An AI Mission for Healthcare: Referral Management

Every referral is a transfer of responsibility between two organisations that keep separate books. The work worth automating is not sending it faster — it is making sure somebody, or something, is still holding it tomorrow.

A referral coordinator's desk is one of the few places left in a modern health system where three generations of communication technology sit side by side and none of them agree. There is a fax line, still running, because a proportion of receiving practices will accept nothing else. There is a stack of specialist portals, each with its own login, its own upload requirements, and its own idea of what constitutes a complete packet. There is a work queue inside the electronic record that shows referrals as placed, which is a status about an order and not a status about a patient. And there is a phone, which is where the actual system lives, because the phone is how the coordinator finds out that the imaging report never made it across, that the authorisation lapsed while the packet sat in someone's inbox, or that the appointment the patient was told to expect was never scheduled at all. The coordinator is not disorganised. They are doing manual reconciliation between organisations that were never designed to reconcile.

What makes this so persistent is that nothing in the process is actually broken in the way that word usually implies. The referring clinician made a judgment, documented it, and placed the order — and from inside that practice, the referral is done. The receiving practice, meanwhile, cannot act on something it never fully received; from inside that building, there is no referral, only an incomplete fax with a missing prior and no authorisation attached. Both sides are behaving reasonably and both sides are looking at accurate information. The patient is the only party who can see the whole shape of what happened, and the patient is the party with the least ability to do anything about it. That asymmetry is not a workflow defect. It is the structural condition of every handoff that crosses an organisational boundary, and it is why the problem the industry calls referral leakage is so resistant to the tools thrown at it.

Leakage is a custody problem wearing the costume of a routing problem

The word leakage does a lot of quiet damage, because it frames the failure as a plumbing issue — volume escaping a network, to be recaptured with tighter directories and better steerage. That framing produces solutions aimed at the moment of send: better provider search, cleaner network lists, one more integration between one more pair of systems. All of it helps at the margin, and none of it touches what actually goes wrong, because what goes wrong almost never happens at the moment of send. It happens afterwards, in the interval nobody owns.

The more useful frame is custody. In any discipline where things must move between organisations without being lost — evidence, freight, financial settlement — the mechanism that makes the transfer reliable is not speed but an unbroken record of who is holding the item and what condition it is in. Every hop is logged, every gap is visible, and responsibility is never ambiguous, because ambiguity is precisely what allows something to disappear. A referral has none of this. It leaves one organisation's system of record and enters a gap where no system asserts ownership, and it stays there until either a person picks the phone up or the patient gives up. There is no shared ledger, no acknowledgement that means anything, and no artefact anywhere that says this handoff is still open and here is who is accountable for closing it. The referral does not leak because it was routed to the wrong place. It leaks because during the period when it most needed a custodian, it had none.

Once you see it this way, the ordinary catalogue of referral failures stops looking like a list of unrelated problems and starts looking like a single one repeating. A packet arrives missing the prior imaging that would let the receiving practice triage it, so it sits. An authorisation is obtained and then expires because the scheduling step took longer than the approval window and nobody was watching both clocks. An appointment is offered and declined, and the decline reaches nobody upstream. A consultation happens and the note never returns to the referring clinician, which means the loop that justified the referral in the first place never closes, and the referring practice learns nothing about whether their patient was seen. In each case the individual task is trivial. What is missing is continuity — a thread that persists across the boundary, knows what state the handoff is in, and is responsible for keeping it moving.

Continuity is the automatable part; the clinical judgment is not

It matters enormously where the line is drawn here, and the line is not subtle. Deciding that a patient needs a specialist, deciding which specialty, deciding urgency, deciding what the consultation should investigate — that is clinical judgment, it belongs to the clinician who examined the patient, and no software should be making, adjusting, or second-guessing it. Nor should anything in the process be optimising which provider a patient ends up with on financial grounds; the referring clinician's determination of what the patient needs is the input, not a variable to be tuned. What sits outside that boundary is everything that happens to a decision after it has been made: assembling the packet the receiving practice actually requires, confirming that it arrived intact, tracking whether an appointment was offered and kept, noticing when an authorisation window and a scheduling delay are on a collision course, and making sure the consultation note finds its way back to the person who asked for it. None of that is medicine. All of it is custody, and custody is exactly the kind of work that has always been done by a person with a phone because there was nothing else capable of doing it.

The reason there was nothing else is that this work does not decompose into a fixed path. Integration engines and rules-based workflows can move a message reliably when the message is well-formed and the route is known, but the referral failures that matter are the ones where the shape is wrong — a packet that is complete by the sending system's definition and incomplete by the receiving one's, a portal that changed its upload requirements, a receiving practice that acknowledges receipt with a fax cover sheet and nothing machine-readable. A workflow with a fixed path hands every one of those back to a human, and in referral management the exceptions are not the edge of the work. They are most of it. This is the same wall that enterprise buyers are hitting across every industry: Gartner has predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, citing unclear value, weak risk controls, and what the firm calls "agent washing" — old rule engines and chatbots relabelled without any change in what they can do unsupervised. A faster router is still a router. It hands the referral off and stops caring, which is the original failure with better latency.

What closing the gap actually requires is a system that treats an open referral as something it is holding rather than something it forwarded. That means keeping a live model of the handoff's state across both organisations' systems, reading what comes back in whatever form it comes back — portal, fax, message, phone — and inferring from it whether the handoff advanced or stalled, then acting on the stall: re-sending the missing prior, re-requesting the authorisation before it lapses, chasing the consultation note, and escalating to a human coordinator when the situation needs a person, or to the clinician when anything touching clinical judgment arises. This is the operating shape behind what a growing body of work now describes as the move toward an autonomous enterprise, and in healthcare operations specifically it is the premise behind running AI Missions on a platform like StudioX — specialist agents working the administrative layer of the handoff, connected to source systems through the Model Context Protocol, with human-in-the-loop gates on anything clinical, financial, or consequential to a patient. The agents do not decide where a patient should go. They make sure that the decision the clinician already made does not quietly dissolve in the space between two organisations.

The reframing worth carrying out of this is that a referral is not a message and should never have been managed as one. It is a custody transfer with an open interval in the middle, and the only meaningful measure of a referral programme is not how many were sent, how fast they went out, or what share stayed in network — it is what proportion of open handoffs had an identifiable custodian at every moment between the order and the returned note. Health systems that start measuring that number will find it uncomfortable at first, because for most of them the honest answer during the critical interval is the patient. That was never a defensible allocation of responsibility, and it is now an avoidable one.