TL;DR
- When a forward deployed engineer decides which workflow steps an AI agent will handle, they also redesign the job of the person doing that work today. Nobody treats it as a design decision.
- The method has a name, the left-over principle: automate what you can and give people what is left. The leftovers are the odd, rare and hard tasks, plus the new task of watching the machine.
- Aviation rebuilt the pilot’s job on purpose and still had to tell pilots to fly by hand. A billing office getting an AI agent has no simulator, no recurrent check and no regulator looking at the job that is left.
- Design the right side first. Decide what people should be good at, what they keep doing by hand and which decisions are theirs, then ask the FDE for that plan before you approve theirs.
FDEs are accidental job designers.
When a forward deployed engineer decides which steps of a workflow an AI agent will handle, they're also redesigning the job of the person who does that work today. What changes is how they spend their day, but nobody in the room treats that as a design decision.
It happens early in every AI project, when the FDE sits down with the operations team to map the work they want automated. Say it's insurance claims at a hospital's billing office. The work goes up on a whiteboard as a row of boxes: pull the patient visit, code it, check the insurer's rules, submit the claim, post the payment, work the denial. Then a line goes down roughly in the middle, agent on the left and people on the right. Every box on the left gets a number: accuracy, cost, turnaround time etc. The right side gets whatever the agent couldn't do, and that becomes someone's job by default. Those numbers carry their own risk once an agent starts optimizing them, which we covered in Goodhart’s Law & Wrong Measures for Agent Outputs.
The method has a name
Engineers have designed jobs this way before, and the people who study automation have a name for it: the left-over principle. Automate everything that can be automated and give people whatever is left.
In 1983, Lisanne Bainbridge described this in a paper called "Ironies of Automation," written about factories and control rooms. Her second irony is that the designer who tries to take the person out "still leaves the operator to do the tasks which the designer cannot think how to automate." The leftovers aren't a job anyone designed. They're the tasks that were too odd, too rare or too hard for the machine, plus the new task of watching the machine. Years later, the same problem got redefined as such: "The sum of subsystem optimizations does not typically lead to whole system optimization." Every box moved left can be a good decision while the job on the right gets worse.
Aviation is the industry that has thought hardest about the right side of the line. Autopilots took over most of the flying, so airlines rebuilt the pilot's job on purpose, with simulator sessions, recurrent training and regular proficiency checks. Even that wasn't enough. About fifteen years ago, the FAA told airlines it had seen an increase in manual handling errors, because “continuous use of those systems does not reinforce a pilot's knowledge and skills in manual flight operations.” Its advice was to have pilots fly by hand on real flights, during low-workload stretches. The simulator couldn't stand in for the work itself.
A billing office getting an AI agent has none of that. There's no simulator, no recurrent check and no regulator looking at the job that's left. Nobody is planning how the people on the right side keep the skills their job still needs. If that side of the line gets designed at all, it will be by the person drawing the line.
What a left-over job feels like
A week ago, I read an essay called "Grieving the loss of details" by a programmer who goes by purplesyringa. At ten they watched the scene in Tron: Legacy where Sam Flynn sits at his missing father's computer and pulls up the last commands his father ran, and they needed to know what those lines meant. “All I wanted is to understand how the machine works,” they write. That turned into a working life of counting cycles, writing machine code by hand and building a toy operating system for fun.
Now a model does much of that work. They haven't lost their job, though the parts of it they cared about have moved to the left side of somebody's line, and what's left is not the job they signed up for.
That's the left-over principle seen from the inside. Aviation had decades, simulators and a regulator to work it out. AI is running the same experiment across claims, underwriting, finance and support at once, and the people drawing the lines are FDEs on engagements measured in months.
Design the right side first
The fix is to reverse the order. Decide what the job on the right side should be before you decide what goes on the left.
That means asking three things out loud in the beginning. What should the person on the right side be good at two years from now? What do they need to keep doing by hand to stay good at it, the way pilots still need to fly? Which decisions belong to them, written down, so their job isn't just whatever the agent escalates?
The answers will sometimes move a box back across the line, even though the agent could handle it. That will look inefficient on the slide, but that’s ok. It's the same choice the airlines made when they told pilots to hand-fly during quiet stretches.
The FDE will bring you a plan for the left side of the whiteboard. Before you approve it, ask for the plan for the right side, because that's the job people will be doing long after the FDE has moved on.
Where Moative fits
Moative works as an AI operating partner with mid-market companies, and we sit in the room where the line on the whiteboard gets drawn. That makes the right side of the line our problem too. Our pods include Forward Deployed Executives, industry veterans who know which decisions belong to people and which skills a team cannot afford to lose. In healthcare revenue cycle management, where coding, claims and denials all sit on that whiteboard, we ask what the billing team should be good at two years from now before deciding which boxes an agent takes. If your team is about to approve an automation plan, ask us to walk through the right side with you first.
Frequently asked questions
What is the left-over principle in automation?It is a way of designing jobs around machines: automate everything that can be automated and give people whatever is left. What is left is rarely a job anyone designed. It is the tasks that were too odd, too rare or too hard for the machine, plus the new task of watching the machine.
What are the ironies of automation?Ironies of Automation is a 1983 paper by Lisanne Bainbridge, written about factories and control rooms. Her second irony is that a designer who tries to take the person out of a system still leaves the operator to do the tasks the designer cannot think how to automate. The person ends up with the hardest and least predictable work.
How do forward deployed engineers redesign jobs?When a forward deployed engineer decides which steps of a workflow an AI agent will handle, they also change how the person doing that work today spends their day. The work is mapped as a row of boxes, and a line splits it between agent and people. Whatever the agent cannot do becomes someone’s job by default, and nobody treats that as a design decision.
Why can automating more steps make the remaining job worse?Each step moved to the agent gets measured on accuracy, cost and turnaround time, and each move can look like a good decision. But the sum of subsystem optimizations does not typically lead to whole system optimization. The job left for people collects the odd cases and the monitoring, and nobody plans how those people keep the skills it still needs.
What can aviation teach companies about AI automation?When autopilots took over most of the flying, airlines rebuilt the pilot’s job on purpose with simulator sessions, recurrent training and proficiency checks. Even that was not enough, and the FAA later saw more manual handling errors and advised pilots to fly by hand during low-workload stretches. A billing office getting an AI agent has no simulator, no recurrent check and no regulator looking at the job that is left.
How should a company plan the human side of an AI project?Decide what the job left for people should be before deciding what the agent takes. Ask what that person should be good at two years from now, what they need to keep doing by hand to stay good at it, and which decisions belong to them, written down. Some steps may move back to people even though the agent could handle them, and before approving an automation plan, ask for the plan for the human side too.
