An automated referral can fail through the wrong destination, missing information, rejected triage, an uncompleted booking or an unnoticed change in the patient's circumstances. A well-written letter addresses only part of that pathway. Effective automation needs to preserve the clinical purpose and show what has actually happened after the request leaves the referring team.
The letter can be correct while the pathway is wrong
Consider a fictional referral for a persistent hand problem. The clinician agrees a specialist assessment and approves an accurate letter. A hypothetical agent selects a service whose name appears appropriate, but the service does not accept that particular presentation through the chosen route.
The provider returns the request with instructions for an alternative pathway. If the agent's task ended when the letter was sent, the patient might believe they are waiting for an appointment while the referring team sees a completed item. The error is not necessarily in the clinical prose. It is in the definition of completion.
This is a constructed scenario, not an allegation about any commercial product. Its purpose is to separate content generation, service selection and successful progression through the referral process.
The distinction is visible in existing NHS systems. NHS England's e-Referral Service guidance, reviewed on 10 October 2026, describes returned triage responses and rejected referrals that require review and action. A sent referral is therefore not interchangeable with one that has been accepted or booked.
Preserve the reason for the referral
A proposed referral specification should include the clinical question, the intended service, the agreed urgency and relevant information needed by that service. The point is not to create another administrative form. It is to make the decision's important assumptions available when software translates it into an action.
For the fictional hand referral, the key issue might be a specific functional problem that determines which pathway is suitable. A summary that retains the broad symptom but drops that distinction may direct the patient to the wrong destination despite sounding clinically reasonable.
Local service information also requires maintenance. The appropriate route cannot always be inferred from a specialty name or an old referral. A buyer should establish which source the agent uses for availability and eligibility, how changes are incorporated and what happens when the information is incomplete.
This is not a claim that a general AI should choose clinical urgency independently. The extent of any automated decision needs to match the product's assessed scope and the organisation's authorisation.
Distinguish booking routes rather than assuming one sequence
Some services offer directly bookable appointments; others assess the referral before deciding the next step. NHS England's explanation of referral assessment services, reviewed on 10 October 2026, describes a pathway in which assessment can lead to onward booking, advice or a returned request.
An automation workflow should reflect the route actually used. It should not claim that the patient has an appointment because a referral identifier exists, or that a specialist has accepted responsibility because an assessment service received the information.
Patient choice and practical arrangements also matter. An offered appointment may be unsuitable because of accessibility, travel or other circumstances. Treating a booking response as a simple success-or-failure field can obscure a need to discuss a different arrangement.
The proposed objective is an appropriate next step that the patient can understand and use, not merely a completed transaction in the referring software.
Keep rejected and incomplete work visible
A useful exception queue should make clear why a referral has stopped progressing. Missing attachments, a returned request, a patient unable to book and an unavailable service require different responses. A generic "failed" status gives little help to the person taking over.
Ownership should follow the actual arrangements between teams. The person correcting a technical attachment problem may not be authorised to change the clinical question or select a different specialty. The workflow must distinguish those tasks and provide a route back to the clinician when interpretation is required.
The GMC's delegation and referral guidance, reviewed on 10 October 2026, supports clear communication when responsibility passes between professionals. The corresponding design question for automation is whether the intended recipient and the outstanding clinical work remain visible throughout that transfer.
The service should also specify who checks the exception queue when the original user is absent. An automated workflow that works only while its enthusiastic early adopter is present is not yet a dependable organisational process.
New information can invalidate a technically successful action
Return to the fictional patient. Before an appointment is arranged, they report a meaningful change in the problem. The correct response may require reassessment of the original plan. An agent that continues optimising the booking process without surfacing the change could complete an action that no longer reflects the current clinical decision.
A proposed control is to link consequential changes to pending actions. The system should show the responsible team that the referral is already in progress, what has changed and whether further review is required. It should not silently cancel or amend an agreed clinical plan beyond its authority.
This introduces a useful evaluation case: the original instruction is valid when given but becomes outdated before completion. A system should not receive full credit merely because it faithfully executed the earlier instruction.
Test receipts, retries and the patient's understanding
A technical timeout can leave uncertainty about whether a referral was received. Retrying without checking may create a duplicate; assuming success may leave the patient un-referred. A proposed evaluation should therefore test reconciliation using the relevant system's identifiers and status information.
The test should also examine what the patient is told. "Your referral has been submitted and is awaiting assessment" has a different meaning from "Your appointment is booked". The wording should reflect the confirmed state, including who to contact when the expected next step does not occur.
Measure the complete pathway: appropriate destination, sufficient information, accepted transfer where required, unresolved work and patient comprehension. Include referrals returned for good clinical reasons rather than counting every rejection as an AI error. Conversely, do not exclude them simply because the software successfully transmitted the letter.
Turn the pathway into a reasoning exercise
A useful educational exercise presents a correct referral letter alongside an incomplete workflow record and asks what remains to be established. The learner must distinguish a clinical decision, an administrative action and evidence that the next service has progressed the request.
As described by iatroX in October 2026, Brainstorm can support structured discussion of a fictional case. It is not being offered here as an NHS referral system or an automated booking service. Reference and learning tools can help clinicians examine the reasoning behind a referral while operational ownership remains in the approved clinical pathway.
For a team seeking drafting support, an accurate letter may be the relevant product goal. For a team seeking end-to-end automation, the evaluation must continue after the letter is sent. Those are different purchasing decisions and should be assessed accordingly.
Frequently asked questions
Does an accurate AI-written letter establish that a referral is safe?
No. The destination, urgency, information requirements, subsequent responses and follow-up arrangements also need to fit the patient's situation.
Is a referral identifier proof that an appointment has been booked?
No. An identifier can establish that a request exists without establishing acceptance, assessment or a confirmed appointment.
Should an agent automatically change a rejected referral?
Only within its explicitly authorised and assessed scope. A returned request requiring clinical interpretation should reach an appropriate professional rather than be treated as a purely administrative retry.
Explore the reasoning and missing steps in a fictional referral →