A manifesto for health system leaders
The Unconducted Orchestra
An Operating Model for Healthcare’s AI Workforce
Every health system now runs a second workforce, one made of AI, and no one is managing it.
01
The workforce no one hired
Your health system took on hundreds of new workers this year. No one interviewed a single one of them.
They hold logins and permissions and take actions, the same as the people you employ, but not one has a badge, a manager, or a personnel file. You may not think of them as workers. You know them as AI features or agents, already drafting inbox replies, working prior authorizations, and running inside workflows someone stood up on a Friday afternoon. Some you switch on with a checkbox inside Epic, some you buy, some your own people build over a weekend. None of it feels like a hiring decision. All of it is one.
Now ask three questions:
- How many are running across your health system?
- What is each one allowed to touch?
- Who answers when one gets it wrong?
For almost every health system, no one can answer. This is a second workforce, the fastest-growing one any health system has ever had. It works at machine speed, around the clock, and no one is in charge of it.
Your people
Badged. Reviewed. On the record.
Your agents
Unbadged. Unreviewed. Unseen.
02
Why this has to exist
Adopting AI is easy. Running it is hard.
Every enterprise technology learned this before AI did. Software earned an operating model before it could scale: version control, testing, staging, the ability to roll back. Enterprise IT earned another: one record of every asset, a change process, a way to answer for what is running. These disciplines are invisible when they work, and they are the only reason a large health system runs thousands of IT systems without chaos. AI has arrived at scale with none of them. It is the most capable technology a hospital has ever deployed, and the first to arrive powerful before it arrived governed.
- Software
- Software development lifecycle
- Enterprise IT
- IT service management
- AI
- No operating model yet
- Overture
- The AI operating model
That gap would matter in any industry. In healthcare it is dangerous. For a generation, bringing technology into a hospital meant installing software that did what it was told. An agent is different. It reasons, chooses, and takes the next step on its own. In healthcare, a wrong step is not a bad sentence on a screen. It is a missed diagnosis, an erroneous order, a canceled follow-up, a denial built on the wrong policy. When an agent acts on a mistake, the error does not stay an output. It becomes an event.
Healthcare never got to move fast and break things. Here, the things that break are people.
The danger is not only that agents are wrong. It is that they are powerful. An agent given full platform access instead of a single task will read an entire chart to answer one question, failing minimum necessary by design, and it reaches across the boundaries your policies work to maintain, into behavioral health records or a department it was never scoped to touch.
An over-entitled employee makes mistakes at human speed. An over-entitled agent makes them by the thousand.
So this is a question of accountability, not of tooling. Picture the board meeting after the complaint escalates, the regulator writes, or an attorney subpoenas two years of prompts. The directors will ask five things. What AI is running here, all of it, not just what IT bought. Who approved it, and against what policy. When one starts to go wrong, can we catch it and shut it down before the mistake repeats a thousand times. Is it delivering the value we were promised. And can we prove what it did, why, and under whose authority.
“The model decided” will not survive an audit.
And the audit is closer than the market thinks. Federal transparency rules for AI in the record, new accreditation expectations, and a widening patchwork of state laws are landing over the next several months, and every one of them assumes you can produce what you ran, why, and to what effect. The reporting bar only tightens from here.
03
What Overture is
No platform can answer for you. Epic governs Epic and Microsoft governs Microsoft, and no one governs the space between them, which is exactly where cross-vendor agents, bought tools, and shadow AI live. A dashboard for one model cannot tell a board whether the whole program is safe and paying off. And a company that sells you the AI is grading its own homework. Hand your AI governance to a vendor that sells AI, and you have asked the fox to audit the henhouse.
Overture is that operating model. We call the category Enterprise AI Management: running all of a health system’s AI as one program, every model and agent seen, measured, and kept safe on a single record. The layer sits above every platform where AI is built or run, owns none of them, and has one job, to make the whole program work, from the first intake decision to the last measure of value. It manages the AI you already run and deploys the AI you do not yet.
The health system
Overture · the operating model
Epic · Agent Factory
Microsoft
Cloud and frontier labs
Point solutions
In-house builds
Shadow AI
What makes it a distinct layer, not a feature of any platform:
Horizontal, not vertical. It spans every platform instead of going deep in one.
Independent, not conflicted. It sells none of the AI it manages, so its read on any model is unbiased, and you stay free to choose the right model for each job and switch when a better one arrives.
Assurance, not capability. It makes no AI of its own. It deploys and governs what others build, on the platforms you already run, so every AI is seen, safe, and answerable.
And Overture does not do this to you, it does it with you. We use the model our teams run in the field: Build. Operate. Transfer. We build the operating model alongside your people, operate it with them in production, and transfer it to your team as they are ready, so the capability ends up yours.
Build
Stand up the operating model with you: the registry, the policies, the first governed deployments.
Operate
Our forward-deployed team runs it alongside your people, in production and at scale.
Transfer
Hand the management to your team as they are ready, so the capability stays yours.
04
How it works
You already know how to do this. A health system runs the most disciplined workforce in any industry. You do not let a cardiologist operate because a diploma arrived in the mail. You verify the training, grant privileges for specific procedures, and review them on a schedule. An agent deserves no less: credentialed for a defined use, privileged to the actions its task requires, supervised where the stakes are high, reviewed on its results, and retired when it stops earning its place.
The operating model
Manage agents the way you manage people.
- Hiring
- Risk-based intake: a defined use case and a committed outcome before any access.
- Role and badge
- Its own identity, with permissions scoped to the task.
- Supervision
- Human approval for anything irreversible or high-consequence.
- Performance review
- Measured against the value it promised: adoption, outcomes, cost.
- Personnel record
- An immutable audit trail of what it saw, did, and under whose authority.
- Offboarding
- An expiration date, a named owner, and retirement when it stops earning its place.
All of it runs on one system of record, the Score: every agent seen, vetted, governed, deployed, adopted, measured, and answerable. Overture does not bolt a new wall onto your systems. It sets your policy once and has each platform enforce it through the gateway it already has, so governance rides on rails that exist. And the work compounds: what your people teach that workforce builds up in a record you own, and every agent that proves out becomes the pattern the next one starts from.
- Seen
- Vetted
- Governed
- Deployed
- Adopted
- Measured
- Answerable
A layer is only as strong as the hands that run it, so Overture staffs it. A forward-deployed team of AI engineers, informaticists, and governance leaders runs the operating model with your people: standing up the registry and the committee, keeping evaluations current, and building the next wave into the platforms you already operate. It is continuous work, and we do it alongside your team so they carry more of it over time, not less. And because the layer spans many health systems, a failure caught in one becomes a warning to all of them, before it reaches a patient.
Do this well and the advantage compounds. The systems that learn to run their AI can take on more of it, safely and faster, while the ones that never got control slow under the weight of their own ungoverned sprawl. Managing the second workforce is not the tax you pay to use AI. It is how you win with it.
05
What we believe
Operating beats owning. How well you operate AI matters more than how much of it you bought or how early you adopted.
Structure is speed. The better a health system can govern its AI, the faster it can put new capability safely to work. Governance is not the brake. It is what lets you move fast at all.
What you cannot measure, you cannot steer. Measuring each agent’s value is how you know where to invest more and where to pause.
The work is never finished. Agents drift, workflows change, and better models arrive every month. Staying ahead is the job, not a phase.
Safe and valuable are the same discipline. They are not competing goals.
Accountability is infrastructure. Every health system will need it, the way it came to need a record for its patients. And it has to be independent to be trusted, because no one should grade their own homework when a patient’s care is at stake.
06
The call
The standard for managing healthcare AI is being set right now, by someone. It should be set on purpose, by a layer whose only job is to keep the AI and the patients safe at the same time, and whose interests sit on the health system’s side of the table.
Every health system now runs a second workforce. The only question left is whether anyone is in charge of it.
We built Overture to be the answer.
Every agent, in concert.