The short version
(Updated 5 October 2026.) ArkCloud Health uses AI wherever a model can save a professional time without taking a decision away from them, and only in three shapes: an answer to a question, a plain-language explanation of what a screen already shows, or a draft a person edits. Every AI panel is labelled AI before it is opened, and nothing it writes reaches a record, a payer or a patient without a person's action. Each application now has some assistive AI; the AI page lists every feature and its status. Claim checks, denial sorting, risk tiers and every dashboard figure still run on written rules, and we label them that way.
This article explains how those pieces fit across a patient's journey, from referral to paid claim, for executives, clinicians, rehabilitation providers and revenue-cycle teams. You can explore each feature, with demonstrations on synthetic data, on AI Across ArkCloud Health.
Why we keep AI narrow
Healthcare software earns trust by being predictable. A clinician needs to know why a note was held from billing; a billing manager needs to know why a claim was stopped. Rules give the same answer every time and can explain themselves, so we use them wherever the decision is a matter of policy.
We use a language model where the work is language: turning facts into a readable answer, or turning rough visit notes into a structured draft. Even there, rules decide what the model is allowed to see, and a person decides what happens next.
Where AI sits in the patient journey
Referral and front door
ArkCloud Access handles referrals, registration, intake, eligibility and scheduling. Its decisions stay with staff: booking requests are approved by people, eligibility is the standard electronic transaction with the payer once a clearinghouse is connected, and nothing is matched or booked automatically. AI helps people read: a referral summary in which every line quotes the referral or an attached record word for word (with a missing-information checklist built by rules, not AI), a summary of a returned intake form that ends each point with the patient's own words, a plain reading of the payer's eligibility response, and an authorization request draft for utilization review to edit and submit.
Evaluation, visits and documentation
ArkCloud Rehab is where AI does its most useful work today.
The Clinical Brain (live). On an episode or assessment page, a therapist can ask questions such as "what was the latest score?" or "what is still outstanding on this episode?". Rules match the question to one of eight approved types and fetch only the facts that type allows. The model turns those facts into a sentence, and a rule-based check compares the answer with the facts and withholds it if it says anything they do not support. A question outside the approved types - "should I increase the load?" - is declined, because treatment decisions belong to the therapist. The Clinical Brain cannot change the record.
The Clinical Brain also prepares a short handoff summary for a covering clinician, from the same recorded facts, and never says whether the patient improved.
The Draft SOAP note (live). From the therapist's own typed visit notes, active goals and the previous note's assessment and plan, the assistant proposes a Subjective, Objective, Assessment and Plan draft. It never writes into the note: the therapist inserts the sections they accept and edits them. Inserting AI text marks the note AI-assisted, and an AI-assisted note cannot be signed until the therapist attests they reviewed it. Progress reports and discharge summaries get the same treatment for their narrative sections, written only from recorded visit counts, goal statuses, published-measure scores and signed notes - never new goals or a plan of care.
Smaller aids sit beside the work: suggestions from an imported outside record that quote it word for word, a plain-language explanation of a published outcome measure (not of the patient's score), reply drafts for patient messages (a fixed word rule routes possibly urgent messages to a clinician instead), and a warmer rewording of missed-visit outreach that staff send themselves.
Outcome scoring, progress-report prefill and billing holds in Rehab are rules, not AI. They work the same way every time and show their reasons. ArkCloud does not yet interpret a score against published thresholds; that is left to the clinician.
Billing and the revenue cycle
The account assistant (live). In ArkCloud RCM, billing staff can open the AI tab on a patient account and ask for a summary, or what to check before a call. The assistant sees the account's balance and ageing, its recent claims with their status and open denial codes, what payers assigned to the patient, any payment plan, the outcomes of recent contacts and the latest account notes. It does not see the patient's name, diagnoses or clinical notes, so it does not offer coding advice - and it cannot post, adjust or contact anyone. Administrators can switch it off for their organisation.
Drafts and explanations (live). Create Appeal on a denial opens with the letter drafted from the records listed beside it; the Script tab tailors the balance discussion before a call; a patient's dispute opens with a plain-language reply explaining the balance; and claim-check findings, remittance posting exceptions and the start of a work queue are explained in plain words. Every amount, date and code in a draft is checked against the records it came from, and a draft that says anything else is withheld. Staff edit and send every one.
Admin agents (released, off until enabled). In ArkCloud Rehab, four agents propose work for a person to approve. Two of them - the appeal drafter and authorisation prep - ask the model to word a letter from the claim and authorisation facts that rules collect. A draft that mentions a code, amount or date not in those facts is discarded and a rule-built letter is used instead, and the patient's name, date of birth and member ID are added by the software, not sent to the model. Every proposal waits for approval; nothing is submitted, sent or signed by an agent. Each agent is switched off until an administrator enables it.
Claim checks before submission, denial categorisation and collections escalation are rules. A claim with a blocking finding is stopped with the reasons listed; that is automation you can audit line by line.
Between visits, staffing and reporting
Care's risk tiers and care gaps are rules your own clinicians write and review, each finding with its reason. AI adds a coordination summary on the patient screen, with up to four proposed follow-up tasks a coordinator accepts one by one, and a plain explanation of each published gap rule. Care plans and pathways are never written by AI.
Insights calculates financial, payer, denial, clinical and quality figures, benchmarks and a cash forecast with fixed formulas over your own records; the forecast is a rule-based projection, not a trained model. Each Insights screen has an Explain with AI panel that describes the totals and rates already on that screen in plain sentences, and a short denial brief sits under the denial figures; both see aggregates only, and text containing any number the screen does not show is withheld.
Workforce schedules and credentials are managed by people, and an expired licence blocks booking because a rule says so. AI writes a short brief on credentials that need attention, with the shift overlap worked out by the software; it never decides who works.
Where the model runs and what it sees
On the hosted service, every ArkCloud AI feature uses a small open model, Gemma 3 (4B), running on a separate on-premises model server ArkCloud operates on its own private network, reached over an encrypted (TLS) connection - not a third-party AI cloud. Before any patient information is read, the software checks where it would go and refuses any endpoint that is not on the same machine or an explicitly acknowledged private-network host over an encrypted link. Patient information is not sent to an external AI provider. Only the facts a task needs are sent: names, dates of birth and member IDs are left out wherever the text does not need them, and added to letters by the software afterwards.
Each feature reads only inside the signed-in user's organisation, after the same permission check as the screen it sits on, and has an hourly limit per person and per organisation. AI queries are recorded as audit events - who, which record, which model and prompt version - without storing the question or the answer. For the Draft SOAP note, the draft, its acceptance and the signing of an AI-assisted note are chained into the tamper-evident audit log.
The technical documentation covers the data flow, configuration and known limits in more detail.
Assistance, not autonomy
Across the platform, AI in ArkCloud does not diagnose, order, prescribe, write a plan of care, a care plan or a goal, suggest a billing or diagnosis code, score a patient's risk, book an appointment, sign, submit a claim, post a payment or contact a patient. It answers questions from governed facts, explains what a screen shows and proposes drafts. People make the decisions, through the same screens and permissions they use for everything else.
We also say plainly what we have not done: these features have not been clinically validated, and one small model on one model server has real limits on answer quality and capacity. When it is busy or unavailable, each screen says so and keeps working without it. We would rather tell you that than imply otherwise.
What this means for your organisation
- Executives get AI whose footprint is known: named features, each listed with its status, all advisory or approval-gated, all on ArkCloud's own private network.
- Clinicians keep authorship of the record. Nothing reaches a signed note without their choice and attestation.
- Rehabilitation providers get help with the questions that slow a session down, without a model making treatment decisions.
- Revenue-cycle teams get a quick read of an account, first drafts of appeals and replies, and dependable, explainable rules for claims and denials.
To see each capability, what it receives and produces, and what a professional reviews, visit AI Across ArkCloud Health, or read the module guides for ArkCloud Rehab and ArkCloud RCM.
