A new drug clears the FDA. Now the health plan has to decide whether to cover it, and that decision sets off a scramble: six or seven specialists, working mostly in sequence, two to three months, and somewhere around $100,000 by the time it’s done. Meanwhile, the patients wait.
For some drugs, the wait is an annoyance. For a schizophrenia drug, it’s something worse. When treatment gets interrupted, people end up hospitalized, and each of those admissions costs a plan somewhere between $8,000 and $15,000. A few hundred of them across one large plan and you’re looking at four to seven million dollars in costs that didn’t need to happen. All from one slow decision about one drug. I’ve seen this movie before. Eleven years at Dell — and a few more at startups — taught me how enterprises buy technology that’s supposed to fix exactly this kind of mess, and how rarely it does.
The COVID pandemic forced me to stop and reflect. As healthcare systems strained under extraordinary pressure, I found myself asking a simple question: “What am I going to tell my grandkids? That I helped people shop and search more with AI? Or can I do something more?” That question ultimately led me to healthcare. The mission became deeply personal when I later lost a friend to breast cancer~~,~~ and heard about their struggle to access experimental treatments buried within thousands of pages of medical records.
Healthcare doesn’t lack expertise. It struggles to get the right expertise to the right people at the right time. It is the same challenge I have seen play out across enterprises for years.
AI is running the same playbook now. MIT researchers analyzed more than 300 enterprise AI deployments last year and found that 95% of generative AI pilots produced no measurable return. Their diagnosis wasn’t that the models were weak. It was that organizations bolted AI onto processes that were never rethought for it.
Right now the entire AI conversation is about agents — how many you can deploy, how autonomous they are, how fast they move. The race is to build more AI workers. Almost nobody is building the thing that manages them. That’s the part that actually matters, and it’s the part nobody’s racing for.
The analysts have already priced in what happens next. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027~~,~~ and cites inadequate risk controls — alongside cost and unclear value — as one of three primary causes.
We’ve been here before. When the cloud showed up, the pitch was transformation. What most companies actually did was pick up their existing workflows, move them off their own servers, and set them down in the cloud. Same work, same thinking, bigger bill. AI is heading down the same road. Companies are automating one step here, one queue there, and calling it a win. But speeding up a broken process just gets you to the wrong answer faster.
Healthcare is hard for AI in exactly the ways that make it useful to learn from. The regulations are brutal. The decisions matter. The data sits in systems that were never built to talk to each other, and the people who can actually read that data — clinicians — are in short supply and getting shorter. The AAMC projects a shortage of up to 86,000 physicians by 2036, and nurses are stretched just as thin.
One health enterprise we work with has 600 nurses whose main job is prior authorization and payment integrity. Six hundred people who trained to care for patients, spending their days inside paperwork. Technology was supposed to fix that. Mostly it just turned the filing cabinet digital. The AMA’s 2024 physician survey found that prior authorization alone consumes an average of 13 hours of physician and staff time every week, and 93% of physicians say it delays patient care.
Healthcare simply makes the problem visible sooner. Financial services, insurance, energy, and government face the same challenge. Expertise is trapped inside workflows, regulations are increasing, and every decision requires traceability. The question isn’t whether AI can perform a task. It’s whether organizations can understand, govern, and trust the decisions AI helps produce.
We ran the math and found that for a pharmacy benefit manager or health plan, a new drug approval can trigger a process involving six or seven specialists — pharmacists, coders, actuaries, compliance counsel — all working largely in sequence over 60 to 90 days, at a cost of roughly $100,000 per drug. A large PBM runs 200 to 300 of these assessments every year. While that process runs, patients are in coverage limbo.
The same assessment, done with coordinated AI, takes four to eight hours. A clinical pharmacist reviews the output rather than producing it~~;~~ , the coverage gap closes, and direct labor cost drops by 97%. And everything the agents did is documented, so when a compliance question comes up later, you can trace exactly what informed decision. No one is hunting for answers buried in an email chain or locked inside someone’s memory.
The goal shouldn’t be automation. It should be changing what’s possible. Most enterprise AI deployments are invisible — accelerating steps a human would otherwise take without changing what that human does. But deploying a thousand agents that can’t be coordinated, governed, or audited is the same thinking that gave enterprises a thousand disconnected point solutions a decade ago~~,~~ and produced the same result.
Healthcare didn’t volunteer to go first. The pressure simply arrived there earlier — people are leaving because of the administrative burden, patients aren’t being cared for, and coverage processes are failing. It is a broken system that urgently needs help. That urgency is forcing the industry to confront something it hasn’t had to yet: not just how to deploy AI, but how to govern it, audit it, and make it work for the people it’s supposed to serve~~–~~ — not just the systems processing data in the background.
Every other regulated industry will face that same pressure eventually. The MIT and Gartner numbers suggest many already are. The question is whether they learn from the lessons of others now or wait until they’re at the same breaking point. What this work has taught me: you don’t need more AI agents. You need an operating system to manage them — and the clarity to change what your people actually do.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.