A referral is a request, not a completed pathway
The unit of orchestration is a person with a goal, a set of needs and a changing context. It is not the message exchanged between two systems. A technically successful referral can still fail because the receiving service lacks capacity, the patient cannot reach it, coverage is unclear or the plan does not reflect the person’s priorities.
I would define orchestration as a management commitment: someone can see the unresolved step, understands why it matters, knows who can act and has authority to escalate. The operating design must reach across organizational boundaries without pretending that one organization owns every clinical decision.
Test the causal chain, not the activity count
The interpretation is not that coordination lacks value. It is that closing one intermediate gap does not prove the entire theory of change. If attendance improves but the intended outcome does not, examine what happened during and after the visit, the availability of treatment, medication continuity and barriers beyond the clinic.
This is a capital-allocation question as well as an evaluation question. The next dollar may belong in receiving capacity, clinical follow-through or practical support rather than another layer of outreach. The choice should follow the failed link in the causal chain.
Design four kinds of continuity
| Continuity | What must travel | What leadership must resolve |
|---|---|---|
| Clinical | Goals, relevant history, current plan and changes requiring review. | Who is the accountable clinician, and how are conflicting plans reconciled? |
| Operational | Acceptance, appointment, capacity, completion and unresolved work. | Who owns the next step when the receiving service cannot act? |
| Economic | Coverage, authorization, payment responsibility and service cost. | Who funds the enabling work, including work that sits between reimbursable encounters? |
| Relational | Patient preferences, language, trust, caregiver context and consent. | How does the person shape the plan, and who helps when the pathway becomes difficult? |
These are not four separate platforms. They are four requirements that the operating model must meet. A shared dashboard is useful only when it makes the next decision easier and the receiving team has the capacity and authority to carry it out.
Give AI bounded work inside the pathway
AI can support record reconciliation, source-linked summaries, administrative classification, gap detection and the preparation of a next-best-action recommendation. Each use should specify what information it relies on, what it may do and what it must leave for an authorized person.
I would separate a proposed action from its execution. A system suggesting outreach has a different risk profile from one changing clinical instructions or resolving a coverage dispute. The latter cannot inherit authority merely because it has access to the data. Decisions with clinical or coverage consequences need the relevant authorized review and a traceable basis.
| Control | Design requirement |
|---|---|
| Context | Identify the person, source, freshness, consent and purpose before using information. |
| Policy | Apply the relevant organizational, clinical and contractual rules to the specific action. |
| Authority | Name the person or role allowed to approve, override or stop the action. |
| Execution | Record what happened, to whom, when, and whether the receiving service accepted it. |
| Recovery | Provide an exception route, correction process and a workable return to manual operation. |
| Evaluation | Measure completed care, workload, access, quality and cost across the whole pathway. |
NIST’s lifecycle approach to AI risk management is a useful foundation for governing these controls; the care-pathway architecture here is my strategic application. [2]
Make an unresolved referral actionable
Consider an illustrative post-discharge pathway with a follow-up visit, medication access and a transport need. Each step needs an accepting service, a completion signal and a clinically appropriate escalation time agreed by the care team. The pathway remains open when a task was sent but not accepted, when a service was unavailable or when the person declined and needs a different plan.
A shared status should describe the actual state: requested, accepted, scheduled, completed, declined or blocked. The blocked state should identify the reason and the person responsible for the next decision. Timing thresholds must reflect local clinical protocols and service agreements; a generic software deadline is not a clinical standard.
The leadership review then sees where capacity, coverage or process repeatedly breaks down. That turns individual exceptions into an informed network investment decision while keeping clinical judgment and patient choice intact.
Measure whether the added layer earns its place
Orchestration adds work of its own: coordination, data exchange, governance and exception management. Its economic case should identify the failure it resolves and compare its cost with the value of resolving that failure. More participating organizations can improve reach while also increasing the cost of alignment.
I would retain an orchestration layer when it improves a defined care pathway and gives accountable teams better control of unresolved work. If it only produces more notifications, more duplicate documentation or a cleaner dashboard without better service completion, the architecture needs to change.
Sources & analytical basis
The strategic recommendations are the author’s interpretation. External research and company observations are attributed below; illustrative scenarios are labeled where used.
- Finkelstein et al., The Camden Coalition Care Management Program Improved Intermediate Care Coordination: A Randomized Controlled Trial. Health Affairs, January 2024; corrected October 2024. Located and fetched through Consensus; corrected public manuscript reviewed.
- NIST, AI Risk Management Framework 1.0. January 2023.