AIGuardianThe accountability layer for clinical AI
AIGuardian monitors the whole imaging chain — the scanners, the PACS and RIS interfaces, every AI you run, and the read itself. Not just whether the model is right, but whether the chain around it is still working.
Your imaging chain, monitored end to end
Studies arrive from the scanners into PACS and RIS. From there they are sent out to every AI you run, and every AI result returns to PACS — which is where the radiologist reads it. We watch each link and each interface between them, and we keep a live performance trace on every tool. When one leaves its agreed bounds, it lights up here.
Scroll the diagram sideways to see the whole chain.
The concept
A correct model on a broken interface is a broken deployment
Most monitoring in this market watches the algorithm. That is the smaller half of the problem. Between the scanner and the signed report there are half a dozen handovers, each one carried by a protocol that can fail without raising an error anywhere.
Studies that never arrive
A modality worklist entry, a changed AE title, a study description that no longer matches a routing rule. The AI is running perfectly and simply never sees the case.
Results that never land
A structured report or overlay comes back but is filed against the wrong series, lands in a tab nobody opens, or is silently rejected by the archive. The AI did its job. Nobody saw it.
Handovers that quietly degrade
Priors that stop being fetched, headers that lose a field after a PACS upgrade, results that arrive after the study has already been read. Everything reports success. The clinical value is gone.
This is why we monitor the chain rather than the model. If the AI is right and the interface is wrong, the patient gets the same outcome as if the AI were wrong — and nobody in the department has the evidence to tell the two apart.
The chain, not the model
Every link between the scanner and the report can fail on its own
A model can be performing exactly as validated while the study never reaches it, the result never lands back in PACS, or the protocol it was tuned for has quietly been changed on the scanner. We watch the links, not only the prediction.
Every study that reaches a tool, and every result that comes back, is accounted for.
Where it breaks
Clinical AI rarely fails as a wrong answer
It fails somewhere along the chain, quietly, in a way that produces no error message. Pick a point.
What goes wrong
The environment moves under the tool
A scanner is replaced. A protocol is adjusted. A service starts seeing a different population. The model has not changed, but what it is being asked to judge has.
What you get
Change is attributed, not guessed at
You are told what shifted in the inputs and when, so a drop in performance is traced to its cause instead of being blamed on the algorithm or on the radiologists.
What goes wrong
The protocols fail silently
A routing rule stops matching after a PACS upgrade. A structured report is rejected by the archive. Results arrive after the read is signed. Every system reports success and the case was never covered.
What you get
Coverage you can state as a fact
We watch the DICOM and HL7 handovers themselves, not only the AI output — so you can say which studies were assessed, which were not, and exactly when that changed.
What goes wrong
Updates arrive without a conversation
Manufacturers ship new versions. The tool your committee approved and the tool running this morning are not always the same one, and nobody downstream is told what moved.
What you get
Continuity across versions
A live trace on every tool against bounds agreed at the start, with an alert the moment behaviour leaves them — which is also the monitoring a change control plan requires.
What goes wrong
Trust shifts in both directions
Clinicians quietly stop using a tool they no longer believe, so the investment is lost without a decision being taken. Or reliance deepens past what the tool warrants.
What you get
Evidence that oversight is real
Visibility of how the department actually uses each tool over time, which is what a regulator means by demonstrating human oversight rather than asserting it.
What goes wrong
Nobody was looking
Most departments find a problem through a complaint, an incident or a renewal meeting. By then it has been true for months, and there is no record of when it started.
What you get
A record that already exists
A continuous, timestamped account of every AI you run, ready for your board, an auditor, a notified body or an insurer.
Controlled deployment
The same measurements that prove a tool works are what get it into routine use faster
Most of the time between signing for an imaging AI and actually relying on it is spent arguing about whether it works. Measuring the deployment against your own data replaces the argument with evidence — what the tool is really reporting, how that should reach the people reading it, and where the thresholds belong.
Grounded in your data
Decisions from your department
Every judgement about the deployment is made against measurements from the population and the systems you actually run, not the validation set the tool was sold on.
Controlled change
Thresholds move in steps
Each adjustment is measured before and after, against live traffic. A change that does not help is visible immediately, and it is reversible.
Faster to routine use
Evidence shortens the argument
Departments reach the point of trusting a tool sooner, and can say why. Adoption stops depending on whoever argues hardest in the meeting.
No extra work
The record writes itself
The measurements that guide the deployment are the same ones a board, an auditor or a notified body will ask for later. Nothing is assembled twice.
Who it is for
Peace of mind for everyone who now depends on this AI
Radiology is where clinical AI actually landed. Of the 1,524 AI-enabled devices the FDA has authorised, 1,164 are radiology devices — 76% — and the people carrying the risk are no longer only the hospital.
Health systems and hospitals
You signed for it
Deployment is nearly universal and confidence is not. You need a standing answer to whether each tool is still working, and a record you can hand to a board or a regulator.
In a 2024 survey of 43 US health systems, imaging was the most widely deployed clinical AI use case, with 90% reporting at least partial deployment — while the same study found reported success with diagnostic use cases was limited.
Radiologists and reading groups
You carry the read
You are asked to rely on a tool you cannot audit. Independent measurement tells you where it is strong, where it is not, and when something has changed.
In the European Society of Radiology's 2024 member survey, 274 of 572 respondents (48%) said they currently use AI, and a further 25% planned to.
AI manufacturers
You need field evidence
Multi-site, real-world performance data you cannot collect from outside a customer's firewall — for post-market surveillance, for renewals, and for a change control plan you can actually execute.
The FDA authorised 92 AI-enabled devices in the first quarter of 2026 alone, 75% of them in radiology.
Imaging OEMs and platforms
You ship AI inside the box
When AI arrives embedded in a scanner or a PACS, the obligation to show it still performs travels with it. An independent layer gives your customers that without you grading your own work.
The five largest holders of radiology AI authorisations are equipment manufacturers — GE HealthCare, Siemens Healthineers, Philips, Canon and United Imaging — ahead of every AI-native vendor.
Sources: FDA AI-Enabled Medical Device List, authorisations through March 2026. Poon et al., JAMIA 2025, cross-sectional survey of 43 US health systems (Scottsdale Institute members, fielded autumn 2024). Zanardo et al., Insights into Imaging 2024, EuroAIM/EuSoMII survey of ESR members — 572 responses from 28,000 members contacted, so respondents are likely to be more AI-engaged than the profession as a whole. Device counts change with each FDA update.
The people carrying the risk
Behind every alert there is someone who has to act on it
The radiologist signing the read, the physicist who validated the tool, the clinical safety officer who signed it off. Each needs a different answer from the same evidence, and each needs it before a problem becomes an incident.
Alerts are written for the person who receives them, not for a dashboard.
Change control
We make a Predetermined Change Control Plan executable
A cleared model is frozen at the version that was approved. The FDA's final PCCP guidance, issued in August 2025, is the way out: a manufacturer can pre-specify what it may change, how it will validate it, and inside what performance bounds — then ship improvements without refiling. The problem is that a plan on paper is not a plan you can run.
Performance bounds
Numeric limits agreed across the many signals we measure, not one headline accuracy figure — measured where the device actually runs.
A validation method
A defensible, repeatable way of establishing what happened in each case, applied identically across every site we monitor.
Working monitoring
Not an intention to monitor. A pipeline already measuring cases inside customer sites, in production, today.
A stop rule that fires
An agreed threshold at which an update is rolled back — and a mechanism that will actually detect it and raise the alert.
For the manufacturer this turns a frozen model into one that can improve under supervision. For the hospital it means the version running this morning is still the version that was assessed. Both sides get the same thing from it: the ability to change something without losing control of it.
The obligation
Approval covers the moment. The duty covers the years after.
Every framework governing clinical AI has arrived at the same requirement: know how the tool performs where it is actually used, and be able to show it.
- In forceEU Medical Device Regulation
Post-market surveillance and real-world clinical follow-up, for the life of the device.
- Aug 2025FDA PCCP final guidance
Marketing submission recommendations for a Predetermined Change Control Plan for AI-enabled device software functions. Pre-authorised change, against declared bounds and real monitoring.
- May 2026ACR-SIIM practice parameter
The first professional standard for imaging AI: selection, pre-deployment evaluation, ongoing performance monitoring and stop rules, with facility recognition for sites that implement it.
- Dec 2027EU AI Act, stand-alone high-risk
Deferred from August 2026 by the Digital Omnibus, which moved the dates and left the duties untouched.
- Aug 2028EU AI Act, device-embedded AI
Where imaging AI sits. Oversight, monitoring of operation, logging and escalation fall on the organisation running the system.
Proof
The contract that nearly was not signed. Then first reader across the network.
A fracture-detection AI in a live urgent-care network was not earning the confidence of the clinicians asked to rely on it, and the contract was close to not being signed. Measuring the deployment against its own data showed what the results actually meant and how they had to be presented to radiologists, and where the thresholds needed to move. The changes were made under control and verified against live data — continuously, over the years since rather than in a single retest.
Too many positives were not holding up on review. Confidence was low and the contract was close to not being signed.
The results were not reaching radiologists in a form that reflected what the tool was actually reporting.
Presentation was corrected, then thresholds were adjusted in controlled steps, each one measured against live data.
The same tool was promoted to first reader for every radiologist and front-line clinician.
Nothing about the algorithm was replaced. A deployment measured against its own data can be improved quickly, in controlled steps, with the evidence for every change.
This runs where the work happens
Installed beside the PACS, watching live traffic. No change to how radiologists read, and no case leaves the hospital.
Find out what your AI is actually doing
Tell us which tools you run and on which systems. We will show you what the monitoring record would look like for them, with a radiologist in the room.