TL;DR: HL7 imaging interfaces deliver clean, structured results straight into the EHR, but only from the few imaging centers willing to build and maintain one. AI document ingestion files outside neurology reports from every source, including fax, PDF, and portal downloads, at the cost of a small human exception queue. For most neurology practices, the better answer is both: interfaces for one or two high-volume imaging partners, and AI ingestion for everyone else.
What's the actual difference between an HL7 interface and AI document ingestion?
An HL7 imaging interface is a direct, structured data feed between an imaging provider's system and your EHR. When the radiologist signs a report, the result message (typically an HL7 ORU) flows to your EHR, matched to a patient and order, with discrete fields for study type, date, and report text. Nobody touches it.
AI document ingestion works on whatever arrives. It takes faxes, PDFs, scanned pages, and portal downloads, reads them using AI document understanding, matches the patient, classifies the study, and files the report to the right chart location. It doesn't need the sender to do anything differently.
The simplest way to think about it: an interface changes how the sender sends. Ingestion changes how you receive. That distinction drives almost every trade-off below.
Both approaches solve the same pain in neurology clinics: outside MRI, CT, EEG, and EMG reports that need to be in the chart before the visit, filed where the neurologist expects them.
How much of your outside volume can each approach cover?
Coverage is where the two approaches differ most, and it's the question most practices skip.
Start by listing every source of outside imaging and neurodiagnostic reports over the last 90 days. In a typical neurology practice, that list is long: two or three local hospitals, several independent imaging centers, a few sleep labs, outside neurologists who did EEGs or EMGs, and dozens of one-off facilities where a patient had a scan years ago.
An HL7 interface covers exactly one sender. If your top imaging partner sends 40% of your outside reports, an interface with them covers 40%. A second interface might add another 15%. After that, returns drop off fast. The remaining senders each account for a small share of volume, and none of them will build an interface for a single referring neurology practice.
AI ingestion covers everything that arrives as a document. That includes the long tail, and it includes hospitals that technically have interfaces but still fax reports to practices outside their network.
The data supports how much still moves as documents. ONC's 2023 interoperability data brief found that 43% of hospitals were routinely interoperable across all four exchange domains in 2023, which means most still weren't. On the practice side, a March 2026 MGMA Stat poll found 24% of practices lacked digital fax fully integrated with their EHR.
How do setup cost and timeline compare?
HL7 interfaces are a project per partner. Each one requires coordination between your EHR vendor, the imaging provider's IT team, and often an interface engine. You'll map patient identifiers, agree on message formats, test, and go live. Depending on who's involved, that can take weeks to months, and there are usually fees on one or both ends: EHR vendor interface fees, interface engine costs, and sometimes charges from the sending facility.
The bigger cost is that you don't control the timeline. An imaging center's IT team has its own priorities. Your interface request competes with every other practice's.
AI document ingestion is one project, not one per partner. You connect your fax lines and inboxes, define your filing rules, test on a sample of real documents, and go live. Implementation usually takes a few weeks, and every sender is covered from day one because nobody on the sending side needs to change anything.
The ongoing cost model differs too. Interfaces have maintenance overhead: when either side upgrades systems, interfaces break and need attention. Ingestion tools charge on volume or subscription, and the vendor handles model updates.
Which approach produces better data quality?
On a per-report basis, a well-built HL7 interface wins. The data arrives structured. Patient matching uses identifiers both systems agreed on. Study type comes as a code, not as text a model has to interpret. When an interface works, it's close to error-free.
AI ingestion is very good but not perfect. Modern document AI reads typed imaging reports accurately, and most neurology reports are typed, which helps. But it's still interpreting a document rather than receiving structured data, so there's a confidence threshold. Documents above the threshold get filed automatically. Documents below it, such as ambiguous patient matches, degraded faxes, and unusual formats, go to a human exception queue.
Two honest caveats cut the other way:
- Interfaces fail silently. When an interface breaks after an upgrade, results stop arriving and nobody notices until a neurologist asks where the MRI is. Ingestion failures show up in the exception queue, where someone sees them.
- Interfaces only carry what was ordered through them. If the patient's MRI was ordered by their PCP, the imaging center's interface may route the report to the PCP, not to you. That report reaches your practice as a fax, which brings you back to ingestion.
How does each approach land results in the chart?
Placement matters as much as accuracy. A report filed in the wrong section is almost as bad as no report.
Interface results typically land as discrete results tied to an order. They show up in results views, flowsheets, and review queues. That's ideal when an order exists. For outside imaging your practice didn't order, there's often no order to attach to, and results may land in an unmatched queue that someone has to reconcile.
AI ingestion files documents according to rules you define: brain MRI reports to the imaging section, EEG reports to neurodiagnostics, everything named consistently. Some tools also extract key fields like study date and impression text, so neurologists can scan results without opening each PDF. The result is a chart that's organized by your conventions, not the sender's.
For neurology, the practical question is: when your neurologist opens a new patient's chart, can they see every prior MRI, EEG, and EMG in one place, labeled clearly? Both approaches can get you there for the reports they cover. Only ingestion covers all of them.
When does each approach make sense for a neurology practice?
Neither approach is right everywhere. The decision comes down to how concentrated your outside volume is and how much control you have over your senders.
When an HL7 interface makes sense
Build an interface when three conditions are true:
- A single partner sends a large share of your outside volume. If one hospital or imaging center consistently sends a large slice of your outside reports, an interface can pay off.
- The partner is willing and able. They have IT capacity and a track record of building interfaces with practices like yours.
- You'll maintain it. Someone owns monitoring the interface, so silent failures get caught.
Hospital-owned neurology practices and large neurology groups with a dominant imaging partner often meet all three. Many independent practices meet none.
When AI document ingestion makes sense
AI ingestion makes sense almost everywhere else, which for most neurology practices means most of their volume. It's the right fit when your outside reports come from many sources, when fax and PDF are still the main channels, and when you want one consistent filing standard regardless of sender.
At Honey Health, this is the layer we build for. Our fax triage agent reads, classifies, matches, and files inbound documents, including outside MRI, EEG, and EMG reports, according to each practice's filing rules. Our data fetching agent pulls reports from hospital and imaging portals when they don't arrive on their own. We work alongside existing interfaces rather than replacing them, so practices that already have a feed from their main hospital keep it and use ingestion for the long tail.
The practical answer: a hybrid model
Most neurology practices that solve this problem end up in the same place: interfaces for one or two high-volume partners, AI ingestion for everything else.
The hybrid works because the two approaches fail differently. Interfaces handle the big, predictable feeds cleanly. Ingestion catches the long tail, the faxed duplicates, and the reports routed to the wrong provider. Together, they cover close to all outside imaging, with humans focused only on genuine exceptions.
If you're starting from zero, start with ingestion. It covers every sender immediately, and its data will show you which partners send enough volume to justify an interface later.
That data is worth more than it sounds. After a few months of ingestion, you'll know exactly how many reports each hospital and imaging center sends, how many arrive as duplicates, and which senders produce the most exceptions. Walk into an interface conversation with an imaging partner carrying those numbers and the discussion gets much shorter. You can show them the volume, show them the manual workload it creates on both sides, and make a concrete case for the feed. Without the data, you're asking for a favor. With it, you're proposing a fix to a problem they can see.
A quick way to sanity-check your own split: if your top sender accounts for less than a quarter of outside reports, an interface probably won't change your team's day much. If one sender dominates, it's worth the conversation.
Frequently Asked Questions
Is AI document ingestion as accurate as an HL7 interface?
Not quite, on a per-report basis. HL7 interfaces receive structured data, while AI ingestion interprets documents. Good ingestion tools close the gap by auto-filing only high-confidence documents and sending the rest to a human review queue. In practice, ingestion's broader coverage usually outweighs the small accuracy difference for neurology practices.
Can AI ingestion and HL7 interfaces run at the same time?
Yes, and they should. Keep interfaces with high-volume partners and use AI ingestion for faxes, PDFs, and portal downloads from everyone else. Good ingestion tools detect duplicates, so a report that arrives both by interface and by fax gets filed once.
How long does it take to build an HL7 imaging interface?
It depends on the partners involved. Each interface requires coordination between your EHR vendor, the imaging provider's IT team, and sometimes an interface engine, plus testing. Timelines often run from several weeks to several months, and you don't control the sender's schedule.
Does AI document ingestion work with any EHR?
Most AI ingestion tools work with major ambulatory EHRs through APIs, interfaces, or screen-level automation. Ask vendors which chart sections they can file into for your specific EHR, and whether they can apply your document naming conventions.
What about the DICOM images themselves?
Neither HL7 results interfaces nor document ingestion move DICOM images. Image sharing requires a separate PACS, image exchange network, or viewer workflow. For pre-visit prep, the written report is usually what neurologists need first.
