What can AI agents actually do in radiology?

Last week we rang 100 Australian radiology practices during business hours and timed every call. Eight reached a person directly. Ninety-one landed in a recorded menu first, and the median wait to a human was 1 minute 30 seconds. We rang the same hundred on a Sunday evening. Not one reached a person.

One clarification before the list, because in radiology “AI” usually means something else entirely.

This is not about AI reading scans. Nothing below touches image interpretation, reporting turnaround, worklist prioritisation, nodule detection or any diagnostic decision. That is a different technology under a different regulatory regime, and it is a different post.

What follows stops at the front desk. Phones, web chat, forms, referrals, bookings, reminders, outbound follow-up and the CRM work around all of it: everything a patient touches before a radiographer does. What happens past reception, in the reading room and the report, we will take separately.

So the question worth answering is not whether a practice should automate something. It is what an agent can genuinely take on in the front office, and what it should be nowhere near.

The short version

Three tiers, roughly the order a practice adopts them.

1. Front office and patient access. Omnichannel triage across phone, chat and messaging. Scheduling that cross-references modality rules, contrast, preparation and funding pathway before it offers a time, written straight into the RIS. Intake and preparation chased before the patient arrives.
2. Operational and revenue recovery. Referrals ingested and parsed, with ambiguous documents routed to a person. No-shows mitigated and waitlists used to fill vacated scanner time. Unbooked referrals and abandoned enquiries followed up rather than lost.
3. Referrer and network support. Professional calls from GPs and specialists recognised and routed separately from patients. Capacity balanced across sites, so a patient is pointed at the clinic that can actually see them sooner.

How much of this is actually AI

Worth separating, because “AI agent” gets used as though the model does all of it. It does not, and the distinction matters when you are comparing vendors.

The AI does the parts that involve language and judgement about language. Working out what a caller wants from how they said it. Holding a conversation through interruption, an accent, a change of mind. Reading a referral that arrives as a phone photo and pulling the referrer, the modality and the body region off it. Summarising a call for the person it is about to be transferred to. Categorising thousands of conversations so patterns become visible.

Everything else is a purpose-built platform, and it is most of this page. Telephony provisioned at double your current line capacity, or as many as you need, transferring with context. The rule engine that knows a cardiac MRI is not a wrist x-ray, which studies need contrast, what preparation each one carries, which site runs it on which day, and what the funding pathway changes. Write-back into the RIS. Per-site configuration and roster rules. Waitlist logic. Quiet hours and contact limits on outbound. Task queues, escalation routing, the audit trail, and per-site reporting.

None of that is a language model. It is ordinary, unglamorous software that has to be built against how radiology actually works, and it is the reason a general-purpose voice agent can sound excellent on a demo call and still be unable to book anything. The model is the interface. The platform is the product.

A useful question for any vendor: which half are you showing me?

What it must never touch

The intro promised this half, so it comes before the inventory rather than after it. Four categories of request are hard stops. Not handled carefully, not escalated when the model is unsure: the agent is not permitted to answer them at all, and the platform is built so that it cannot.

Acute symptoms. Chest pain, sudden breathlessness, facial droop or one-sided weakness, a sudden severe headache, heavy bleeding, a significant fall. The agent stops working the booking, tells the caller to hang up and ring the emergency number for the clinic’s country (000 in Australia), stops talking, and leaves a flagged record for the practice. That is the one response, at every site. It does not grade severity, ask further questions to decide how urgent the symptom is, or conclude that it is probably nothing. Recognising that a caller has described a red flag is a routing decision. Deciding what the red flag means is a clinical one, and the only safe route is off the platform.

Protocol decisions. Whether a study needs contrast, which phases it runs, whether the history changes the sequence, whether a suspected pathology warrants a different examination from the one requested. The agent holds the practice’s rules about what a given study normally involves, and it captures the history in the patient’s words. It never assigns a protocol, never varies one, and never resolves a disagreement between the referral and the rule by choosing. Where those two disagree, a person decides.

Medication, screening and preparation exceptions. The agent delivers the practice’s approved preparation sheet as written. The moment a patient asks something the sheet does not answer, it stops and hands over: insulin before a fasting scan, metformin around contrast, anticoagulants before a biopsy, an implant, a pacemaker, pregnancy, breastfeeding, kidney function. The preparation sheet is a document the agent delivers. It is not a document the agent interprets.

Results and interpretation. Any question about findings, a report, an incidental note or what a scan means goes to the referring practitioner or the clinical records desk. One line is worth drawing carefully, because two questions sound alike. “Has my report gone to my GP yet?” is administrative, and the agent can answer it. “What did it say?” is not, and no amount of careful wording makes it so.

Two things make a hard stop real rather than a paragraph in a policy document.

The first is that it is enforced in the platform rather than requested in a prompt. The booking path is closed for those intents, so there is no phrasing, no persistence and no unusual way of asking that talks the agent into a different answer. A guardrail that lives only in the model’s instructions holds until somebody words the question differently.

The second is that every stop is logged and lands somewhere with an owner. An escalation into an unattended queue is not an escalation. If a practice cannot say who picks up a flagged call and within what time, and what happens to one raised at 9pm on a Saturday, the boundary has been described rather than built.

And one thing the agent never does at any hour: pretend. It says it is an AI assistant, says what it can do, and offers a person.

The inventory

Everything below is administrative and none of it is clinical. It is not an exhaustive list. It is what we see practices ask for and what the platform is built to do, written out so the shape of the work is visible. Some of it will not apply to your practice, and you will almost certainly have jobs on your desk that are not on this page.

Seventeen categories, grouped into three pillars. If you only want the argument rather than the detail, skip to the bigger opportunity, and if you only want the boundary, it is directly above.

Pillar one: patient access and scheduling

Everything between a patient deciding they need a scan and arriving prepared for the right one, at the right site, with a valid referral.

Patient access

• Answer patient phone calls 24/7
• Handle multiple calls simultaneously
• Respond to website chat and messaging enquiries
• Identify returning patients
• Understand why a patient is contacting the practice
• Answer common questions about scans, preparation, locations, opening hours and parking
• Provide information in multiple languages
• Send information by SMS or email during the conversation
• Escalate complex or urgent enquiries to staff

Appointment booking

• Identify the examination requested on the referral and apply the clinic’s booking rules
• Collect the information required to make a booking
• Find suitable appointments across multiple clinics
• Offer alternative locations when the preferred clinic is full
• Offer alternative appointment times
• Book directly into the RIS or booking system
• Reschedule appointments
• Cancel appointments
• Manage appointment waitlists
• Fill appointments created by cancellations
• Prioritise available scanner capacity based on predefined business rules

This is a short list and the hardest thing on the page to build, because before an agent can offer a single time it has to resolve a stack of things that have nothing to do with holding a conversation.

Get one of those wrong and you have not saved the practice a phone call. You have created a second one, plus an annoyed patient. It is why we built the rule engine before the conversation rather than the other way round.

Patient preparation

• Send examination preparation instructions
• Explain fasting, hydration or arrival requirements
• Send appointment reminders
• Confirm whether preparation instructions have been understood
• Remind patients to bring referrals or previous imaging
• Send location and parking information
• Provide links to forms that need to be completed
• Follow up incomplete forms before the appointment

There is a category of loss here most practices never separate out. A patient who arrives without fasting, without the required hydration, or without disclosing an implant is recorded as an arrival and treated as a failure. They consumed the slot, the staff time and the goodwill, and they still have to come back. Counting those inside ordinary no-shows hides the one cause that is entirely fixable with better instructions at a better moment.

Forms and patient intake

• Collect patient details conversationally
• Pre-fill administrative forms
• Collect administrative consent, such as SMS and email contact preferences
• Validate required fields before submission
• Identify missing information
• Route completed forms to the correct workflow
• Push collected information into existing systems

Pillar two: referral and revenue recovery

Inbound demand makes noise. This pillar is the demand that does not: the referral that never became a booking, the enquiry that stopped halfway, the slot that emptied at 4pm.

Referral management

• Receive referrals from patients and referrers
• Extract information from uploaded referral documents
• Identify the requested modality and examination
• Detect missing referral information
• Ask patients or referrers for missing information
• Route referrals to the appropriate team
• Match referrals with existing patient records
• Create administrative workflows around incoming referrals
• Follow up referrals that have not yet resulted in a booking
• Identify referrals that may otherwise be lost

One behaviour matters more than the rest here. When a document read is poor, the referral goes to a staff queue with the original attached. It is never guessed at. An ambiguous referral is a clinical risk rather than a parsing problem, and a system that resolves ambiguity by picking the most likely option is the one that eventually books somebody for the wrong study.

No-shows and cancellations

• Send automated appointment confirmations
• Ask patients to confirm attendance
• Identify bookings with a raised risk of cancellation, from administrative patterns
• Automatically offer cancelled appointments to patients on a waitlist
• Contact patients who have missed appointments
• Make it easy for patients to rebook
• Analyse patterns associated with no-shows

A no-show costs twice: the study that did not happen, and the capacity that evaporated. Recovery has to pay back both, which is why the waitlist item above sits next to the rebooking one rather than in a different section.

Outbound patient communication

• Call patients automatically
• Send personalised SMS messages
• Send email
• Conduct appointment reminder campaigns
• Follow up outstanding referrals
• Contact patients about changed appointments
• Notify patients when earlier appointments become available
• Conduct administrative post-appointment follow-up

Outbound is where most of the recoverable money sits, because nothing rings when a patient fails to arrive. Inbound demands attention by making noise. Outbound sits in a queue reception reaches only after the counter is clear and the phones are quiet, which is to say rarely.

Revenue and referral leakage

• Identify calls that never became bookings
• Identify referrals that were received but never booked
• Identify patients who abandoned the booking process
• Detect missed enquiries
• Quantify potential revenue leakage
• Follow up recoverable bookings
• Attribute recovered bookings to the AI workflow
• Show the financial impact of operational bottlenecks

The quietest leak in the building is the referral that never became a booking. There was never an appointment to miss, so it appears in no no-show report and no revenue report. It looks like nothing at all, which is precisely why it persists.

Marketing and patient acquisition

• Track which advertising generated an enquiry
• Connect marketing activity to actual bookings
• Identify which campaigns produce completed scans rather than just clicks
• Automatically follow up unconverted enquiries
• Adjust marketing activity around available capacity
• Promote specific modalities or locations with unused capacity
• Measure cost per booked patient
• Measure cost per completed scan
• Identify where prospective patients are dropping out

Most advertising reporting stops at the click, which is the wrong end of the funnel to measure. A booked scan and a completed scan are different numbers, and the gap between them is usually where the budget is actually going.

Pillar three: network and operational intelligence

The pillar that earns most across a group rather than a single site, and the one practices usually reach only once the first two are running.

Referrer support

• Answer calls from GP and specialist practices
• Identify the referring practice
• Route professional calls differently from patient enquiries
• Provide administrative information about referrals
• Confirm whether referrals have been received
• Request missing referral information
• Direct referrers to the appropriate radiology team
• Assist with practitioner-specific administrative workflows

This section is routinely missed. A call from a GP is not a patient call: the intent differs, the urgency differs, and the person on the other end has a relationship with the practice that a hold queue damages faster than any patient interaction. Treating referrers and patients as one inbound stream is a decision most practices have made by accident rather than on purpose.

Contact centre operations

• Triage every incoming call before it reaches reception
• Transfer calls to the correct department
• Determine whether a call actually requires a person
• Handle routine enquiries without staff involvement
• Capture information while a patient is waiting
• Return a caller to the AI agent if a transferred call is unanswered
• Manage after-hours calls
• Handle overflow during peak periods
• Create callbacks instead of leaving patients on hold
• Provide staff with a summary before transferring a conversation

Workflow automation

• Trigger workflows based on what a patient says or does
• Assign tasks to staff
• Route enquiries between departments
• Escalate unresolved cases
• Follow up outstanding tasks automatically
• Coordinate workflows across phone, messages, chat, web, forms and email
• Maintain a history of patient interactions
• Detect when a patient journey has stalled
• Automatically initiate the next appropriate administrative action

Capacity management

• Identify clinics with available capacity
• Identify underutilised modalities
• Direct patients towards available appointment capacity
• Promote earlier appointments at alternative locations
• Match patient demand with scanner availability
• Help fill unused CT, MRI, ultrasound or other modality capacity
• Identify recurring periods of underutilisation
• Support network-wide capacity balancing

Note what every item in this section has in common: it moves demand towards capacity that already exists. None of it creates capacity. An agent books into availability rather than manufacturing it, so answering more calls can surface a capacity problem the busy tone was hiding.

Operational intelligence

• Analyse thousands of patient conversations
• Identify the most common reasons patients call
• Measure call demand by clinic, hour and day
• Measure hold and transfer rates
• Identify recurring patient complaints
• Identify booking friction
• Measure enquiry-to-booking conversion
• Measure referral-to-booking conversion
• Measure booking-to-completed-scan conversion
• Identify where patients abandon the journey
• Compare performance between clinics
• Surface operational problems that might otherwise remain invisible

Calls answered tells you the system ran. The three conversion measures above tell you whether it worked, and reported per site they tell you which centre is leaking capacity.

Staff assistance

• Summarise patient conversations
• Provide reception staff with relevant information before a transfer
• Suggest the next administrative action
• Search internal policies and procedures
• Help staff answer common questions
• Draft patient messages
• Reduce repetitive data entry
• Automatically create administrative notes
• Handle repetitive enquiries so staff can concentrate on patients requiring human assistance

After-hours operations

• Answer every after-hours call
• Provide preparation and appointment information
• Capture new booking requests
• Send online booking links
• Collect referrals
• Create follow-up tasks for the next business day
• Distinguish urgent from routine enquiries using predefined rules
• Escalate appropriate calls to an on-call pathway

This is the section our own research speaks to most directly. Of 100 practices rung on a Sunday evening, none reached a person, eleven were answered by an AI agent, and the remaining eighty-nine returned a closed message.

Enterprise radiology networks

Agents get disproportionately more useful across a large group, because they can:

• Operate across dozens or hundreds of clinics
• Understand different rules for each location
• Search capacity across the entire network
• Route patients between clinics
• Apply modality-specific workflows
• Integrate with existing RIS, telephony, websites and booking systems
• Provide one operational view of patient demand across the group

What connecting to your RIS actually involves

This is where projects of this kind fail, and rarely for an interesting reason. The system of record will not accept a write, or it will accept one nobody can audit, or the interface sits eighteen months down a vendor roadmap.

Your RIS remains the system of record. Triagents keeps the administrative work around it moving. Which means the write matters as much as the conversation, and there are four patterns for it.

The fourth pattern matters more than vendors admit. Plenty of radiology sites run a version with no write path worth using, or an IT policy that will not open one until the platform has been live a quarter. A product that only works with the first pattern spends six months waiting on somebody else’s release notes. A product that treats the fourth as a real outcome starts recovering referrals in week two and earns the write later.

Two positions worth stating plainly, because both come up in the second meeting. Direct database access is a last resort we avoid: a write that bypasses the RIS’s own validation also bypasses its audit trail, and the practice inherits both problems. And nothing here touches PACS or the reporting worklist. The agent’s write ends at the appointment and the administrative record, which is the same boundary the rest of this page keeps.

Where integrations stand today: practice management write-back shipped first for Best Practice and Genie, in the TriageWorkflow rostering release. The connector list moves month to month, so ask for the current one rather than trusting a web page, ours included. On the radiology side, Karisma, Comrad and Voyager are scoped per site rather than assumed, because two practices running the same RIS version rarely share the same appointment types, room rules or funding configuration. The gap between those two facts is where a demonstration stops matching a practice.

Questions worth putting to any vendor, including us:

• Which of the four patterns do you support at my RIS version, in production, today?
• Does the booking appear in the RIS under a distinct service account, or under a staff login?
• What does the patient hear when the write fails? A confirmation you cannot keep is worse than a callback.
• Who owns the interface when the RIS is upgraded, and what is tested before it goes back into service?
• What runs on day one if the write path is not open yet?

Where patient information goes

A front office handles the most sensitive administrative data in the building: voice recordings, transcripts, and referral letters carrying clinical history. So the governance questions are not a procurement formality, and they have specific answers.

Residency. Hosting is in the country where the clinic operates, and the region is fixed when the clinic is provisioned. Australian clinics run on Australian infrastructure, United States clinics on US infrastructure, UK clinics on UK infrastructure. Logs, backups, recordings and transcripts are stored in the same country, and the region cannot move without a documented migration request from the customer. Vendor processors, telephony carriers included, are bound by contract to the same requirement.

Encryption. TLS 1.2 or above for network traffic. AES-256 at rest.

Model training. Customer data and identifiable personal or health information are not used by Triagents, or by any third party including our AI service providers, to develop, improve or train the software or any AI model without the customer’s prior written consent. Aggregated, de-identified operational data, which cannot reasonably identify the practice or any individual and contains no personal or health information, is used to provide the analytics, reporting and benchmarking the platform exists to deliver, and to operate, secure and improve it. That is the whole exception, and it is not permission to train models on patient data. Ask any vendor for the equivalent in writing rather than on a web page, and ask it twice about the inference path: a recording filed in Sydney and processed by a model endpoint in Virginia is not onshore, whatever the storage diagram says.

Retention. Recordings and transcripts are retained per clinic configuration, typically 90 days, then deleted. Operational logs run 12 months and are then de-identified. The configuration belongs to the practice, not to us.

Access and audit. Role-based access, MFA on every administrative account, and an audit log of every access and every change. Each action the agent takes carries its own identity, which is why the service-account question sits in the integration list above. A practice needs to be able to ask what the agent did and get an answer that separates it from what a receptionist did.

Consent and identification. The agent identifies itself as an AI assistant at the start of a conversation, says what it can do, and offers a person. Recording notification is configured per site, because the obligation differs by jurisdiction.

When something goes wrong. A notifiable breach means notification to the affected clinic within 24 hours, and the obligations of Part IIIC of the Privacy Act 1988 (Cth) towards affected individuals and the OAIC.

The frameworks are the Privacy Act 1988 (Cth), including the Australian Privacy Principles and the Notifiable Data Breaches scheme; HIPAA for United States customers, with a Business Associate Agreement signed before any covered data is processed; and UK or EU GDPR with standard contractual clauses where applicable. Detail sits on the security and privacy pages, and residency specifically in sovereignty by design.

One limit, stated because vendors rarely state it. Residency addresses cross-border disclosure and nothing else. It is not a security control and it is not compliance. Access control, retention, encryption and audit each have to hold on their own.

What to measure, and what we can prove

This is the point in a page like this where a vendor prints a percentage. A recovery rate, a no-show reduction, a headcount saved. We are not going to, and the reason is worth more to you than the number would be.

An uplift figure from a vendor’s other customers is a claim about their referrer base, their appointment types, their configuration and their staffing. It travels badly. Anyone quoting a referral recovery rate before they have seen your unbooked referral queue has worked backwards from a sales target, and ours would be no better. What we can show is what we have measured ourselves.

Everything else on this page should be measured in your practice, against your own starting point. Six numbers are enough, and producing them is itself diagnostic.

1. Unbooked referrals. Referrals received minus referrals booked over 90 days, counted per referrer. Most practices cannot produce this from one query, and that difficulty is the finding.
2. Abandonment in the morning peak. Calls abandoned between 08:30 and 10:30, from the telephony data rather than from impression.
3. Prep-failure arrivals. Patients turned away for fasting, hydration or an undisclosed implant, separated out of the general no-show number. It is the one loss that better instructions at a better moment reliably fix.
4. Referral-to-booking and booking-to-completed-scan conversion, per site. Two numbers rather than one, because the gap between them is where the money goes.
5. Vacated capacity refilled within 24 hours. Cancellations are inevitable. A scanner that stays empty after one is a choice.
6. Desk hours spent on enquiries that never needed a person. Parking, preparation, directions, simple reschedules. Sample a week rather than estimating.

Measure the same six again after 90 days on one modality, and the difference is your number. It is also the only one that will survive a board meeting.

The bigger opportunity

The real opportunity is not an AI receptionist.

Read the inventory again as one thing rather than three pillars and seventeen lists. It describes an administrative patient operations layer sitting across phone, messages, chat, web, forms, email, referrals, marketing, booking and the RIS. It identifies what needs to happen, takes the next action, moves information between systems, and keeps the patient journey progressing from first enquiry to completed scan.

That is the point at which AI stops answering questions and starts running parts of radiology operations.

A four-step readiness check

Nothing here argues for an all-at-once rollout. The practices that get value quickest tend to work in this order.

1. Ring your own line. At 8:05 on a Monday, from a mobile, on mobile data, and time how long it takes to reach a person. Then ring it on Sunday evening. Whatever those two numbers are, they are your baseline, and everything above is only worth discussing against them.

2. Count what is already lost. The unbooked referral queue over 90 days, and after-hours call volume from the telephony records. If nobody in the practice can produce the first number, that is the project, because no agent fixes a queue nobody can see.

3. Establish which integration pattern your RIS supports. Ask the vendor in writing, before any product demonstration: documented appointment API, HL7 through the interface engine, FHIR gateway, or none of the three. That one answer shapes scope, timeline and price more than any feature list will.

4. Pilot on one rule-bound, high-volume modality. General x-ray, bone densitometry or routine ultrasound. The rules are simple, the volume is high enough to read a result within weeks, and a mistake is recoverable. MRI, contrast studies and interventional scheduling come last, not first.

Two things to settle before go-live, whichever modality you start on: who owns the escalation queue during the day, and what happens to a flagged call at 9pm on a Saturday. Both answers are administrative. Both are the reason a patient or a referrer trusts the front desk afterwards.