A clinical AI can know the literature and still misunderstand the patient. The next competitive advantage may therefore lie less in producing another general answer than in obtaining the right clinical context, representing it faithfully and recognising its limits. More information is not automatically better information, particularly when an old interpretation is mistaken for a current fact.
OpenAI's September 2026 announcement about connecting authorised health records to ChatGPT illustrates the direction. Its Epic plugin documentation, checked on 15 October 2026, describes organisation-controlled, read-only access in eligible healthcare environments, not unrestricted access for every account or permission to change the record. This is a concrete example of context becoming a product capability with separate technical and organisational conditions. OpenAI's health-record announcement and Epic plugin documentation establish those boundaries.
General evidence and a decision about this patient are different products
A general question can be answered from an appropriate evidence source: what factors should a clinician consider when reviewing a particular condition? A patient-specific recommendation additionally depends on which of those factors are present, absent, uncertain or not yet assessed. It may also depend on previous treatment, the patient's priorities and the services available locally.
That distinction creates a product-design choice. An independent reference tool can help a clinician interrogate a specific knowledge gap without collecting the full record. A record-connected system may reduce repeated data entry and identify relevant historical information. Neither approach resolves the other half of the problem automatically: evidence still needs to apply, and the record still needs interpretation.
The competition is therefore not simply between products with and without an integration. It is between different ways of assembling sufficient information for a defined task, while avoiding unnecessary access and false certainty.
A fictional record that becomes less helpful as it grows
Consider a fictional patient attending with recurring fatigue. Their record includes a previous investigation, an old specialist letter, a provisional explanation recorded during an earlier consultation and a recent statement that their circumstances have changed. The most frequently repeated interpretation is not necessarily the best explanation of the present encounter.
A weak contextual assistant might count repeated mentions as corroboration and foreground the old explanation. A more useful design would distinguish the original observation from later copies, show the dates, identify unresolved findings and ask whether the patient's current account changes the working interpretation. This is a proposed design comparison, not an observed result from a named product.
The clinician also needs to know what the system could not see. Care received elsewhere, a result held in another organisation and an unrecorded patient preference cannot safely become negative findings merely because the connected chart contains no corresponding entry.
Context has several dimensions
A practical context specification should cover identity, time, source, status and relevance. Identity establishes whose information is being used. Time separates current information from history. Source distinguishes a patient report, an examination finding, a copied summary and an external document. Status separates a proposed plan from an authorised decision or completed action. Relevance explains why a particular item matters to the current task.
HL7's FHIR Provenance specification, checked on 15 October 2026, provides a structured way to represent aspects of information lineage. It does not establish that a source was clinically correct or that a later recommendation is appropriate. Authentic information can still be outdated, misunderstood or irrelevant.
This creates an important evaluation question: does the assistant improve clinicians' understanding of the case, or merely produce a more fluent account of what the database contains? Those outcomes require different tests.
EHR access is an advantage, not a complete answer
An electronic health record vendor starts near the information and the operational workflow. That may make contextual launch, user permissions and display within the correct encounter easier to arrange than a separate application. However, technical proximity does not prove that every relevant item is available, correctly interpreted or current.
An independent supplier may work across several record systems or offer a distinct evidence perspective. It also faces integration costs, access restrictions and the risk that its context is narrower than users assume. OpenAI's current Epic documentation is useful precisely because it states deployment and permission boundaries rather than treating the word integration as a complete specification. Epic plugin documentation.
For buyers, the relevant demonstration is not a polished answer from a complete fictional chart. It is what happens when records conflict, a document is missing, the user changes patient or a previously reasonable plan has been superseded.
Privacy-preserving context is a design problem
A proposed architecture could retrieve only the minimum information required for a task, keep access scoped to the authorised patient and encounter, and attach visible source references to the resulting summary. Separate permissions could govern reading, preparing an action and committing a change. These are design options, not a description of an existing iatroX integration.
Data minimisation must also reach secondary systems. A carefully scoped record connection is undermined if extensive patient information is copied into debugging logs, support tickets or unrelated analytics. Access controls, retention, onward processing and the purposes of use need assessment together.
The ICO's AI guidance provides the UK data-protection starting point. It does not make a particular architecture compliant merely because it uses encryption or stores information in the UK. The proposed processing and the organisations' responsibilities still matter.
How to test whether context helps
A useful evaluation would compare the same task under deliberately different information conditions: complete records, incomplete records, conflicting documents and changing circumstances. It should assess clinically important omissions and incorrect assumptions, not only whether the answer contains more patient-specific detail.
Review burden belongs in the assessment. An assistant that surfaces every historical item may create more work than a concise, transparent presentation of relevant uncertainties. Conversely, an exceptionally short summary may conceal the fact that the model never considered a crucial contradiction. More concise and more complete are not interchangeable goals.
Performance should also be examined for patients with fragmented care. A system that works well only for people with extensive, consistent records may not deliver its benefits evenly. The proposed comparison needs enough evidence to distinguish a genuine difference from unstable results in a small subgroup.
The independent reference role should remain narrow and honest
This article is published by iatroX and includes its reference approach among the options discussed. As described in its product materials on 15 October 2026, Ask-iatroX provides source-linked clinical reference grounded in UK guidance and medicines information. It is not presented here as an EHR-connected patient-context service.
For a general knowledge gap, an appropriately framed reference question may be sufficient and involve less information sharing. For reviewing a longitudinal patient plan, authorised record access and clinical assessment may be essential. For conflicting or incomplete information, the priority is resolving the uncertainty, not selecting the product that produces the most personalised-sounding answer.
Patient context could become commercially valuable because it connects knowledge to work. Its clinical value depends on a harder achievement: making clear what is known, what is inferred and what still needs to be established.
Frequently asked questions
Does access to the full patient record guarantee a better AI recommendation?
No. More context can help, but outdated information, copied assumptions and missing external records can also make an answer more confidently wrong.
Is an EHR integration the same as permission to update the record?
No. Read access, drafting and committing a change are separate capabilities and should have separately defined permissions.
When is a standalone reference tool useful?
It can help answer a general clinical question or inspect the evidence behind a recommendation without requiring a complete patient record.