The definition
An Agentic Operations System (AOS) does the work of the imaging order, from the referral to the payment. It brings the order in, books and prepares the patient, settles coverage, benefits and authorization before the exam, delivers the report, and works the claim until it is paid. The patient and the referring office can do their part themselves, and the center’s people handle only the exceptions, on one record.
An AOS is measured by how many tasks finish without your staff touching them.
Within it, agents perform the operational work of an imaging order from the referring office to the payment: bringing the order in and keeping the referring provider informed, intake and booking, eligibility, prior authorization and estimates, patient communication and check-in, delivery of the signed report and images back to the referrer, charge creation, claim submission, remittance and collection. Rather than queue that work up for staff, the system completes it.
A CRM is a system of relationships. A RIS is a system of orders. An RCM is a system of claims. An AOS is a system of work.
A relationship records who sends. An order records what should happen. A claim records what did. An AOS makes it happen.
Information systems record the work. An AOS does the work.
Every incumbent category in imaging is organized around storing and displaying operational information so a person can act on it. An AOS is organized around completing the action.
Two boundaries make the definition precise. An AOS begins before the order exists: the referring office is where imaging demand is created and lost, and the system’s agents work there on the imaging center’s behalf, receiving the order in whatever form it arrives, answering the office’s questions, reporting the order’s progress, and returning the report and the images when the exam is read. And an AOS has no role in the clinical middle: it does not acquire, store or interpret images, and it does not write the report. It carries the order to the scanner and carries the result back out.
- Order inReceives the order, in whatever form it arrivesAnswers the office’s questionsReports the order’s progress
- The clinical middleNot an AOSImages acquired, stored and interpretedThe report written
- Result outReturns the report and the imagesCharge, claim and payment
Why a new category exists now
For thirty years, imaging software has been divided into three products along lines that do not exist in the actual work.
A CRM, where there is one, sees the referring physicians. The RIS sees orders, demographics, slot inventory and schedules. The RCM sees claims, payments and balances. All three categories are mature and well understood, and the boundaries between them are treated as facts of nature.
They are not. The lines are artifacts of how the software was sold, and an imaging order crosses all three.
The consequence is that every order can be lost three times, in three different systems, watched by three different teams. It can be lost before it arrives, when a referring office quietly sends elsewhere or a fax is never worked; call that referral leakage. It can be lost before it is scanned, when nobody reaches the patient and the slot goes empty; order leakage. And it can be lost after the exam, when a missing authorization or a wrong insurance surfaces as a denial forty-five days later, when it is expensive and often unrecoverable; revenue leakage. A denied claim is, overwhelmingly, a front-end mistake that the back end discovers. Unused capacity is the same story: an imaging center filling about 65 percent of its slots while patients wait weeks for an appointment is not short of scanners. It is short of orders that make it all the way through.
- Referral leakage: lost before it arrives
- Order leakage: lost before it is scanned
- Revenue leakage: lost before it is paid
Historically the gaps between the three systems were staffed. Booking staff, intake coordinators, authorization specialists and eligibility clerks existed to carry information across boundaries the software would not cross. That was the only available answer, because the work requires judgment, phone calls, portal navigation and the reading of badly handwritten faxes, none of which software could do.
That constraint lifted. Agents can now hold a phone conversation, read a smudged order, navigate a payer portal that has no API, and decide what to do next. Once that is true, the gaps no longer need staffing. They need a system that treats the order as one piece of work from the referring office to the payment, and that system is an AOS.
The five tests
Not every system marketed with agents, automation or AI is an Agentic Operations System. These five criteria separate the category from adjacent ones. A system that fails any of them is doing something else, which may still be useful but is not this.
It performs the work, not the workflow
The measure is completed transactions with no human involvement, not clicks saved or screens consolidated. Workflow software makes a person faster. An AOS completes the task; people handle the exceptions.
Ask: What percentage of prior authorizations complete end to end with no human touch? A vendor who answers in time saved rather than completion rate is describing workflow software.
It spans the whole order, from the referring office to the payment
An AOS treats the order as one process rather than three modules with handoffs between them. The test is not whether referral tracking, booking and billing all appear on the price list. It is whether a problem discovered at one point gets resolved at another, by the same system, without a person carrying it across: a coverage problem found at booking is fixed before the patient arrives, and a report signed by the radiologist reaches the referring office without anyone faxing it.
Ask: When a coverage problem turns up at booking, what resolves it, and does the patient arrive with it already fixed? When the report is signed, how does the referring office get it?
It operates systems it does not own
Most of imaging operations runs through systems the imaging center does not control and cannot change: payer portals, clearinghouses, referring-physician fax lines, a legacy RIS with no modern interface. An AOS has to work through whatever interface exists, whether that is an API, HL7, FHIR, or driving a browser the way a person would. A system that requires every counterparty to be integrated is not an operations system. It is an integration project.
Ask: What happens when the payer has no API? “We’d need them to build one” is a disqualifying answer.
It scales with volume, not headcount
An AOS decouples operational capacity from staffing. Doubling volume in a week should not require hiring, and should not require notice. In this category, cost per transaction falls as volume rises. A staffed operation works the other way: fixed cost, and hiring ahead of demand.
Ask: What happens to our operating cost if volume doubles next month? A seat-based license answers this question badly.
It escalates by exception, not by default
Humans do not disappear from an AOS. They move. Routine work completes on its own, and genuine exceptions go to a person with the full context attached: ambiguity, clinical judgment, an angry patient, a payer behaving unusually. Maturity shows up as a falling exception rate, and a vendor should be able to state theirs.
Ask: What is your exception rate by flow, and how has it moved over the last twelve months?
AOS compared to adjacent categories
| CRM | RIS | RCM | AOS | |
|---|---|---|---|---|
| Answers the question | Who refers to us, and how much? | What is booked and who is the patient? | What was billed and what was paid? | Is the work done? |
| Unit of value | A relationship | An order | A claim | A completed transaction |
| What staff do | Log visits and calls | Enter and retrieve | Submit and chase | Handle exceptions |
| Scope | Before the order | The order and the schedule | After the exam | The whole order, referring office to payment |
| When volume doubles | Hire liaisons | Hire | Hire | Cost per transaction falls |
| Primary constraint | Staff capacity | Staff capacity | Staff capacity | Exception rate |
The important rows are the last two. How many people you can hire and train is what limits all three incumbent categories. Consolidating them into a single system is a real improvement (fewer products, cleaner data, better visibility) but it does not touch that limit. Consolidation makes a staffed operation more efficient. An AOS changes what the staff are for.
The AOS maturity model
Autonomy is a property of a workflow rather than of an organization. A single imaging center will usually run different processes in different modes at once. There are five modes. The first two involve no artificial intelligence; the last three do.
No AI
Manual
A person performs the task in a system of record. The software’s job is to store the result.
A person does the workStoredAutomated
The software performs the task by fixed rules, where nothing needs deciding: a fax arrives and is received, a signed report goes out. It takes the person out of work that never needed judgment, and it involves no artificial intelligence.
Fixed rules do the workDone
Uses AI
Assisted
The software pre-fills, suggests or flags. A person still performs and owns every transaction.
A person does the workSoftware suggestsSupervised
An agent performs the transaction end to end and a person reviews before release. Throughput rises sharply, but headcount is still coupled to volume.
An agent does the workA person reviewsAgentic
An agent performs and completes the transaction. No person is in the routine path. People are in the exception path, and increasingly in the design path, building and tuning and supervising the operation rather than executing it.
An agent does the workExceptions go to a person
The modes are not a ranking. Automated work is rule-based and complete, and it does not become “more autonomous” by adding AI. The last three form the progression that matters where judgment is required: assisted, then supervised, then agentic. Most imaging organizations run their judgment work manual or assisted today, whatever their software is marketed as. Agentic also arrives workflow by workflow: usually document processing first, then eligibility, then booking, with prior authorization last because payer behavior is the least predictable input in the whole process. A vendor claiming every workflow runs agentic on day one is describing an ambition rather than a deployment.
Beside the ladder sits the other way a task leaves the center’s desk: the patient or the referring office does it themselves, through a doorway such as a portal, a link or an upload. A doorway has no mode, because the patient or the office did the task, not the software.
What an AOS is not
It is not a PACS or a reading tool. An AOS does not acquire, store or interpret images, and it does not write the report. Its work pauses when the patient enters the scanner and resumes when the report is signed: delivering that report, and the images, to the referring provider.
It is not diagnostic AI. Most AI investment in radiology addresses detection and interpretation, and an AOS addresses none of it. The two are complements and do not compete for the same problem, though they frequently compete for the same budget.
It is not a marketing tool, though it works in the referring office. An AOS does the clerical half of the referring relationship (receiving orders, answering status questions, returning results) so that the people who own the relationship can spend their time on it.
It is not a RIS, a CRM or an RCM. CRM, RIS, PACS and RCM are information systems: as the name says, they record the work. It can run on top of the systems you already run.
It is not RPA. Robotic process automation replays a recorded sequence of steps and breaks when the screen changes. An agent interprets intent and adapts. The difference becomes obvious the first time a payer redesigns its portal.
It is not a chatbot. A conversational interface may be one surface of an AOS, but a system that answers questions without completing transactions fails test one.
It is not an “agentic operating system.” The phrase is used in other industries for software that runs other software. An Agentic Operations System runs an operation.
How to evaluate an AOS
Nine questions worth putting to any vendor in this category, including in an RFP where the answers can be held to.
What percentage of each flow completes with no human touch today, in production, at a customer of our size?
Ask for the number by flow rather than in aggregate.
What is your exception rate, and what is the trend?
A vendor who cannot state it is not measuring the thing that governs the value.
What happens when a payer or referrer has no API?
The answer reveals whether the system can operate in the real world or only in an integrated one.
How are we charged: per seat, per transaction, or per outcome?
Seat-based pricing in this category signals a vendor who still assumes staff.
What is the time to first fully autonomous workflow, as opposed to time to go-live?
These are very different dates.
Who is accountable when the system gets it wrong, and what is the audit trail?
Ask to see a real one.
What happens to our data, and can we leave with it in a usable form?
Is the vendor also an operator of imaging centers?
If so, understand clearly what governance separates your operational data from a competitor’s commercial interest. This is a reasonable question, and a straightforward vendor will have a straightforward answer.
What does a referring office experience?
How does it send an order, how does it learn where the order stands, and how soon after the read does it have the report and the images? A system that starts at intake has no answer to this question.
Origin of the term
The term Agentic Operations System was proposed by Alpha Nodus in 2026 to name a category that had begun to exist in practice before it had a name. We have an obvious commercial interest in the category, and we would rather state that plainly than pretend to neutrality.
We also think the category is bigger than any one vendor, and we would consider it a good outcome if others adopted the term, competitors included. Categories are more useful than acronyms. If the five tests above turn out to be the right tests, they will be applied to us too, and they should be.
This definition is versioned and will be updated as the category develops. Disagreements, corrections and additions are welcome: aos@alphanodus.com.