How to Design Agentic AI Workflows for Healthcare in Dubai

A referral can fail before the appointment begins

A referral may be delayed because a test result is missing, the receiving specialty is unclear or no one owns the next action. It may also reach the wrong patient record, omit an urgent symptom or secure an appointment that the receiving service has not accepted.

Connected health records make information available across facilities, but access does not establish that two records belong to the same patient or that the clinical evidence is complete. The workflow must therefore verify identity, consent, provenance and referral state before it coordinates the next step.

Use the agent for administrative completeness

An agent can locate permitted records, preserve the reason for referral, identify missing evidence, track deadlines and find feasible appointment options. It can keep the initiating and receiving teams informed about the same case state.

The agent should not diagnose, change clinical priority or communicate a material clinical decision without professional review. Research on healthcare agents consistently treats recommendation, authorisation and execution as separate responsibilities, particularly for high-risk or irreversible action.

The first useful scope is therefore controlled administrative coordination. The agent prepares a complete case and makes unresolved clinical questions visible; licensed professionals decide urgency, diagnosis and acceptance.

Define the minimum complete referral

The pathway should specify the evidence required before a receiving team can review the case. Dubai Health Authority’s current referral policy identifies required referral fields such as the initiating and receiving facilities, patient identifiers, diagnosis, procedures, relevant history, medications, reason for referral, material findings, patient status, follow-up instructions, insurance information, the referring professional and the referral time. The workflow should represent those required fields accurately and add pathway-specific evidence where the receiving service needs it.

The agent can check the administrative presence and provenance of these items. However, it cannot infer that an absent test is normal or that a brief note captures every material clinical constraint. Missing evidence must remain visible and return to an accountable owner.

The referral record should identify who supplied each clinical fact, which policy or pathway applied and which licensed professional accepted responsibility. This record supports correction if the summary omits a constraint or a later result changes the appropriate route.

Preserve time and state across the pathway

Referral evidence changes while the case is moving. A new laboratory result may increase urgency, a patient may withdraw consent or capacity may disappear before booking. A workflow that relies on a stale snapshot can produce a decision that was once plausible but is no longer safe.

The agent should record when each fact was obtained and verify material state before acting. A booking request, acknowledgement and confirmed appointment are different events. The system must not tell a patient that the referral is accepted until the receiving service records that state.

Test realistic clinical and administrative failures

Static question-answer tests cannot establish whether a healthcare agent uses electronic records correctly. MedAgentBench, for example, evaluates agents through patient-specific tasks in an interactive health-record environment. The broader lesson is that workflow performance must be tested against records, tools and state rather than medical prose alone.

A referral evaluation should include similar patient identities, missing consent, contradictory results, urgent symptoms embedded in long notes, inaccessible records and receiving-capacity failures. It should also test whether the agent retains the clinical reason during summarisation and refers uncertainty to a licensed professional.

State checks establish which patient record the agent accessed, which evidence it retrieved and whether an appointment exists. Clinical reviewers assess whether the summary preserves the information required for safe review. Every clinically consequential miss should be investigated separately rather than averaged into a general score.

Pilot one non-emergency pathway

The pilot should begin with one non-emergency referral pathway that has clear evidence requirements and immediate escalation rules. Protected or synthetic cases allow the team to test sensitive failures without using unnecessary patient information.

The first phase can run in shadow mode. The agent prepares the referral while staff follow the existing process. Reviewers compare completeness, correction, routing and elapsed time before any live transaction authority is granted.

Expansion should depend on the failure pattern. If errors concentrate in patient matching, the programme needs stronger identity controls. If summaries lose clinical constraints, the agent needs a different evidence and review design. A larger model will not correct a broken responsibility or handoff rule.

Example: a referral with a late test result

A primary-care clinician refers a patient to a specialist and records a non-urgent reason. The agent verifies identity and consent, retrieves the required investigations and finds an available appointment. Before the receiving team accepts the case, a new result introduces a potentially urgent concern.

The workflow detects the new result because the referral remains open and material evidence is time-stamped. It stops the routine booking path, alerts the initiating clinician and preserves the original referral and new evidence for professional review. The agent does not independently change the clinical priority.

The test verifies the patient record, timing of the new result, escalation event, professional decision and final appointment state. A fluent patient message cannot compensate for a missed escalation or a booking based on stale evidence.

Measure safer coordination

The scorecard should include referral completeness, clinician correction, inappropriate routing, elapsed time, repeated contact, delayed appointments and verified appointment state. Safety events and privacy exceptions require individual review.

Results should be segmented by pathway, urgency and material patient conditions. An overall improvement can otherwise conceal a failure pattern among cases with incomplete records or urgent evidence.

How Marketways designs the referral workflow

Marketways maps one referral pathway, measures avoidable rework and waiting, defines the minimum complete evidence set and designs a non-diagnostic coordination pilot. We specify identity, consent, evidence, capacity, escalation and final-state controls.

The evaluation uses protected or synthetic cases and licensed clinical reviewers for clinical preservation. Marketways provides workflow design, analytical evaluation and implementation guidance; healthcare providers and authorised professionals retain clinical, legal and regulatory responsibility.