Manu Khetan is Founder and CEO of Rolling Arrays, a LinkedIn 'Top Voice' Influencer and Creator of the R7 Framework.getty​Every job in your company is about to be split in two: the part an AI agent can do, and the part only a person can do. The splitting is not the hard part. Knowing where to cut is.At a roundtable I hosted in Kuala Lumpur in July, a group HR leader from an oil and gas firm told the table what his board had already told him. They had seen what AI could do, and their conclusion was to cut the workforce by a third. He did not agree. He also could not say why not in a way that would survive a board meeting.That gap, between what a leader believes and what they can defend, is where expensive decisions get made.Seventeen years inside enterprise workforces have shown me the same pattern, and it is not "AI takes jobs." It is quieter. A company automates the visible part of a role, books the saving and two quarters later something it was quietly relying on is gone. That thing was never a line item, so nobody watched it go.It keeps happening because our most trusted instrument, the P&L, cannot see the difference between the two kinds of work humans do.Two Kinds Of Human WorkSome human work is low variance. Four people in a meeting to make a one-person decision. The report nobody reads. The reconciliation an agent now does faster and cleaner. Whoever does it, the output is about the same. Which is why it is safe to automate.Other human work is high variance. The judgment on the deal that does not fit the model. The relationship that renews a contract a weaker operator would have lost. The read of a situation that stops a mistake nobody will ever know was avoided. Here the result swings hugely with who is doing it.The low-variance work is the part you can see. The high-variance work carries the value. On a profit and loss statement they look identical. Both are a salary. The four-in-a-meeting and the one irreplaceable judgment sit on the same line, often inside the same person.So when you redesign by cost, you take out the part you can see and leave the part you cannot. And if you measured that role by its execution all along, you have no way to tell whether the valuable part still works.Run a rough version. A role costs 100. Perhaps 60 is execution you can automate and 40 is judgment and trust. Automate the 60 and the cost sheet looks excellent. But if a slice of that 40 was riding on the same person and walks out with the 60, you did not save 60. You paid to remove value you were never pricing. Those numbers are illustrative. The shape is the point.When It Happens To A Good CompanyKlarna ran this experiment in public. The first half worked, and they had the courage to correct the second half.In February 2024, Klarna reported that its OpenAI-powered assistant had handled 2.3 million conversations in its first month, which the company estimated as the equivalent work of 700 full-time agents. Resolution time fell from 11 minutes to under two, on a projected $40 million benefit for the year. Every number said the redesign was a success.Fifteen months later, CEO Sebastian Siemiatkowski said cost had become "a too predominant evaluation factor" and the result was lower quality. Klarna began recruiting human agents again for the moments that needed them.Read that through the two kinds of work. Execution moved to the agent and went well, because that part was overhead. Judgment and trust moved too, and that part was the value. The spreadsheet could not tell them apart, because on a spreadsheet they were the same number.The customers told them. Customers always do. The question is whether your measurement lets you hear it before the cut, or after.Where The Value Is Already VisibleAt that same roundtable, the CEO of a specialist hospital group described her AI program. Assisted radiology, so reads come back faster and more accurately. A robotic pharmacy, valued less for speed than for driving medication error toward zero. Precision medicine on large clinical datasets.She never mentioned reducing headcount. Not once.Not because healthcare leaders are more principled than the rest of us. Her sector already measures the thing a spreadsheet cannot see. Medication error rates exist. Diagnostic accuracy is tracked. The value of a human judgment sits on a dashboard, so nobody proposes automating it away.Most functions never built the equivalent measure. Their value is not less real. It is undefended.The Wrong Layer Of The Right RolesThe real risk is not that we automate too much. It is that we automate the wrong layer of the right roles, and find out after the number that matters has moved.That boundary is what I call the Human Line, the boundary between the work in a role that can move to an AI or robotic agent and the work that must stay human—judgment, trust and the relationships that carry value. It has to be drawn deliberately, task by task. If you do not draw it, a cost model draws it for you, working with bad information.Cut the work, not the people. The work splits. The value-creating humans move above the agent line, not out the door.The leader with the board instruction left the room with that sentence. He will not argue against the third. He will change the question to which work moves, which work stays and what the business loses if we get that wrong.Look at the salary line of the next team someone proposes to optimize. Can anything you measure today tell you which part is overhead and which part is value? If not, this is not a cost decision. It is a blind spot.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?