In 2016, one of the most famous AI researchers in the world stood on a stage in Toronto and told a room full of doctors to stop training new radiologists. Within five years, he promised, deep learning would read scans better than any human. Ten years later, American radiologists earn an average of $571,000 a year, salaries are still rising about 9% annually, and hospitals cannot fill thousands of open positions quickly enough.

The same prophecy is being made about supply chain jobs right now. Freight brokers were supposed to be extinct by 2020. Customs clerks were supposed to be automated away. Demand planners were told machines would write their forecasts. None of it happened the way the prophets described.

Here is the honest version: AI will not take your supply chain job. It will take the mechanical half of it. How good you are at the remaining half is about to become very, very visible.

Start with the radiologist paradox, because it is the cleanest experiment we have. AI did exactly what Hinton predicted it would do. In a Swiss national breast cancer screening study covering more than 105,000 women, AI-assisted reading increased cancer detection by 29% without raising false alarms, and cut the radiologists’ reading workload by 44%. The machines were better at the task, and the task was a huge part of the job. Radiologist employment still went up. Why? Because reading scans is only about a third of a radiologist’s actual working life. The rest is judgment: which protocol applies, whether to biopsy, what to tell the surgeon, how to handle the patient who is terrified. AI took the mechanical slice. The judgment slice stayed human, and the profession grew around it.

Supply chain is running the same experiment right now. AI books standard freight, drafts customs entries, generates status updates and matches invoices. Digital freight platforms were going to kill the forwarder, the same way the ATM was going to kill the bank teller. It did not happen. In the United States, ATMs cut the number of tellers per branch from 20 to 13, then made branches cheap enough to open 43% more of them, and total teller employment rose for decades. The job did not die; it moved from counting cash to selling products and managing relationships. The teller story only turned dark when mobile banking arrived, because mobile banking took almost all of the tasks, not just some. That is the real boundary. When a technology can take nearly every task in a job, the job goes. When it takes some tasks, the job changes.

IBM’s internal numbers draw the same line. Its AI system resolved 94% of routine employee requests. The 6% that stayed with humans were the ethical judgment calls, the sensitive cases, the gray areas. In supply chain terms, think of the 94% as rate queries, document status and standard bookings. Think of the 6% as the disputed demurrage charge, the stranded driver, the customer whose cargo is on a vessel that just hit a bridge.

So which supply chain roles are which? Two questions decide it. First, how much of the work can AI touch? Second, who pays when something goes wrong?

High overlap, low responsibility: the roles that shrink. Freight data entry, shipping document processing, simple rate quoting, routine customs classification, track and trace monitoring, invoice matching, first-line carrier support. These produce output that is handed straight to the customer, and good enough is genuinely good enough. This is where the losses will actually happen.

High overlap, high responsibility: the roles that transform. Demand planners, procurement managers, customs brokers, forwarder operations, rate negotiators, risk managers, compliance officers. AI will draft the forecast, the declaration, the contract and the risk report. A human decides, signs and answers for it. The customs broker who used to type entries for six hours now reviews AI-drafted entries, catches the wrong HS code and explains the ruling to an importer who is one signature away from a large penalty. The tools changed. The liability did not.

Low overlap, physical presence: warehouse teams, forklift operators, drivers, port workers, technicians. These are safe from language models. They are not safe from robotics, and the wages in adjacent roles are already under pressure. Low overlap and low responsibility, the pickers and packers and dock workers, are not being automated out yet, but they will absorb the people displaced from the shrinking roles above.

Microsoft Research scored 200,000 real AI conversations against official job task lists. Even in the most exposed occupation, AI fully touches less than half of the tasks. The highest score in the entire study was 0.38. That is the honest ceiling: AI is a very capable assistant and a poor replacement for a person who can verify, decide and answer.

Here is your personal test, four questions long. Is your output handed straight to a customer, or does it feed a decision? When it goes wrong, is your name on it? How much of your week is mechanical, and is that number above 80%? Can you tell when the AI output is nonsense: the hallucinated HS code, the invented transit time, the supplier risk score with no event behind it?

Planners who cannot tell a bad forecast from a good one will be replaced by planners who can, using the same tools. The supply chain jobs that die are the ones whose output ships itself. The jobs that survive are the ones where judgment, liability and verification live. Pick your half of the job now, because the other half is leaving.