Agentic AI as the new infrastructure for interoperability

Interoperability in imaging has long relied on a simple bargain: When systems speak the same language — DICOM, HL7, FHIR — data moves. When they don’t, staff step in: logging into portals, copying values, reconciling IDs, scheduling exams, chasing priors. Standards cover the happy path; people bridge the rest. 

Rishi Nayyar.Agentic AI changes that bargain. Instead of waiting for every vendor to expose the perfect API, purpose-built agents behave like credentialed teammates. They see the screen, understand the form, follow the workflow, and take the next safe step. The effect isn’t a replacement for standards, but an extension — bringing interoperability to the messy edges where real work happens.

AI in imaging is often framed around interpretation. Yet many delays, repeat scans, and missed follow-ups occur before and after the read: when orders can’t be created from outside requisitions; when priors live in a vendor portal with no reliable export; when a self-scheduling slot is free but not exposed via the API; when reminders require data from three systems that don’t talk to each other. Today, staff resolve these seams by clicking. That’s coordination, not cognition — and it’s exactly where agentic AI excels. 

How it works

Agentic AI acts on a user’s behalf across software. It signs in with named credentials, reads what’s on screen, executes the workflow as a registrar or technologist would, and documents what it did. Standards remain the first choice, but where they fall short, agents faithfully mimic human workflow, safely and at scale. They handle the deterministic 80% to 90% and hand off the rest. 

DICOM/HL7/FHIR remain the backbone and should be used whenever possible. But when capabilities are missing or implemented idiosyncratically, humans take over. Agents offer a third option: Preserve standards wherever they exist and execute staff workflows wherever they don’t. This creates more resilient, governed operations without brittle interfaces. 

Why this moment is different 

Graphical user interface (GUI) automation isn’t new; reliability is. Modern agents combine language understanding, computer vision, tool use, and policy memory to adapt as interfaces shift, maintain context across multiple systems, and follow institution-authored policies (“if glucose > X and modality = CT with contrast, branch to Y”). Every action generates a complete audit trail — who did what, when, and with which inputs. This is scaffolding for governance, not a shortcut around it. 

Where it works today 

• Requisition → order creation (no interface available). The agent extracts data from faxed or portal-submitted requisitions, validates patient and provider details, applies site-specific orderable mappings, and creates the order directly through the EHR/RIS UI — exactly as a registrar would. Ambiguous cases route to a human review queue.  

• Priors retrieval from specialty PACS (non-DICOM quirks). Many specialty PACS (cardiology, ophthalmology) use nonstandard exports. Agents sign in, locate the study, convert or export using vendor tools, and route it into the enterprise PACS/VNA with the correct accession and medical record number — no vendor-to-vendor API required.  

• Reminders and prep. Appointment reminders depend on scattered data: order metadata from the RIS, contact preferences from the EHR, and modality-specific prep from policy manuals. Agents bring these together to deliver personalized reminders across SMS/email/IVR, then write back confirmations or reschedules by following the clinic’s UI flow. 

• True self-scheduling without exposed availability. Patient-facing agents navigate scheduling applications like staff do, respecting site rules, block types, and prep constraints. If a slot is free, they book it, update the order, and issue instructions, all without needing a bespoke scheduling API. 

• Follow-up closure. Agents monitor finalized reports for follow-up recommendations, reconcile them with existing orders, reach out per policy, and create the downstream order through the UI, escalating clinically complex cases for human judgment. 

Here, the workflow becomes the integration. We’re not bypassing the enterprise; we’re operationalizing it. 

Guardrails 

Agents only matter if they’re safe, auditable, and controllable. That’s not just security posture, it’s an operating model. The guardrails below make agent behavior predictable, reviewable, and reversible. 

• Least-privilege access and role-based access control (RBAC). Use of named, scoped identities — never shared logins. 

• Policies as code. Orderable mappings, scheduling rules, and prep criteria — explicit and version-controlled. 

• Human-in-the-loop for ambiguity. Confidence thresholds route edge cases to staff, with one-click resolution and learning feedback. 

• Environment isolation and change detection. Validating in staging: They watch for UI drift and pause gracefully if uncertainty spikes.  

• Comprehensive auditability. Every field read, decision taken, and record written is captured for compliance review. 
   

With those in place, the interoperability question shifts from “Does Vendor X expose API Y?” to “Can a trained human do this safely today?” If yes, an agent can too — with an audit trail and a pause button. 

Where to modernize first 

Start with high-cost, well-understood manual workflows. If agents can’t improve these without adding risk, they’re not ready: 

• Requisition intake → order creation for high-volume modalities and common payers