Curbside Health

AI is an authoring and maintenance layer on a pathway platform whose value stands without it. The company's pages say AI structures guidelines, PDFs and policies into decision logic, proposes structure, tagging and cross-references, summarizes long documents for review, runs consistency and completeness checks, retrieves evidence for reviewers, and monitors sources to flag guidance that may be outdated; its governance page shows AI turning a 48-page guideline into branching logic for a named clinical owner to direct. The clinical logic clinicians see is authored and approved by people, and the company says AI never diagnoses, prescribes or acts on a patient.

Drafting decision logic from clinical prose and summarizing documents are functions whose working form is a trained language model, and the company refers to improving prompts, so the learned component is found. Remove it and the authoring, governance, EHR delivery and community sharing remain a complete product.

What the AI may and may not do is published, and every AI output routes to a named person before it can reach care. The responsible AI and philosophy pages list what AI does (structures documents into decision logic, summarizes, runs consistency and completeness checks, retrieves evidence, monitors sources and flags outdated guidance) and what it never does (diagnose, prescribe or make treatment decisions, act on a patient without a clinician, publish or update guidance without human approval, override clinical judgment). The route is published step by step: a named clinical owner directs the draft, two independent clinicians review every branch with comments resolved against the source, a multidisciplinary committee signs off and sets the effective date, and new evidence or a policy change reopens the pathway on a defined cadence as a new version. The company calls review a gate, not a suggestion; AI surfaces candidates for review, not silent edits.

At the point of care the clinician confirms eligibility and chooses, and the pathway is shown as explicit logic, so a clinician can disagree with a specific branch. The company does not describe how AI drafted passages are marked for reviewers, or how a clinician reports a branch that looks wrong in use.

The role of the AI is described in detail and the technology is not identified. The company lists what AI does in drafting, monitoring and review support, and refers to improving prompts, checks and controls, which points to a language model; no model, provider, version or update practice is named, and nothing describes how guideline prose becomes decision nodes or how AI suggestions are checked before reviewers see them. The decision engine and content framework are described as proprietary software.

The architecture is described and no party in the chain is named. The EHR launch reads scoped patient context over SMART on FHIR, the AI drafts and summarizes from guideline documents under prompts the company maintains, and the compliance statement lists vendor and subprocessor due diligence as a control; no hosting provider, model provider or subprocessor list is published.

Content also moves between institutions by design: health systems publish pathways to the community network and others adapt them, and the intellectual property page says nothing changes hands by default and customers keep ownership.

Named customers and a reach figure are published, and no result is. The home page says the platform is live in more than 40 health systems and shows Intermountain Health, University Hospitals Rainbow Babies and Children's, Children's Hospital Los Angeles, Nemours Children's Health and the American Academy of Pediatrics, with more than 1,000 pathway decisions modeled. The insights library argues that pathways fail at delivery rather than documentation and that most health systems underestimate the cost of variation two to three times, without a source; the variation figures on the clinical programs page are labeled fictional and illustrative.

No adherence change, variation reduction, outcome or time saving is published for any named customer, and nothing measures whether AI drafting saves authoring time or introduces errors. For a product whose case is that guidance fails at execution, an adherence figure with a denominator at a named site is the expected proof.

General assurances are published, and what reaches a model is not. The compliance statement says protected health information is encrypted in transit and at rest, with access logging, monitoring and periodic review; the privacy page summarizes principles (use only for disclosed purposes, least-privilege access, no sale of personal information) and says the full policy is provided during contracting. The intellectual property page says customers keep ownership of the clinical content they author, the company takes no rights to it, and nothing changes hands by default.

No retention period, training position or de-identification method is stated. The EHR launch reads patient context, and nothing says whether any patient data reaches the AI or its provider, or whether guideline documents uploaded for drafting are used to improve models. Safety engineering, meaning how a wrong AI draft or a misfiring pathway is caught after release, is not published beyond the review gate.

HIPAA alignment is claimed and the documents are held back. The compliance statement lists HIPAA-aligned safeguards and says the platform governs how protected health information is stored, encrypted, transmitted, accessed and logged; it adds that alignment does not imply certification and that current documentation is shared on request. No business associate status, agreement or scope is stated, and the privacy page is a summary, with the full policy provided during contracting.

The product reads patient context from the EHR, so a buyer will need a business associate agreement.

Controls are described and nothing independent stands behind them. The security page lists defense-in-depth architecture, role-based least-privilege access, encryption in transit and at rest, secure development with code review and dependency scanning, and continuous monitoring, and calls itself an illustrative summary with documentation on request. The compliance statement lists SOC 2-aligned controls and practices, then says alignment does not imply certification and that the company shares where it holds formal attestations on request. No certification, audit, penetration test or trust portal is named.

For a platform that reads patient context inside the EHR at enterprise health systems, a SOC 2 report is the document a vendor review will ask for.

No device claim is made, and no regulatory position is stated. The terms summary says the clinical content and tools are intended to inform and support qualified clinical decision-making and do not replace professional judgment, and the responsible AI page says the AI does not diagnose, prescribe, triage or act on a patient. The product delivers pathways authored and approved by the customer's own clinicians, with every branch linked to its source, which is the kind of reviewable basis that clinical decision support outside device regulation relies on.

The company does not state that reasoning. Its demonstration includes a time-critical chest pain pathway with cath lab activation, and the clinical decision support boundary is less settled for time-critical guidance.

Content governance is specified; model governance is a commitment without results. The responsible AI page says the company tests AI-assisted workflows for consistency and quality, watches for drift after release, and improves prompts, checks and controls as evidence changes; no method, test set, error rate or finding is published. Content governance is concrete by contrast: named clinical owners, two independent clinical reviewers, a multidisciplinary committee sign-off, version history with audit trails, a defined review cadence, and status showing what is live, outdated or under review.

The community layer adds a risk of its own. When institutions publish pathways to the network and others adapt them, the practice patterns and resource assumptions of the most active publishers can become defaults elsewhere, and a pathway that assumes staff or equipment a hospital lacks is not best practice there.

Mechanisms let reviewers catch a wrong draft, and nothing stands behind the output once approved. The company says no AI-assisted output is published, updated or surfaced at the point of care until a qualified clinical reviewer has approved it, links every AI-assisted recommendation to the guideline section it draws on, and versions and logs AI-assisted changes with their provenance; its governance workflow adds two independent reviewers and a committee sign-off before a pathway goes live.

Those are review and traceability controls. No drafting error rate is published, so the load the reviewers carry is unmeasured, and no warranty, remedy or route is described for a clinician or patient harmed by a pathway that went live with an error. The customer's own clinicians own and approve the content, which places responsibility with the health system.

Named record systems and a two-way model are claimed, with production unverified. The EHR integration page says Curbside works with Epic, Oracle Health, MEDITECH, athenahealth and other FHIR-capable systems, launches in context over a SMART on FHIR-ready architecture with OAuth 2.0 authorization, reads scoped patient context and writes back orders and notes inside the session; it says the platform is built with modern Epic environments in mind and demonstrates an embedded activity in a simulated Epic Hyperspace session. Pathway logic stays under clinical control after the initial integration, and the mobile apps read the same published version as the EHR launch.

The phrasing is ready rather than live, no FHIR resources or scopes are listed, no marketplace listing is cited, and no customer is tied to a production connection.

The delivery model is described and where the data rests is not. Curbside is a hosted platform that launches inside the EHR over SMART on FHIR with OAuth authorization, runs on the web, and ships iOS and Android apps with single sign-on and managed deployment through Intune, Jamf or Workspace ONE; downloaded pathways stay available offline in the app sandbox and are removed with the app. No hosting provider, region, data location or tenant separation is stated, and the customer base shown is in the United States.

No price is published, and how the product is sold can be pieced together. Buyers are named by role (health systems, clinical programs and service lines, professional societies, physician editors, education and training programs), deployment is enterprise (EHR launch, single sign-on through the customer's identity provider, managed mobile deployment), and every route leads to a demo request; security documentation and the full terms are provided on request or during contracting.

No rate, unit or tier appears.

Coverage is named by audience and content type, with named customers behind part of it. The company sells to health systems, clinical programs and service lines (oncology, trauma, pediatrics), professional societies, physician editors and education programs, across pathways, guidelines and protocols, policies and regulatory content, calculators, and antimicrobial stewardship with local antibiograms. It names Intermountain Health, University Hospitals Rainbow Babies and Children's, Children's Hospital Los Angeles and Nemours Children's Health, a concentration in children's hospitals that its demonstration pathways reflect. Those deployments support the health system and children's hospital settings; the professional society and education uses have no named customer, and no outcome is published for any setting.

It is listed for workforce and training because its education and training page sells the platform to academic medical centers, residency and fellowship programs, clinical education leaders and teaching hospitals, for onboarding and orientation and for keeping trainees and faculty on one current standard. No training outcome is published.