What the AI does and does not do

Professional-sounding wording is not proof that a draft is correct. The checks described below catch text the facts do not support; they do not check clinical judgement. That is the clinician's job, and the signed note is the clinician's responsibility. See Clinical safety and accountability.

Several Rehab features look intelligent and are plain rules. The difference matters: rules behave the same way every time and can be checked line by line.

How to use each one: AI drafts in Rehab, worked examples and Visit recording and dictation.

AI drafts in Rehab are checked against the facts they were given before you see them.

• Every sentence in a visit-note draft must name its source: today's visit notes, your assessment or plan, goals, the plan of care, precautions, confirmed findings, the previous signed note, and so on. Subjective and Objective may cite only what was recorded for today's visit.
• A sentence with a number, side, assistance level or unit that is not in its source, or that drops a "not" or a "possibly", is removed. You are told how many sentences were removed and why.
• A section with no usable source is left empty rather than filled in. In a progress report or discharge summary, a section that fails the checks is left blank with the reason.
• Suggestions from outside records quote the document word for word.

These checks catch invented facts. They do not catch a draft that is accurate to its sources and still clinically wrong.

• It does not diagnose. A diagnosis in a draft comes only from the diagnosis already documented.
• It does not send anything: no message, letter, claim or referral leaves ArkCloud because the AI wrote it.
• It does not fill a gap with a plausible guess. Missing information stays missing.
• It does not score or interpret outcome measures. Scores are what the clinician enters.

On the hosted service, AI runs on a separate on-premises model server that ArkCloud operates on its own private network, reached over an encrypted (TLS) connection. It is not a third-party AI cloud. The software sends patient information only to a model on the same machine or to a private-network host an operator has explicitly acknowledged, over TLS; any other endpoint is refused, and no patient information is sent to an external AI provider.

One feature is the exception, and it carries no patient information: exercise illustrations for home programmes are drawn by OpenAI's image service from the library exercise's description alone - its positions, steps, side, equipment and camera view - never from a chart. Text that looks like a name, date or record number is refused before anything is sent, and every illustration is approved by a clinician before a patient can see it.

Each feature checks your role and your organisation's AI switch first, then reads only the minimum facts it needs, inside your organisation. The visit-note draft receives what you recorded for the visit, up to six active goals, the plan of care, the documented diagnosis, precautions, cognitive status, confirmed findings and earlier measurements labelled with their dates, today's recorded services, the home exercise programme, outcome scores and the previous signed note's assessment and plan. It receives no name, date of birth, record number, address, payer or charge amounts.

AI requests are recorded in the audit trail with the user, the record and the kind of request, never the text of the question, the facts or the answer.

An administrator can switch every AI feature off for the organisation under Settings → AI. Rule-based features, including guided questions, keep working. Visit recording has its own organisation setting.

A clinic can also install the patient AI agreement, so each patient chooses whether AI helps draft their visit notes and whether their visits may be recorded. See The patient AI agreement.

RCM, Access, Care, Insights and Workforce use the same model and the same rules: drafts and explanations a person reviews. The list below shows the features switched on for the hosted service, and whether each uses AI or fixed rules.

More detail: AI drafts in the revenue cycle, AI assists on your screens, referral summary and Care coordination summary.

• One small model, shared by every AI feature, limits answer quality and how many requests run at once. When it is busy or unavailable, the screen says so and keeps working without it.
• Each AI feature has an hourly limit per person, so a burst of requests can be refused for a while; write the note as usual meanwhile.
• The AI features have not been clinically validated, and drafts were tested on synthetic data.
• Questions typed to the Clinical Brain or the help assistant are sent as typed; do not include patient details the question does not need.