This is a guest post by By Oliver Kharraz, MD, Zocdoc founder and CEO of Zocdoc.
AI agents are getting astonishingly smart. But I believe the biggest leap is still ahead of us: a world where we stop prompting agents to do things, and they start prompting us.
Today, we tell agents what we need. Find me a dermatologist. Book me a flight. Buy a ticket to a show. But imagine agents that anticipate our needs before we prompt them, determine what to do about them, and then just ask for our approval.
That's where the agentic economy is headed, and ultimately, it's what the best agents will compete on. But achieving this requires something fundamental: the ability to reliably get things done in the real world. And that depends on infrastructure the AI industry is dramatically underestimating.
You Cannot Prompt Your Way to Reality
Even the smartest agents are only as good as the real-world systems they interact with, and many are fragmented, outdated, or inaccessible to autonomous software. An agent can find a dermatologist in seconds. But can it verify that the doctor is real and licensed, accepts your specific insurance plan, has an appointment Thursday at 2:45 p.m., and actually book it?
It turns out, an agent can’t reason an appointment into existence or guarantee a hotel reservation without access to actual inventory. It needs verified information, permission to act, and confirmation that a transaction succeeded.
As agents become more autonomous, the stakes rise. An agent relying on unverified information could book an appointment that doesn't exist or mistake a scam for a legitimate business. And consumers will pay the price. Ultimately, agents will too when they lose consumers’ trust. And solving these real-world challenges is harder than it looks.
The Last Mile Is the Hardest
I learned this the hard way building Zocdoc. Nearly twenty years ago, we set out to do something seemingly simple: help people book doctor's appointments online. We quickly found out that the front end was easy. What sat beneath it was the hard part.
Unlike Sabre for travel or Stripe for payments, there was no universal scheduling infrastructure to plug into. We had to build it ourselves, spending nearly two decades and millions of dollars connecting disparate provider organizations, their fragmented scheduling systems, accepted insurance plans, and highly bespoke booking workflows.
How does all of this complexity show up in practice? Let me count the ways.
A doctor might accept Aetna but not a patient's particular plan. A 3:30 opening might be reserved for existing patients. An elbow consultation might require fifteen minutes while a knee requires 30. The front desk might stick a Post-it note on their computer saying "Doctor Out Friday" without blocking the scheduling software.
Multiply that by thousands of idiosyncrasies. Solving these problems took years of unglamorous work. But underneath our marketplace, we built the access infrastructure healthcare was missing.
AI agents now face the same reality: knowing what needs to happen is very different from making it happen. And until agents can reliably connect to real-world systems, verify what's available, and complete transactions successfully, they can't deliver on the possibility of proactively getting things done.
From Responding to Anticipating
Reliable infrastructure is what will unlock the next generation of agents. Today, we judge agents by how well they respond to our prompts and requests. Tomorrow, I believe they'll compete on how well they prompt us. How much can they anticipate our needs and bring us solutions that are ready to act on?
Consider preventive healthcare. Today, patients must remember when they're due for screenings, find doctors, verify insurance, and schedule appointments. Now imagine an agent that, with your permission, knows when you're due for a screening. Because it can connect to verified healthcare infrastructure, it can find an appropriate in-network doctor with real availability and say, "You're due for this screening. I found an appointment next Thursday at 11:45 am, when your calendar is open. Want me to book it? Say the word."
You haven't prompted the agent. It has done the work and come to you with a solution; you just confirm. This kind of proactive support could extend far beyond healthcare. The best agents will anticipate what we need, what's actually possible, how to reliably get it done, and when to ask for our approval.
The Real World Is the Next Frontier
I get lots of concerned questions about what AI agents mean for Zocdoc. I see it differently. I believe agents will make transaction infrastructure more valuable, not less. Agents may bypass consumer interfaces for platforms like ours, but the infrastructure beneath them becomes more important as they take on greater responsibility and transact.
That's why we recently launched Welcome, Agents, opening the healthcare access infrastructure we've spent nearly two decades building to agents acting on patients' behalf. It's an extension of our mission to give power to the patient: why make patients come to our marketplace when we can help them reliably get access wherever they start their search?
And as agents evolve from responding to prompts to anticipating needs, that infrastructure will help them to go much further. The same systems that make it possible to book an appointment when prompted could help an agent proactively line up care before a patient thinks to seek it.
Across industries, the next generation of AI agents will compete on how well they anticipate our needs and bring us solutions ready for approval. But to do this, they’ll need the proven infrastructure in place to deliver on those solutions in the real world.

About the Author
Oliver Kharraz, MD, is CEO and founder of Zocdoc. Oliver is the most recent doctor in a 300-year family tradition. Over the course of his wide-ranging career, Oliver has accrued comprehensive experience effecting change and building efficiency in large scale healthcare organizations using information technology.

