Superintelligent Agents for Imaging Centers: Scheduling and Prior Auths
Which scheduling and prior authorization tasks an imaging center can hand to an AI agent, and where the agent must stop before anything reaches a payer.

An AI agent for an imaging center is most useful on the scheduling and prior authorization tasks that are repetitive, documented and easy to reverse: confirming that an order is complete, tracking where each request sits, and drafting the outreach that staff would otherwise write by hand. It should not write clinical justifications, submit requests to payers or move a confirmed exam without a named person approving the step. The practical starting point is one queue, one set of stop points written in advance, and one success criterion the schedulers already recognize.
Key takeaways
- Start with status work: order completeness, request age and reminder drafts are the safest first tasks.
- Keep clinical justification, payer submission and appeals with trained staff every time.
- Write each stop point in advance, including who receives the alert and what the agent logs.
- Limit the first engagement to one queue and one modality so the team can check every output.
- Review permissions and evaluations on a schedule, because payer rules and staff roles both change.
Why scheduling and prior authorization suit an agent first
Imaging scheduling combines three kinds of work. Calendar logistics such as room, modality, contrast timing and patient transport follow rules the center already knows. Eligibility and prior authorization depend on the ordering provider's documentation, the payer's current criteria and a deadline that sits outside the building. The first kind is where an agent adds steady value with limited risk. The second needs human judgment at the moment something is sent.
Prior authorization is also a visibility problem more than a decision problem. A request is submitted, waits, receives a question, and ends in approval, denial or expiry, and the exam date depends on which of those happened. Delays often come from nobody looking at the queue at the right time, not from schedulers being unable to decide. An agent that reads the queue, flags aging items and drafts the next message fits that gap. The same boundary appears in AI for Behavioral Health Clinics: Intake and Scheduling Limits, where the useful question is not what the agent can reach but what it may do with what it reads.
Sort every task into one of three lanes
Before any build, list the tasks your schedulers and authorization staff perform in a typical week, then place each one in a lane. The lane decides who is accountable for the outcome, which matters more than how capable the model is.
- Lane one, the agent acts and logs: confirm reminder text against the booked appointment, flag orders missing a required field, and compile open requests into a daily list sorted by age.
- Lane two, the agent drafts and a person sends: a message asking a patient for missing demographics, a request to the ordering office for a specific missing note, or a proposed new slot when a room or contrast window changes.
- Lane three, a trained person only: clinical justification, any payer submission or appeal, any change to the requested service, and any decision to schedule before authorization is confirmed.
Lane three is not a verdict that the model is weak. It records that a named clinician or authorization specialist owns the outcome and must be able to explain it later. If a task is hard to place, put it in lane two until staff have reviewed enough drafts to move it.
A first engagement plan for one queue
Keep the first engagement narrow enough that a scheduler can check every output. The sequence below is the one Be Human uses to scope a first role, adapted here for an imaging center.
- Step 1: Choose one queue, such as orders awaiting prior authorization for a single modality.
- Step 2: Write the current process as staff actually perform it, including the workarounds that never made it into the procedure manual.
- Step 3: Mark every point where the work touches a payer, a patient commitment or a clinical field, and name the role that approves each one.
- Step 4: List the memory and sources the agent needs, such as the schedule, order status, the center's own payer criteria and approved message templates. Leave out anything it does not need for this queue.
- Step 5: Set one success criterion in operational terms, for example that no request reaches a payer with a required element missing.
- Step 6: Run the agent in shadow mode, comparing its flags with what the team found without it, before any draft reaches a patient.
- Step 7: Review what it can read and write, and confirm that every action leaves a record someone can audit.
Illustrative scenario: a knee MRI order arrives with an authorization request returned for a missing clinical note. The agent flags the order the morning the return is logged, drafts a message to the ordering office naming the missing item, and marks the slot as tentative on the calendar. A scheduler reviews the draft, sends it, and decides whether to hold the slot. The agent does not resubmit anything to the payer.
Where the agent must stop before anything reaches a payer
A stop point is a rule that halts the agent and hands the case to a named person, with the reason recorded. Write the triggers before deployment, not during the first incident. Reasonable triggers for an imaging center include:
- The payer asks for clinical detail or attachments beyond the standard packet.
- The requested service does not match the order as written.
- The authorization is denied or expires, or its date window does not cover the proposed exam.
- Patient financial responsibility or consent is unclear in the record.
- The order was revised after the request was prepared.
When a trigger fires, the agent holds the case, notifies the owner named in the procedure and logs the trigger. It does not retry, reword or work around the block. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a useful reference for documenting these controls as part of ordinary risk management, and OWASP's authorization guidance explains why each action should be checked against explicit permissions rather than the agent's own reading of what it may do.
Payer portals, faxed notes and emails are inputs, and they can contain text written to steer an automated reader. Treat anything received from outside as data to summarize, never as an instruction. The OWASP prompt injection cheat sheet covers the basic defenses, and they matter most in the queues where the agent can send messages. Denials that arrive after submission are a separate workflow, covered in Superintelligence Prep for Medical Billing Firms: Denials and Follow-Up.
How Be Human's managed approach handles this
Be Human designs and operates the agent as a managed role. Its identity, memory, channels, tools and permissions are defined in writing and kept under the center's control. The first engagement sets the one role, the memory it needs, the approval boundaries and the success criterion described above. Managed operations then cover monitoring of queue behavior, recurring evaluations against reviewed cases, permission reviews whenever staff or systems change, integration maintenance, and controlled updates to templates and rules, with incidents handled under the same stop points. The service is remote and available in English and Portuguese.
Measuring the agent and choosing the first task
An agent that worked in its first month can drift when a payer changes criteria or a template is edited. Keep a fixed set of past cases, including clean ones and difficult ones, and score the agent's flags against the decisions staff actually made. The [Anthropic guidance on evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) makes the same point: test behavior on realistic tasks rather than trusting a demonstration. Useful measures for this queue include:
- Requests that reached a payer with a required element missing.
- Exams moved because authorization was not confirmed in time.
- Re-keying that staff still do from the schedule or the order.
- Flags that staff dismissed, reviewed with the team each month.
If you are not sure which of these tasks belongs first at your center, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent submit prior authorization requests to payers on its own? For most imaging centers, it should not, at least at the start. The agent can assemble the packet, check it against the order and the payer's current requirements, and draft the submission for a trained person to review. Submission and any clinical argument stay in lane three until the team has evidence that the drafts are reliable.
Which scheduling tasks are safest to hand over first? Reminder text checked against the booked appointment, detection of orders missing a required field, and a daily list of open requests sorted by age are usually the safest. They are rules-based, easy to verify and easy to reverse. Anything that changes a confirmed exam or contacts a patient about a clinical matter should come later, with a person approving each change.
Does the agent need full access to the electronic health record? No. Start with the minimum the queue requires, such as scheduling fields, order status and the authorization record, and leave the rest out of its memory. Permissions should be written as explicit scopes and reviewed when staff roles or systems change, which is the principle the OWASP authorization guidance describes.
How do we know the agent still works correctly after payer rules change? Rerun the fixed set of past cases after every rule or template change, and compare the agent's flags with staff decisions. Keep the monthly review of dismissed flags, because a rise in dismissals often signals drift before anyone notices a missed deadline. Managed operations include this recurring evaluation, so the check does not depend on someone remembering to run it.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


