AI Solutions for Healthcare

A hospital can answer a patient quickly while leaving the actual appointment blocked. The referral is missing a required attachment, the insurer response is still pending, or the service has no suitable appointment. For a patient-access director, the useful AI solution is a reliable readiness record: what has been received, what remains unresolved, who can resolve it and what the patient can truthfully be told.

This is a bounded administrative use case. It can prepare evidence and route work; a clinician retains decisions about urgency, clinical appropriateness and care. A payer decision about coverage is also distinct from a clinical decision. The broader AI solutions service can be scoped around this separation, with connections to existing systems confirmed during discovery rather than assumed.

Choose between access, capacity and clinical-supply workflows

Referral readiness is one operating choice. A hospital with reliable referral records but recurring congestion may gain more from a capacity coordination service. This would combine expected arrivals, confirmed departures, bed status, cleaning and the availability of the staff and equipment needed to operate a service. Its work product is an exception list for the capacity team: a bed awaiting cleaning, a discharge task awaiting an authorised decision or a service whose nominal capacity exceeds staffed capacity. Clinical discharge and patient priority remain clinical decisions.

Evaluate capacity support against decisions made early enough to act. A forecast produced after the morning allocation meeting cannot improve that allocation. Preserve the information available at the planning cut-off, the range of plausible demand and the actions actually available. Compare delays, avoidable transfers and coordinator workload rather than claiming that a more accurate occupancy estimate has automatically created beds.

A second alternative is clinical-supply availability. The useful record links an approved procedure or service requirement to suitable stock, expiry, supplier commitment and replenishment authority. It should distinguish physically present stock from stock held for quality review or already allocated. A proposal to transfer or reorder a clinical item must respect the hospital's approved substitution and procurement processes. The assistant cannot select a therapeutically equivalent product on its own.

Supply and access workflows share identifiers but have different success measures. Access work should improve complete handovers and truthful patient updates. Capacity work should improve feasible allocation. Supply work should improve availability of suitable material while accounting for expiry and emergency purchasing. A hospital should choose the decision with a named owner, an observable bottleneck and usable evidence, rather than commissioning all three because they appear in the same presentation.

The first scope also needs a recovery rehearsal. For referral readiness, introduce a corrected patient identifier after a draft has been prepared and demonstrate that the wrong association is removed from the queue and any dependent draft. For capacity, withdraw an apparent discharge and show that the bed is no longer offered. For supplies, place a lot on hold and show that proposed allocations stop. These are distinct operating tests, and passing one does not establish the others.

During buyer review, ask which existing system remains authoritative for each state and who will maintain the mappings. If the proposed benefit depends on a new approval team, undocumented data access or a capability the hospital does not possess, include that dependency in the scope and cost. Otherwise the pilot may demonstrate a useful interface while leaving the real constraint untouched.

Start with one referral pathway

Choose a service with a recognisable referral package and an accountable coordinator. Define the entry event, such as receipt of an outpatient referral, and the exit event, such as an authorised scheduling handover. Avoid mixing emergency presentations, routine appointments and complex inpatient pathways in the first evaluation: their consequences and escalation routes differ.

The assistant reads the permitted documents, identifies explicit facts and assembles a checklist against the pathway's approved requirements. It must distinguish an absent document from an unreadable document and a missing answer from a negative answer. A blank allergy field cannot become “no known allergies”. Where text implies urgency, the record goes to the designated clinical team without waiting for administrative completeness.

Document processing explains the document-to-case mechanics. In healthcare, the extra operating requirement is to keep patient identity, encounter identity and referral purpose together. Two people with similar names, a duplicate referral and an attachment from an earlier encounter belong in the test set.

Keep readiness, payer response and care decisions separate

The published nphies prior-authorisation guide distinguishes request validation, insurer adjudication and responses including requests for information. Implement against the version and interfaces confirmed for the organisation. A technically accepted message is not evidence of insurance approval, and insurance approval is not an appointment or treatment decision.

Proposed referral readiness record

Field	Evidence to retain	Responsible action
Identity match	Verified patient and encounter identifiers; matching exceptions	Registration resolves uncertainty
Required material	Document type, date, source and readable passage	Coordinator requests specific missing material
Clinical review	Named reviewer and recorded instruction	Clinical team decides priority and suitability
Payer state	Actual response, reference, scope and validity	Authorised payer team checks restrictions
Patient update	Approved wording and delivery record	Access team confirms the next step

Store the original response alongside its explanation. If authorisation covers only part of a request, the summary must name that limitation. A resubmission should reference the existing case where appropriate and avoid creating conflicting duplicate requests. Data extraction controls help verify identifiers and dates; customer-service workflow design covers the subsequent conversation and handover.

A worked readiness calculation

Illustrative example: a clinic receives 240 routine referrals in a week. At the initial review, 150 have the required administrative evidence, 54 need a specific attachment and 36 need identity or pathway clarification. These are mutually exclusive categories totalling 240. The first-pass readiness rate is 150 ÷ 240 = 62.5%. This says nothing about whether treatment is appropriate.

Suppose the assistant prepares 54 attachment requests and reviewers accept 45 without material correction. The acceptance rate for that task is 45 ÷ 54 = 83.3%. It would be misleading to report 83.3% automation of all referrals. Nine requests required correction, and every request still needed the permitted communication and follow-up process.

Measure the waiting time from receipt to a complete handover separately from active staff minutes. If drafting becomes faster but clinical review capacity is unchanged, the overall waiting time may remain flat. Track repeat requests to patients and cases reopened after being marked ready; apparent speed can otherwise conceal avoidable work.

Test the cases that can hurt a patient

Use a review set containing Arabic and English referrals, mixed scripts, scanned forms, expired attachments, abbreviations, conflicting dates and incomplete clinical information. Define critical errors before testing: a wrong-person attachment, invented clinical fact, omitted urgent escalation or a false statement of approval should block release. Do not let a high average extraction score conceal them.

Access to source material should follow the person's operational role. A scheduling coordinator may need readiness status without access to an entire clinical record. Specify what enters the model, what is logged, how long it is retained and which processors receive it. Confirm applicable privacy, security and healthcare obligations for the actual deployment rather than assuming one hosting location resolves them all.

If the intended use expands into diagnosis, prediction or clinical recommendations, reassess the product boundary before proceeding. The SFDA publishes guidance on digital health products; a proposed administrative assistant is not a claim of medical-device authorisation. Clinical safety and regulatory review belong with qualified specialists.

Buy a complete access workflow

A practical scope names one pathway, permitted sources, interface responsibilities, a readiness schema, an exception queue, review rules, bilingual test cases and a support owner. Require evidence that corrected source records propagate to the assistant and that a failed connection leaves cases visible for manual work. Translation controls are useful where the patient-facing explanation changes language; they must preserve restrictions and uncertainty.

Acceptance should combine correct case assembly, appropriate escalation, fewer unnecessary requests and manageable review effort. Include a comparison with the current workflow and sample the cases the assistant calls complete. Expand only when the access team can explain why each case moved and recover when it moved incorrectly.

For related commercial sector reading, Trusted Data's Saudi hospital AI guide discusses a wider range of hospital operating decisions. Those offerings are separate from the proposed OpenForAI scope.

Discuss a referral-readiness workflow with a masked referral package, the current checklist and a few examples of delayed handovers. These are enough to make the first scope concrete without submitting identifiable patient information in an enquiry.