AI Readiness for Healthcare

A healthcare AI proposal can look technically straightforward while depending on several different clinical and administrative truths. A referral exists, a clinician has authorised it, a receiving service has accepted it and an appointment has been attended are separate facts. An assistant that treats them as equivalent can give staff a confident but operationally wrong answer.

AI readiness for healthcare starts with the precise assistance being proposed and the evidence needed to use it responsibly. For a Saudi provider, that means examining patient identity, source meaning, permitted access, Arabic and English records, workflow states, clinical responsibility and the people who will resolve exceptions. The assessment should end with a bounded next-step decision rather than a general declaration that the hospital is ready for AI.

This guide is written for healthcare operations leaders, clinical informatics teams, data owners and digital-transformation sponsors. Its worked example is an outpatient referral-routing assistant that prepares administrative suggestions for authorised staff. It does not determine clinical urgency, diagnose a condition, choose treatment or independently accept or reject a patient.

Define the intended assistance before assessing the dataset

Write one sentence describing the user, the decision and the proposed output. In the example, a referral coordinator needs a source-linked summary of the receiving department, administrative completeness and unresolved routing questions. A qualified person retains responsibility for interpreting clinical instructions and taking any action. The assistant works inside that existing process.

Then record what a user might reasonably mistake the output for. A label saying “suggestion” offers little protection if the interface presents a department as confirmed, hides the source date or makes acceptance the default. Readiness includes the output design and operating procedure, not only the underlying records. Staff should be able to distinguish a recorded referral instruction from a model-generated interpretation.

Intended use also matters to regulatory classification. SFDA's published guidance on AI and machine-learning medical devices explains the relevance of intended purpose to medical-device regulation. A project team should obtain the appropriate regulatory and clinical assessment for its actual function. Calling a system administrative does not settle its classification if its behaviour influences clinical decisions.

Keep proposed future capabilities outside the initial conclusion. Adding urgency prediction, patient-facing advice or automated referral rejection changes the question being assessed. The evidence sufficient for administrative summarisation should not be carried over as approval for those functions. Capture the boundary in the decision pack and in the change process for the pilot.

Separate referral, task and appointment states

The HL7 FHIR R4 ServiceRequest specification provides a useful example of a request resource and distinguishes it from fulfilment tracking. It also separates the intention behind a request from its status. These distinctions help frame a mapping discussion; they do not establish that a particular Saudi provider uses FHIR R4 or any specific implementation profile.

Ask the local system owner to demonstrate the real lifecycle. Which field means the referring clinician signed the order? Which event means the receiving team accepted responsibility? Does “closed” mean completed care, an administrative cancellation, a duplicate or a transferred request? A local label can hide several meanings, especially after system migrations or departmental customisation.

Referral-routing evidence map for the illustrative assessment

Question	Evidence to examine	Readiness consequence
Is this the correct patient and encounter?	Approved identity crosswalk, merge history and encounter relationship	Unresolved matching prevents a patient-specific summary
What was actually requested?	Signed instruction, request intent, relevant version and amendment	Do not replace the authorised request with an inferred service
Who currently owns the next administrative step?	Receiving-team acceptance and current routing rules	A proposed destination remains a proposal until accepted
Has the requested service occurred?	Appropriate fulfilment or attendance evidence	An order or booking alone cannot establish completed care
Was the request cancelled or superseded?	Status reason, event sequence and replacement reference	Exclude obsolete instructions from current routing suggestions

Make the map specific to the selected facilities. Two hospitals in the same group may have different department identifiers and responsibility boundaries. A shared warehouse column does not prove a shared operational definition. Record who can confirm the meaning and which source wins if an export conflicts with the transaction system.

Use a referral sample that exposes the routing problem

Consider a hypothetical provider reviewing 60 deliberately selected referral episodes before designing a controlled pilot. The selection includes Arabic and English notes, amended requests, cross-facility transfers and incomplete administrative records. It is a challenge set for discovering prerequisites, not a random estimate of the provider's overall data quality or patient population.

The reviewers classify 36 episodes as having a traceable instruction and unambiguous current administrative owner. Ten contain a later amendment that the existing export does not expose clearly. Eight have an unresolved link between the source request and receiving-service record. Six lack the evidence needed to interpret a local closure state. These four mutually exclusive groups total 60 episodes.

The result does not mean the model achieved 60% routing accuracy. No model has been evaluated by those counts. It means 36 of the selected episodes presently support the agreed evidence question, while 24 need clarification or repair. Report the categories separately so the sponsor does not purchase prompt tuning to solve an identity or lifecycle problem.

A reasonable proposed decision is to assess a narrower pilot population with verified request and ownership mappings while holding the affected categories. That still requires proving the team can identify the included population before an answer is generated. Manually choosing the 36 clean examples for a demonstration does not establish a reliable production inclusion rule.

Build the reference answer with the right reviewers

Have an authorised referral coordinator and an appropriate clinical representative resolve ambiguous source instructions through the provider's established process. Preserve their reason and source references. A disagreement between reviewers is evidence to investigate, not a label to settle by majority vote without examining the underlying responsibility.

The reference should identify the allowed administrative output, any information that must remain unresolved and the required escalation. It should not manufacture clinical certainty to make a benchmark easier to score. Some cases correctly end with a request for human clarification, and the evaluation must recognise that as the intended outcome.

Check Arabic meaning without changing clinical authority

A bilingual record may contain an Arabic narrative, an English specialty name and an abbreviated local department code. The assessment should test whether the pipeline preserves those relationships. A fluent translation is insufficient if it changes a negation, treats a historical instruction as current or attaches a note to the wrong referral.

Start with Arabic OCR and document quality when the authoritative instruction is scanned. Record which fields come from structured data and which require interpretation of an image. If the scan is illegible, the correct readiness finding is an unresolved source, not permission for the model to fill the gap from a typical clinical pathway.

Use bilingual reviewers for cases where language changes the routing interpretation. Define how an approved amendment in one language supersedes an earlier version in the other. The bilingual evidence and source-conflict method helps make this explicit. Neither language should gain authority merely because the retrieval system finds it more easily.

Retain meaningful distinctions in the staff-facing summary: “the source requests”, “the receiving service recorded” and “the coordinator must confirm” communicate different evidence states. Test them with realistic users. A well-written Arabic summary must preserve the uncertainty and professional responsibility present in the source, not simply sound natural.

Assess privacy and access at the patient-workflow boundary

Health information is a sensitive-data category under Saudi Arabia's Personal Data Protection Law. The provider must establish the appropriate purpose, basis, safeguards and responsibilities for its specific processing. The assessment should identify what personal information is necessary for the routing task and what would be an unnecessary addition to the model context.

Use PDPL data classification to distinguish the information involved and data residency and hosting analysis to examine the complete processing flow. A Saudi application endpoint alone does not describe the location or access arrangements for extraction, model processing, logs, support or retained evaluation material. Obtain evidence for the actual deployment proposal.

Test user and facility boundaries. A coordinator's ability to search one clinic's records should not automatically become access to all records in a group-wide index. Investigate cross-cover arrangements, temporary assignments and revoked roles. The system must preserve the approved care and administrative access model rather than creating a new broad permission through convenience.

Assessment evidence itself needs controlled handling. Prefer an authorised demonstration, appropriately prepared samples or a bounded review environment over sending full patient histories to a supplier. Record what the assessor could inspect and what remains unverified. Masking can reduce exposure, but poorly masked records may still identify individuals through context or linked identifiers.

Preserve identity corrections and repeated referral episodes

Inspect what happens when the organisation merges duplicate patient records. A warehouse may retain both historic identifiers while the operational system presents one current record. The assistant needs an approved relationship, not a name-based guess. Ask whether previous referral links are updated, whether the change reaches every extract and how a mistaken merge is corrected. A familiar patient name does not establish that two episodes belong together.

Repeated referrals require a separate check. The same patient may have several requests for the same specialty at different times. A later acceptance cannot automatically be attached to an earlier cancelled request. Preserve the request identifier, relevant encounter, replacement relationship and event order. Reviewers should be able to reconstruct why a particular instruction is current without relying on the latest timestamp across unrelated records.

Consider a synthetic case in which one request is cancelled and replaced after an administrative correction. The export contains the replacement but omits its link to the original. The assistant might present two active requests or attach the old receiving-team response to the new request. The readiness action is to repair or explicitly resolve the relationship before that category enters the pilot. A polished combined summary would conceal the defect.

Keep the patient's clinical history distinct from the administrative evidence needed for this task. More context can add unrelated episodes and sensitive information without resolving the missing link. The assessor should ask for the smallest sufficient evidence set and document the conditions under which an authorised reviewer needs additional information. This supports both interpretability and proportionate access.

Rehearse a source update before allowing current-state answers

Use an authorised synthetic episode to demonstrate an amendment moving from the source system into the extract and then into the assistant's evidence. Record when each stage becomes available. The required freshness depends on the intended workflow and its consequences; the assessment should not invent a universal refresh interval. It should establish what staff will see while an update is pending.

If a source becomes unavailable, a previous snapshot may still support historical reporting while being unsuitable for a statement about current referral ownership. Distinguish those uses in the response design. A visible date is useful, but it does not by itself make an outdated instruction appropriate. Agree when the assistant must withhold a current-state answer and direct the coordinator to the established source.

Finally, reconcile the pilot's sample with the operational queue. Count requests entering, remaining unresolved, leaving the eligible population and receiving human handling. This does not require publishing identifiable records. It requires a controlled account of what the assistant actually covered, so that apparently faster handling is not explained by quietly moving difficult referrals outside the measured process.

Measure a useful administrative outcome

Define the denominator before selecting a headline metric. For the proposed referral assistant, an eligible episode has the approved source set, verified patient relationship and supported workflow state. Report excluded and unresolved episodes separately. A high score on included cases becomes misleading if difficult referrals disappear from the measurement without explanation.

Proposed measures for a referral-assistance pilot

Measure	What it should reveal	Necessary qualification
Supported administrative suggestions	Whether the suggestion follows the authorised current instruction	Independent review against the source, not agreement with the model
Incorrect current-state assertions	Whether cancelled, superseded or unaccepted requests are misrepresented	Separate consequential errors from harmless presentation differences
Appropriate escalations	Whether unclear identity, source or ownership reaches the right person	Measure unnecessary escalation as well as missed escalation
Total coordinator handling effort	Whether preparation, checking and correction change the workload	Include review time and subsequent rework within the agreed observation window
Coverage of the intended population	Which referral categories can actually use the assistant	Show exclusions, facilities and language segments

Set acceptance conditions with clinical and operational owners before observing the pilot results. The conditions are local design decisions, not universal regulatory thresholds. The responsible team may require a certain class of identity or instruction error to stop the pilot regardless of average performance. Document who can pause use and how staff revert to the established workflow.

Do not claim patient-outcome improvement from a short administrative demonstration. Reduced preparation time, fewer routing corrections and clinical outcomes are different propositions with different evidence needs. Begin with the outcome the bounded workflow can actually measure and maintain a record of any downstream effects that require further investigation.

Establish who can operate the pilot on a difficult day

Ask the service manager what happens when the receiving department changes its coverage rules, the source feed is delayed or the nominated reviewer is absent. Readiness requires a maintained route for those situations. A successful test attended by the project team does not show that routine staff can recognise and manage a stale or incomplete answer.

Define the human review workload. If every suggestion needs several systems opened and a senior clinician contacted, the proposed assistance may move work rather than reduce it. Observe a small authorised workflow rehearsal and document the steps. The decision might be to improve the source handover first, which can be valuable even if no model is deployed.

Use ownership, evaluation and adoption readiness to assign responsibilities for source corrections, referral rules, application support and evaluation updates. Avoid one undifferentiated “AI owner”. A technical team can maintain the interface without having authority to reinterpret a clinical instruction or change an administrative acceptance policy.

Preserve a way to challenge the assistant's suggestion. The reviewer should see the relevant source and be able to record why a suggestion was rejected. Those corrections are not automatically reliable training labels: some reflect new information or a changed policy. Keep their timing and reason so later evaluation does not mistake changed circumstances for model error.

Prepare a healthcare readiness decision that can be acted on

The final pack should state the proposed function, selected facilities, users, patient-data boundaries, source mappings and evaluation population. Attach the referral-state map, sampled exceptions, privacy and hosting decisions, unresolved dependencies and named operating responsibilities. Identify the evidence that supports a limited next step and the conditions that still prevent it.

For the illustrative provider, the immediate backlog might include exposing amendment references in the export, resolving request-to-receiving-record links and documenting closure reasons. Each item needs an owner and a test showing the repaired data reaches the proposed assistant. A database correction that never enters the relevant export does not remove the readiness gap.

Separate assessment work from implementation commitments. Building a connector, changing referral processes, obtaining specialist approvals or conducting a wider evaluation can require further scope. The sponsor should understand which decisions the assessment supports and which delivery activities remain. A usable conclusion can be a narrower pilot, a remediation phase or a decision to defer.

For related commercial reading, Trusted Data's healthcare AI overview discusses broader sector applications. The readiness question here remains the evidence for the provider's own defined workflow. Start with an enterprise data and AI readiness assessment and bring a masked process map, system inventory and known referral exceptions to a healthcare readiness assessment discussion. That creates a concrete basis for agreeing the next decision.