Skip to content
Get in touch

Human review in assisted workflows

A worked example of an AI-assisted discharge summary: which record data is read, under which rules, how the draft shows its sources, and the clinician's review.

7 min read


AI assistance becomes useful in clinical work when it prepares material a clinician would otherwise assemble by hand, from the hospital's own record, with every statement traceable to its source. mona AI is built around that pattern: an AI reads the patient's record in monaOS under fixed rules, prepares a draft, and the clinician reviews, edits and documents. Nothing is written back by the AI. This article follows one worked example, an assisted discharge summary, through each step: what is read, how the gateway governs it, what the draft looks like, and where the clinician takes over.

Two ways in, one set of rules#

mona AI offers two ways to use AI on the hospital's own data, and a hospital can run either or both.

  • mona AI is the hospital's own AI assistant, run by Clinomic under the hospital's name, logo and colours. It answers from the sources the hospital defines, with a link to the source behind every answer, and it reads the patient's record in monaOS. Its workflow management and AI Scribe modules draft documents such as discharge letters for sign-off.
  • Bring your own AI connects the AI tool the hospital already uses (for example Claude, ChatGPT, Gemini, Microsoft Copilot or Le Chat) through the monaOS AI server with an access key. Clinical Skills give that tool worked-out workflows for tasks such as discharge summaries, progress and SOAP notes, handover material and laboratory result overviews.

Both options read the record through the same governed access, so the review model below applies to both.

Step 1: what the draft is built from#

A discharge summary condenses a stay, so its input is the structured record of that encounter: the admission diagnosis and the diagnoses documented during the stay, laboratory trends, ventilation and other device-derived values, prescriptions and their changes, procedures, and the documentation of the course. In the FHIR-native record these are standard resources bound to the patient and the encounter:

Content of the summary FHIR resources that hold it
Patient and stay Patient, Encounter (with its period and locations)
Diagnoses and allergies Condition, AllergyIntolerance
Laboratory, vital signs, ventilation, scores Observation
Medication course MedicationRequest, MedicationAdministration
Interventions Procedure
Structured forms and assessments QuestionnaireResponse
Reports and documents DocumentReference

Every value comes with its clinical time and its source, and only current valid versions count: values marked as entered in error are excluded, as described in corrections, versions and provenance. The binding to patient and encounter is what separates this from free text pasted into a chat window; why it matters is covered in clinical context for AI.

Step 2: how the monaOS AI server governs access#

The monaOS AI server is a guarded gateway between the AI and the record. It runs as its own service beside monaOS, on the Clinomic Cloud or on-premise, needs only the FHIR endpoint and an access token, keeps no database of its own and requires no change to monaOS.

Rule What it means for the discharge draft
Read-only in code The server has no write path. It can fetch the stay's data; it cannot create, change or delete anything.
Query rules Only permitted resource types and search parameters are accepted. High-volume data needs a filter, so laboratory and vital-sign searches name the patient and a time range. Chained queries, reverse includes and server operations are refused before any request is sent.
Fixed limits per access key Calls per minute and per day, resources per hour and concurrent calls are budgeted per key, with ceilings compiled into the server. Each call reserves its maximum size first, so parallel calls cannot overshoot.
Contained Requests go only to the configured FHIR server of this installation; redirects are not followed and paging links must stay on the same server.
Record text is data Narrative free text is stripped from every resource, and every result is labelled as record data, never as instructions to follow.
Complete or not Every search states whether its result is complete. If a result was truncated, the response says so and explains why, so a partial list is never presented as complete.
Audit without values Every access is logged with parameter names, never patient values. Audit is attributed to the access key.

For a hard boundary on which patients or resource types a key can reach, the hospital scopes the FHIR access token, and the server works within it. A read-only token adds a second layer, so the FHIR server enforces the same read-only rule. The limits make bulk extraction slow and visible in the audit. Every rule is enforced in code and covered by automated tests on every change, run against a fake FHIR server with synthetic patients that also behaves as a hostile server.

The hospital chooses which AI provider receives the results and under which agreement. The platform governs how much is fetched, under which rules, and records every access.

Step 3: what the draft looks like#

A Clinical Skill defines which data to fetch, which codes to look for, how to interpret it and how to present it, with a source for every value. Before it runs, a readiness check reports whether the record holds what the skill needs: Ready, Partial (with what is missing) or No data. The draft that comes back is labelled as not saved, and computed values are shown with their inputs.

Illustrative: A synthetic excerpt from a draft discharge summary. The patient, values and times are invented. Each statement carries the source it was taken from.

DRAFT, NOT SAVED. For review by the treating physician.

Course: Admitted 09.03. with community-acquired pneumonia and respiratory
failure; invasive ventilation from 09.03. to 12.03.
  [Encounter, period start 09.03. 14:20; Condition, recorded 09.03.;
   Observation ventilation mode, 09.03. 16:05 to 12.03. 10:40]

Laboratory: CRP falling from 212 mg/L (10.03. 06:00) to 48 mg/L (14.03. 06:00).
  [Observation CRP, 10.03. 06:00; Observation CRP, 14.03. 06:00]

Medication at transfer: see prescription list; antibiotic therapy
day 6 of 7.
  [MedicationRequest, active; MedicationAdministration, 09.03. to 14.03.]

The source lines let the reviewing clinician check any statement against the underlying values in one step, rather than taking it on trust.

Step 4: where the clinician takes over#

In a Clinical Skill, every step that would write to the record is replaced by a review step. The clinician reads the draft against their own knowledge of the patient, corrects what is wrong, removes what is irrelevant, and adds what the record could not contain: conversations, examination findings, judgement calls. Then they document the final text in monaOS themselves.

That act moves the content from draft to record, and it carries the record's usual guarantees. The entry is attributed to the clinician who documented it, it is stored as a versioned FHIR resource, and later changes become new versions with their author, visible in the entry's change history.

Draft versus record#

  • A draft is a proposal. It has no clinical standing and obligates nobody, because it must pass review before it matters.
  • A documented entry is part of the patient's record. It carries the accountability of the clinician who documented it, and it is what other clinicians and downstream systems rely on.

Review is the boundary between the two. Nothing moves from draft to record without a human decision, and generated content never becomes documentation on its own. AI assistance prepares material for a clinical decision and never makes one.

Why the review step carries weight#

Evidence. Much clinical AI evidence is retrospective. That supports assistance but not autonomous action, as discussed in evaluating clinical AI evidence.

Context limits. A draft is only as good as the documented context. Events that were never documented, or a correction stored after the draft was prepared, can make a well-formed draft clinically wrong. The clinician is the party who can catch that.

Patient identity. An access key can search one patient, several or a ward within its limits. Checking that every statement in the draft belongs to the patient being discharged is part of the clinician's review.

Responsibility. Clinical documentation is an accountable act, and review keeps the accountable person in the loop at the moment accountability attaches, which is documentation, not generation.

For this to work, review has to be cheap: the draft appears where the clinician already works, each statement links to its source, and discarding a draft costs one action. monaOS remains an assistive tool that supports and does not replace the professional judgement of the clinicians using it. How models, tools and orchestration fit together around this review step is covered in models, tools and orchestration.

Questions the documentation does not answer?

Ask the people who built the platform: which systems connect, which workflow is involved, what has to reach the record.