Hiring the Infinite Employee

In the last post we laid out the gap: AI adoption outside health care is the fastest ever measured, while adoption inside health systems moves on a different clock. The easy explanation is bureaucracy. We think the real explanation is more interesting, and more useful.

Here it is: deploying true AI in a health system is not buying software. It is hiring an employee.

Not metaphorically. Procedurally. Walk through what a health system's AI governance process actually requires, and you will find you are running a hiring process - every step of one - for a candidate who happens to be a machine.

A machine hired to think

Start with what AI actually is. The OECD definition - the one adopted into the EU's AI Act - describes a machine-based system that infers, from the inputs it receives, how to generate predictions, content, recommendations, or decisions. Britannica's is older and blunter: the ability of a machine to "perform tasks commonly associated with intelligent beings" - to reason, to discover meaning, to generalize, to learn from experience.

Those definitions could seem to describe a human worker more than a simple tool.

Every previous technology a hospital bought - the CT scanner, the IV pump, the EHR itself - was equipment. It did exactly what it was told, the same way, every time, and judgment stayed with the human operating it. That is why we credential the radiologist and not the scanner.

AI is the first technology that arrives as the judgment. It absorbs information, reasons over it, and produces recommendations and decisions - the exact activities a health system has only ever been able to acquire one way: by hiring people. And the machinery a century of organizational life built for acquiring judgment is not procurement. It is hiring. So when a governance process starts asking hiring questions of a software product, that is not confusion. It is the institution correctly recognizing what is "walking" in the door.

The six hiring questions

1. What problem are we trying to solve, and what type of employee do we need to solve it? Every real hire starts with a vacancy, and "this tool is impressive" is not a vacancy. But notice what happens the moment you ask the question honestly inside a health system: everyone has one.

The physician is drowning in the EMR - so the vacancy is an ambient scribe. The patient cannot figure out how to get an appointment - so the vacancy is an intake agent that interviews the patient and books the right visit. A nurse spends her week combing charts to fill quality registry submissions - so the vacancy is an abstraction assistant. The payor wants patients discharged sooner and to lower-cost settings - so the vacancy is utilization review, increasingly automated. Finance, HR, supply chain, and compliance each have their own version.

This is not new. It is how health systems came to look the way they do: for a hundred years, every part of the organization solved its problems by hiring staff to work on them. That is why a health system today serves so many internal "customers" - patients, physicians, nurses, administrators, payors - each with a legitimate claim on resources. AI does not simplify that reality; it walks straight into it. Every corner of the organization has a real vacancy an AI could fill, and every one of those requisitions lands on the same governance committee's desk. The prioritization fight is fully underway before the first interview is ever scheduled.

2. What is their job description? Once a vacancy is chosen, someone has to write the role: the exact tasks the AI performs, and - just as important - what it may not do without a human sign-off. Vague job descriptions get vague results from people. They get dangerous results from AI.

3. How do we know they can do the job safely? For a physician, we call this credentialing, and we would never skip it. For AI, the same instinct produces tiering: the more clinical the work, the higher the credential bar, and the more validation, references, and supervised practice required before we trust the work. A note-drafting tool is a Tier 1 hire. A tool that touches diagnosis is a Tier 3 or 4 hire, and it gets interviewed accordingly. Stanford Health Care has formalized this into its FURM assessment - an ethics review, usefulness simulation, financial projection, and monitoring plan that now governs every AI adoption there. The industry as a whole is behind: by one recent analysis, roughly 75% of US health systems have deployed or plan to deploy AI, while only 18% have mature AI governance.

4. What does this employee cost? Not the salary - the total compensation. The license fee is the salary. Implementation is recruiting and onboarding. Infrastructure is the equipment. And then there is the piece almost nobody budgets: the ongoing management time the hire consumes.

We went looking for defensible numbers, and here is our working estimate, triangulated from published frameworks and industry planning guides. A lightweight tool - no protected data, no clinical workflow - needs a named owner and periodic check-ins: roughly 200-300 loaded staff hours a year, or $20K-$40K. A standard tool touching protected data or operational decisions adds subcommittee review, monitoring dashboards, and IT support: roughly 800 hours, or $75K-$150K a year. A deep clinical deployment - EHR-integrated, shaping patient-facing decisions - requires the full apparatus: multi-stakeholder assessment, drift monitoring, quarterly review, and compliance documentation, on the order of 1,500-2,000 hours a year, or $150K-$300K. Every year. Before the vendor is paid a dollar.

If those numbers sound high, consider the anchors. A peer-reviewed staffing analysis puts the loaded cost of frontline clinical review staff at roughly $53 an hour - and AI supervision draws on informatics, compliance, and clinical leadership time that costs considerably more. Industry planning guides put AI monitoring systems alone at $30K-$100K a year and ongoing regulatory compliance at $50K-$200K. And the monitoring is not optional: a 2025 study across seven hospitals found a meaningful accuracy drop in a deployed model after a routine lab-test change - exactly the kind of failure a one-time vendor review would never catch. AI does not manage itself. Someone reviews the drift reports.

5. Who manages them after they are hired? Every hire needs a manager and a performance review that never stops. In practice this becomes a governance subcommittee (the hiring manager) reporting to a full AI governance committee (the HR of record), with re-review triggered by performance drift, regulatory change, a near-miss, or the tool expanding its own job duties.

6. When does the budget allow it? Now stack all five questions on top of the annual budgeting process most health systems run. An AI opportunity identified in the spring may be reviewed in the summer, credentialed in the fall, budgeted in the winter, and deployed - in cohorts, with a pilot first - the following year. None of the individual steps is unreasonable. Run sequentially, they take twelve to twenty-four months. The technology, meanwhile, ships a new generation every six.

Every AI hire is an outsourcing decision

Hiring a person is simpler than hiring an AI in one final respect, and it makes the AI decision harder, not easier.

Every AI "employee" is work performed by a vendor's system instead of a local person. That is a spreadsheet question in most industries. In ours, it is a community question: health systems are the largest employer in many of the markets they serve - we are in ours. An institution whose mission includes the economic health of its region will not, and arguably should not, be casual about outsourcing functions its neighbors currently perform. The honest framing is not "AI replaces jobs" but "AI hiring is offshoring to the cloud," and mission-driven systems feel that tension in a way a national retailer never will.

An infinite workforce - but only part of one

Here is what makes this moment genuinely new. For the first time, health systems face something like an infinite labor supply: for a price, we can hire as much knowledge work as we want, in any function, starting next quarter.

The italics matter. A job is a bundle of tasks, and AI can only take over the knowledge-work portion of the bundle. For some roles, that portion is nearly the whole job: scribing, coding, claims follow-up, chart abstraction, prior authorization. For others - physicians, nurses, therapists - it is a fraction. Boston Consulting Group estimated that more than half of US jobs will be reshaped by AI within two to three years but far fewer replaced, precisely because most workers do a range of things and AI can only do some of them. Stanford's labor data says the same thing from the other direction: the least AI-exposed work in the economy is relational and physical - the home health aide's job barely registers on exposure indices, and employment there is growing.

So the infinite workforce is specifically the knowledge workforce: infinite scribes, schedulers, coders, abstractors, and analysts. It is not infinite nurses, infinite surgeons, or infinite hands to hold. And that reframes the opportunity in a way we find genuinely hopeful: for clinicians, an infinite knowledge workforce means something closer to infinite support staff. Every physician with a scribe. Every nurse with an virtual expert consult. Every care manager with an analyst. The humans who do the irreplaceably human parts of care, surrounded by as much cognitive support as they can productively use - more effective and more present, not more optional.

The binding constraint is no longer whether talent exists. It is our capacity to write the job descriptions, run the credentialing, fund the total compensation, supervise the work, and decide - deliberately - which work stays human.

So the strategic problem becomes one health systems have never had to solve: given an infinite applicant pool, where do you hire first? And how do you make each hire economically viable once you count the management time nobody budgets? We have started work on this concept of building internal tools to see our AI portfolio exactly this way: every tool read as a hire, with a job description, total comp, a timesheet, and a performance review, on one page. The hiring questions are not the problem - they are legitimate, and most of them protect patients. The problem is answering them sequentially, artisanally, one requisition at a time, on an annual budget clock, in an era of infinite applicants.

The next post is about how to answer them faster - what startups owe us, what we owe ourselves, and where Epic fits.