 44% said they had overridden AI outputs because the underlying process was not set up correctly. And 65% said they could use AI much more effectively if it was implemented differently.
Those are not people resisting change. They are people telling you where the design went wrong.
What mapping the real work finds
When we map a workflow with the people who run it, the same things come up:
- Handoffs where work waits for someone to notice it
- Workarounds that exist because two systems do not talk to each other
- Exceptions that are rare on paper and frequent in practice
- Checks that nobody remembers the reason for
Some of these are where AI removes real work. Others should simply be removed. A few are the decisions that must stay with a person, and they are easier to protect once they are written down.
Three operations workflows, mapped first
Approvals and exceptions. Work routes itself, the next step and its owner are visible, and anything outside the rules stops and finds a person instead of guessing.
Shift and workforce scheduling. AI proposes roster changes from demand, skills and constraints. The manager approves anything outside limits agreed in advance.
Maintenance scheduling. AI ranks maintenance work by risk and suggests windows with the least impact on production. The planner decides.
Assisted first, autonomous later
Neptune's whitepaper draws a useful line between two modes. In assisted mode, AI informs and a person decides. In autonomous mode, AI acts within boundaries agreed in advance.
Start assisted. Move a step to autonomous only when it has a track record, the limits are written down, and someone reviews what it did every month.
Where to start
Pick one workflow and spend two weeks with the people who run it. Map it, agree one number to move, and write down what AI may never do. That is our diagnostic, and it is the cheapest part of any AI project to get right.