Healthcare workflow intelligence for fragmented care operations.
MahCare AI is an AI-native healthcare intelligence and orchestration platform for ambulatory, community, home, pharmacy-linked and hybrid care — unifying care execution, patient engagement, a governed AI workforce, predictive risk and an audit-ready evidence trail on one core with UK and US operating profiles. It catches the patient slipping before it becomes a crisis: turning referrals, plans, alerts and visits into managed work prioritised by predicted risk, where a qualified clinician always makes the final call.
Early access — selecting UK/US pilot partners. No autonomous prescribing, diagnosis, or silent writes to the clinical record: consequential actions always pause for human review.
What MahCare AI does
The working surface — what the engine produces, controls and proves in practice.
Care Graph — a connected operational-clinical-commercial substrate
A longitudinal graph linking patients, episodes, tasks, medications, messages, documents, risks and outcomes — not siloed EHR tables. AI workers reason over the full patient journey (missed message, omitted dose, high-risk observation, owning team, open approval) instead of isolated text chunks, with grounded context for fewer hallucinations.
AgentOS — a governed AI workforce of named workers
Task-specific agents including Referral Intake, Patient Concierge, Documentation, Care Coordination, Medication Adherence, Prior-Auth, Coding & Revenue, Inbox Triage, Quality Auditor, Executive Analyst, plus RPM Monitor, Virtual Ward and a Safety Reviewer that adversarially reviews clinically-influential outputs. Each has a defined job, a hard boundary and measured performance.
Adaptive autonomy with human-review gates
Every agent runs at a configurable autonomy level — from Suggest (no PHI leaves the regulated boundary) through Draft (human edits and signs) and Propose (human accepts before it takes effect, where most agents sit) to bounded reversible execution within tenant guardrails. Autonomy, model routing and review thresholds are set per tenant in Studio with no code.
Calibrated prediction and risk, with honest uncertainty
Predictive models flag deterioration, no-show risk, medication non-adherence, rising care needs, escalation risk and readmission/crisis risk before they become emergencies. Every score arrives with a plain confidence band — a strong signal versus an early hint — never a bare number dressed up as certainty. The system informs; a clinician decides every consequential step.
Hard safety boundaries enforced at platform level
Four rules the AI workforce never breaks: no AI autonomously prescribes, discontinues or alters medication; no AI silently writes to the legal clinical record; clinically-influential outputs require human review before action; all submissions to external parties remain human-accountable. These cannot be overridden by tenant policy or agent behaviour.
Evidence Ledger — compliance productised, not bolted on
A tamper-evident record of what happened, who did it, with what permission and under which policy — covering administrative actions, sign-ins, policy diffs with approvers, AI decisions (model, prompt version, evidence, reviewer, effects), workflow transitions and patient-data access. Entries are locked together so any after-the-fact change shows immediately; auditors can verify independently without application access.
Studio, Marketplace and Benchmark Network
No-code/low-code configuration of pathways, forms, roles, policies, automations, prompts and integration mappings; a Marketplace for apps, connectors, templates and pathway packs; and an opt-in, de-identified Benchmark Network for cross-tenant operational and pathway comparisons under strict governance and consent controls.
Standards-based interoperability and developer platform
FHIR R4-aligned canonical model, OpenAPI 3.1 contracts with deterministic SDKs, signed webhooks and replayable event streams, SSO/SCIM and SMART/OAuth, CSV/NDJSON/FHIR-bundle import-export with row-level error reporting, and a developer sandbox with tenant-scoped API keys, webhook replay and synthetic data.
A controlled, evidence-led flow
Ground in the Care Graph
Patients, episodes, tasks, medications, messages, documents, risks and outcomes are connected into one longitudinal graph. AI workers reason over this grounded context — the full patient picture — rather than isolated records or raw text.
AI proposes within governed autonomy
Named agents draft, prioritise, route and predict at a configured autonomy level. Consequential actions stop at human-review gates; a Safety Reviewer adversarially checks clinically-influential outputs, and PHI is minimised and redacted before any model invocation.
A human reviews and the Evidence Ledger records everything
A clinician or operator approves or rejects. Every AI action stores model, provider, version, prompt template, input references, decision path, reviewer and timestamp to a tamper-evident ledger that can be verified independently and exported.
Outcomes feed back as a compounding loop
Agents are measured on closed-loop outcomes — contact success, overdue reduction, time saved, override and false-alert rates — not output volume. Workflow and AI policy changes can be simulated against historical event streams before release.
What it's used for
Referral to first contact
Capture referrals from forms, portals, APIs, CSV and call-centre workflows; AI-assisted extraction and triage suggestions; eligibility, coverage and service-area checks; waitlist ranking and automatic refill — reducing time-to-first-contact and unworked referrals. AI never silently accepts or rejects a referral.
Care-plan execution and work orchestration
Versioned care-plan templates and pathways emit tasks, forms, reminders and alerts into a unified work inbox with pull/push/round-robin/priority/skill-based queues, SLA policies with breach prediction and escalation ladders — driving higher pathway adherence and fewer overdue tasks.
Medication coordination with safety boundaries
Reconciliation at referral, start-of-care, review, transfer and discharge; administration, omission, refusal and delay capture with reason codes; refill, adherence and pharmacy follow-up tasking using dm+d (UK) and RxNorm/NDC (US) terminology — for fewer medication-related misses. No AI alters a regimen without human confirmation.
Deterioration, RPM and virtual wards
Ingest remote-monitoring streams and patient-reported inputs; evaluate thresholds, deterioration and trend rules; RPM Monitor and Virtual Ward agents propose who to escalate, visit or step down — with the reasoning — for faster time-to-acknowledge and time-to-action, with mandatory human review on clinically-influential cues.
Prior auth and revenue prep
Prior-auth workspaces assemble evidence, complete checklists, draft medical-necessity narratives and track submission status; Coding & Revenue and Revenue Leakage agents propose ranked, evidence-cited codes and flag leakage across schedules, charges, coverage and denials — for faster submission turnaround. Submissions remain human-accountable.
Compliance response and patient engagement
DSAR/SAR search-compile-redact-approve-export, legal holds, access reviews, break-glass and incident/CAPA workflows turn compliance from days into hours; multi-channel engagement (app, SMS, email, voice, letter) with preference centre, reply triage and proxy handling lifts digital delivery and response rates.
What you receive — and how it's proven
Every job ships with an evidence record, not just an output.
What you receive
- Coordinated, tracked care workflows from referral to discharge
- Governed AI-worker output (drafts, triage, predictions) gated by human review
- Calibrated risk flags with honest confidence bands — deterioration, no-show, non-adherence, escalation, readmission
- Patient communications across app, SMS, email, voice and letter
- Prior-auth packets, evidence-cited coding suggestions and revenue-leakage flags
- Tamper-evident Evidence Ledger entry on every action, independently verifiable
- Automated evidence packs, DSAR exports and access reviews
- Operational, clinical, commercial and AI-performance dashboards and board-ready operating packs
Evidence & proof
- Mandatory provenance on every AI action: model, provider, version, prompt template, inputs, decision path, reviewer and timestamp
- Human-review gates on all consequential and clinically-influential actions; four hard safety boundaries enforced at platform level
- Tamper-evident Evidence Ledger, independently verifiable without application access and exportable in standard formats
- Reliability targets: 99.95% core API/admin and audit-write SLO, 99.9% portal/mobile; RPO ≤ 15 min, RTO ≤ 4 hours; quarterly restore tests with evidence (figures are targets, subject to contract)
- Encryption in transit everywhere and at rest for databases, files, backups and search indexes, with managed key rotation and audit
- AI agents measured on closed-loop outcomes — time saved, approval/override rates, false-alert rates — via evaluation harnesses, not output volume
Run it your way
Common questions
Is MahCare AI certified?
No — it's an early-access product and it isn't certified, and we won't imply otherwise. It is built to the principles that matter in care settings from day one: encryption in transit and at rest, role-based access control, comprehensive audit logging, the tamper-evident Evidence Ledger, and hard safety boundaries on AI. The evidence trail is designed to support your own information-governance review.
Does AI replace clinicians in MahCare AI?
No. AI workers assist, not replace. Every agent runs at a configurable autonomy level with human-review gates, and four hard safety boundaries are enforced at platform level: no autonomous prescribing or medication changes, no silent writes to the legal clinical record, mandatory human review of clinically-influential outputs, and human accountability for all external submissions. A qualified clinician makes the final call on every consequential step.
What is the Care Graph and how does it differ from an EHR data model?
The Care Graph is a longitudinal operational-clinical-commercial graph connecting patients, episodes, tasks, medications, messages, documents, risks and outcomes into one fabric. Unlike siloed EHR tables, it lets AI workers reason over the full patient journey with grounded context — which improves relevance and reduces hallucination compared with stand-alone copilots working over raw text.
How do AI autonomy levels work?
Each agent operates at a level you configure in Studio: Suggest (no PHI leaves the regulated boundary), Draft (a human edits and signs), Propose (a human accepts before it takes effect — where most agents sit), and bounded reversible execution within tenant guardrails with operator override. Potentially-regulated functions are not enabled by default and require a separate clinical-safety and regulatory workstream.
Is MahCare AI available in the UK and US, and where is data stored?
It is one core with UK and US country packs — no product forks. UK deployments are built around NHS login, NHS Notify, dm+d and DTAC/DSPT evidence; US deployments around HIPAA Security Rule principles, US Core / SMART on FHIR, NPI and RxNorm/NDC. It supports UK/EU and US regional data residency, with production data region-separated; dedicated single-tenant environments are available for isolation requirements. DCB0129/DCB0160 clinical safety requires a customer-side Clinical Safety Officer; we provide the platform evidence to support it.
How does the Evidence Ledger support audits and DSARs?
Every action — administrative changes, sign-ins, policy diffs with approvers, AI decisions, workflow transitions and patient-data access — is written to a tamper-evident ledger whose entries are locked together, so any after-the-fact change is visible. Built-in workflows cover DSAR/SAR search-compile-redact-approve-export, legal holds, retention scheduling, break-glass and incident packs, and records can be verified independently and exported in standard formats.
How does MahCare AI relate to ITLOX's other products?
RadMah Sentinel can wrap healthcare AI workflows with runtime governance and audit-ready evidence; RadMah AI generates synthetic FHIR data for safe testing before real patient data (example benchmark: 95.69% overall on the public mostlyai-qa benchmark — Adult, 24K rows, 200 epochs — results vary by dataset, not a guarantee). AegisWire can provide secure transport for healthcare networks; it is hybrid post-quantum (X25519 + ML-KEM-768) designed to resist harvest-now-decrypt-later, never described as quantum-proof, and its hardware appliance is on the roadmap, not available now.
How MahCare AI relates to the rest of ITLOX
RadMah Sentinel
Evidence and review gates around healthcare AI workflows.
RadMah AI
Synthetic FHIR data for safe testing before real patient data.
Next step