Skovos

No learned component, and none is claimed. Skovos does not present itself as an AI product; it is the control layer around other vendors' AI agents. The agent registry, permission checks, hash chained audit trail, recall and signed evidence exports are rule based software, and the screen that rejects patient identifiers in audit records matches formats such as record numbers, phone numbers and dates.

The plain language feature runs through the customer's own Claude or ChatGPT, which reach Skovos through a connector, so the language model belongs to those providers. For a buyer, Skovos is bought as control over AI systems rather than as AI itself.

The boundaries are published; the failure behavior is not. Skovos does not act on clinical systems itself. Clinical, compliance and IT leaders set each agent's scopes, and granting access, expanding scopes and recalling agents require a separate operator credential. Connected Claude or ChatGPT clients cannot approve themselves or recall other agents, organization codes give read only observer access, and workload codes can only check permissions and report activity for one bound identity.

What is missing is what happens when a check cannot complete or denies an action partway through a workflow: whether an agent fails closed or open when Skovos is unreachable, and how a clinician overrides a denial that blocks needed work.

The building blocks are named; versioning is not. Audit records are hash chained, evidence exports are signed with a per organization Ed25519 key, access uses OAuth 2.1 with one hour tokens, records are partitioned by organization, and the service runs on Cloudflare Workers, D1 and KV. The connector uses the Model Context Protocol, with separate endpoints for Claude and ChatGPT, and each tool is documented, including which ones change data. No model sits in the product. The connector is in early access and the terms say it may change, but no version history or change log is published, so a buyer cannot tell when behavior changed.

A short chain with its parties named. Skovos runs no model of its own and is designed to keep patient data out of its records. The connector privacy policy names Cloudflare as the host and states that Claude and ChatGPT, when used to reach Skovos, process prompts and tool results under their own terms, which Skovos does not restate. The website privacy policy names MailerLite for the newsletter and Google Analytics for site visits, neither of which touches governance records. No subprocessor list is published for the main application.

No named deployment and no figure a reader can check. Skovos describes itself in its own comparison with Censinet as an early stage company, and no health system is named as a customer of the governance product. The home page logos, including Aledade, Palm Beach ACO and Doctor's ACO, sit under recognition and collaborators rather than customers and match the accountable care focus the company had as Haver AI; VBCare Network, also shown, is led by co-founder Anthony Rodgers.

The paid pilot aims to deliver an owner validated access map within two business days of receiving inputs, which the site labels a delivery target rather than a guarantee. Nothing is published on denial rates, recall timing or audit verification at a live site.

Categorical commitments and a retention position for the connector, with a conflicting website policy. Skovos states that connector data is used only to provide the service, is not sold or used for advertising, and does not train models. Records are kept until the organization asks for deletion, and because audit records are append only, deletion removes an organization's data as a whole. Evidence exports carry keyed commitments instead of agent names or content.

The website privacy policy, which covers platforms that link to it, allows de identified and aggregated data to be used for any purpose and collected information to improve products, so which terms reach the main application is not settled. Nothing is published on how a wrongly allowed or wrongly denied action is handled as a safety event.

A plain position that patient data stays out, with a service privacy document behind it. The connector privacy policy (13 October 2026) says the service is designed for governance metadata rather than patient data, tells customers not to send protected health information, and states that no business associate agreement is signed for the connector at this time.

Audit records matching common identifier formats are rejected, but names and addresses are not detected and the screen runs only on self reported activity. The company website privacy policy reads differently: it describes collecting electronic protected health information through affiliated healthcare providers and refers to a HIPAA notice of privacy practices, language that does not fit a governance product. Ask which document governs the main application.

Controls are described with nothing independent behind them. The security page sets out OAuth 2.1 with PKCE and one hour tokens, storage of access codes as SHA-256 hashes only, per organization partitioning, Ed25519 signed evidence, and TLS with provider encryption at rest, and gives a vulnerability reporting route with a two business day response aim. No SOC 2, HITRUST or ISO 27001 attestation, penetration test or trust portal is published, and the connector is in early access.

No device claim, and the product is scoped accordingly. The connector terms state that it records governance decisions and does not diagnose, treat or make clinical decisions. Skovos separately runs AI PSO, listed by the Agency for Healthcare Research and Quality as patient safety organization P0300 from 14 October 2026 under the Patient Safety and Quality Improvement Act; the company's own page says the listing applies to the PSO service and is not an endorsement of the software.

Its guides cover HTI-1 decision support transparency, the Section 1557 rule on patient care decision support tools and the Joint Commission's responsible use of AI certification as context for customers.

Evidence built for a governance committee to check the product's own controls, with no outside audit. Skovos exports signed, hash chained evidence that its controls ran, verifiable with a per organization key, and the audit trail can be recomputed to detect tampering. The evidence follows an OVERT v1.1 style profile that the company labels self declared and not a conformance claim.

With no model inside the product, bias and subgroup evaluation do not arise for Skovos itself; for the agents it governs, it records ownership and scopes but does not test their behavior or performance. AI PSO gives health systems a federally listed route to report and review AI incidents and near misses. No independent audit of the controls is published.

Mechanisms to check what happened, nothing standing behind the result. A tamper evident audit trail, integrity verification and signed evidence let a health system reconstruct what an agent was allowed and denied. The connector terms provide the service as is during early access, without warranties, and exclude liability for indirect or consequential damages.

No commitment covers a wrongly allowed or wrongly denied action, and no route is described for a patient or clinician affected by one. AI PSO offers incident reviews, but as a separate paid service rather than recourse against Skovos.

Clinical system connections are shown as examples with nothing to verify. The solutions page displays Epic, Cerner, athenahealth, Snowflake, Microsoft, AWS, Google Cloud and OpenAI as ecosystem examples and says supported connections and control coverage are confirmed per environment. No marketplace listing, integration guide or named production connection to a record system is published.

The documented integration is the governance connector for Claude and ChatGPT, which reaches Skovos's own registry and audit trail rather than the clinical record. Control over an agent depends on that agent checking through Skovos, or on an execution boundary integrated separately.

A single hosted service, isolation stated, location not committed. Skovos runs on Cloudflare Workers, D1 and KV with TLS in transit and the provider's encryption at rest, and partitions every record by organization so that no tool accepts an organization identifier as input. No region, residency commitment or customer hosted option is published. When governance staff use the connector from Claude or ChatGPT, prompts and tool results are processed by those providers under their own terms.

No price, and the way to buy is discoverable. A free trial can be started from the site, and paid work begins with a scoped pilot on one workflow: the buyer brings a sample or sanitized example, Skovos maps its agents, data access and accountable owner, then tests an allowed and a denied action in a safe environment. Scope, fee, access dependencies and acceptance criteria are agreed before work starts. AI PSO incident reviews are sold separately at a fixed fee for one provider, one event and one clinical workflow. No price, tier or unit of charge is published for either.

Stated clearly, with nothing validating it yet. Skovos is built for US health systems, hospitals, accountable care organizations and health plans, and frames permissions in clinical terms, with scopes such as reading vitals, drafting messages and writing orders. Its examples cover ambient scribes that draft orders, prior authorization agents, patient messaging agents and agents a health system builds itself on Claude or ChatGPT. Coverage is limited to agents connected to Skovos or checking through it, and supported connections are confirmed per environment. No live deployment in any of these settings is named.