If you've done ops for any length of time, you know the ticket that dies in somebody else's queue. You file a problem with the team that owns the thing, and then you stop checking. They have less context than you and a dozen louder fires. Nobody decides to drop it. It just sits in the gap between two queues.
Medicine has the same failure, and the stakes are higher. Kids with diabetes need regular eye exams, because the disease can quietly damage the retina long before a child notices anything wrong. The standard move is a referral: here's an eye clinic, here's a handout, please go. Only 35 to 72 percent of young people with diabetes actually get the recommended exams.
A trial called ACCESS, run at two Johns Hopkins pediatric diabetes clinics in Baltimore, tried skipping the referral altogether. It enrolled 164 patients aged 8 to 21, nearly half of them on Medicaid, the public insurance for low-income families. Half got the usual referral. The other half had their eyes photographed during the diabetes visit they were already sitting in. An AI program read the images in about a minute, and families got the result before they walked out of the building. (An eye specialist also reviewed every image as a backstop.)
Of the kids analyzed, 18 of 82 in the referral group got checked. In the same-visit group, 81 of 81 did. For every kid whose result came back normal, the second trip never had to happen.
This section has argued, over and over, that automation mostly moves work around: into checking output, handling the weird cases, mopping up the mistakes. Usually that's right. ACCESS is a clean case where a step got removed rather than relocated, and it deserves as much attention as the cautionary tales get.
It has limits. Twenty-five kids got results suggesting eye disease and still had to see a specialist; within six months, 16 went and 9 didn't. The referral group never produced a matching set of flagged kids, so there's nothing to compare that number to. The study also changed two things at once, an AI reader and an in-visit exam, so we can't say which one did the heavy lifting. It counted exams, not eyesight saved. And one author has financial ties to the device maker.
Still, the trip that's left is a different trip. Instead of everyone leaving with a handout, a small group goes, each of them carrying a specific finding. Nine kids who didn't follow through is a short list with names on it, and a clinic can chase a list like that.
Most of the agent projects I see are trying to smooth handoffs: faster routing, automatic booking, reminder texts. Useful! But all of that assumes the handoff has to exist. The harder thing to figure out is whether a step exists because the person needs an expert, or because the expertise used to live in another building, another team, another system with its own damn login. Think of the customer told to open a separate ticket with billing, or the employee filing a form so some other team can approve routine access. Some handoffs protect something real. A lot of them are leftovers from wherever the know-how happened to be sitting.
If agents make that know-how portable, a lot of "go see someone else" starts to reflect an old org chart more than the work itself. One possibility is services redrawn around where the person already is, with referrals kept for cases where somebody else really does have to act.
So walk your workflows, mark the handoffs that exist only because of where the skill used to be, and cut them. Then staff the ones you keep like they matter, because 16 of 25 says that's exactly where people still fall out.