Voice control on a workstation showing a patient chart.

The risk is disclosure, not automation

A voice assistant on a clinical workstation is usually assessed as an automation risk - what if it clicks the wrong thing. That is the smaller problem, because a wrong click is visible and recoverable. The larger one is quiet: a tool that reads the screen to decide what to do has, by construction, read the chart, and whatever it sends to a model provider has left the covered entity.

So the design question is which bytes leave the workstation, under whose approval, and whether there is a record. Fluent answers it by routing every model request through a single controlled path that applies the egress allowlist and writes the audit record in the same step, so a blocked request becomes evidence rather than a silent failure.

On-device inference is what makes offline possible

In HIPAA mode, planning runs against a loopback model endpoint on the workstation instead of a hosted gateway. Because on-device inference is not egress, a hospital can run the whole product with no outbound model traffic at all - which is the only configuration some networks will ever approve.

Screenshots are treated separately and more strictly than text. Vision requests carry their own classification so they cannot be sent to an endpoint approved only for text, regions the detectors identify are masked before an image becomes a payload, and when masking cannot be completed the capture is refused and the run falls back to the accessibility tree rather than sending the original pixels.

• Planning on a loopback endpoint; no outbound model traffic required.
• Screenshot payloads classified separately from text so they cannot be misrouted.
• Masking failure refuses the capture instead of degrading to the raw image.
• Accessibility-tree reading preferred over screen capture generally - cheaper, more reliable, and it keeps pixels on the device.
• Telemetry is absent in HIPAA mode, and that specific rule is not delegable to a machine policy.

An audit trail you can verify, not just read

Every model call, every tool execution and every policy refusal is appended to a hash-chained log. Records are never rewritten or reordered, the chain can be verified independently of the application that wrote it, and it is checked against an anchor at start-up so that deleting the tail of the file is detectable - a chain alone cannot see a rollback, because what remains still verifies.

Values are passed through redaction before anything reaches disk. That is defence in depth and is deliberately not described as de-identification or Safe Harbor: no automated detector reliably finds a personal name, and a vendor claiming otherwise is telling you something about the vendor.

Ask any vendor the same two questions: can the audit record be verified by something other than your own software, and what happens when your redaction fails? An answer that treats the second case as impossible is the answer to worry about.

What is not certified, stated plainly

HIPAA has no product certification, so no software is "HIPAA certified" and any vendor using that phrase is describing marketing rather than a status. What exists is a covered entity's risk analysis, a business associate agreement where one is required, and evidence that controls work.

Fluent has a published control matrix, a data map, a CycloneDX software bill of materials with each release, and a business associate agreement template. It does not have a completed SOC 2 Type II report, and its accessibility conformance report is a self-assessment rather than a third-party audit. Both of those are stated here because a procurement office will find them anyway, and finding them after a claim is worse than reading them in the vendor's own document.

What an IT deployment actually looks like

Policy is deployed as a machine-level configuration file rather than a user setting, and it is resolved once per process - an assessor needs "the policy in force for this session" to be one auditable value that a running agent cannot see change underneath it. The floor fields an administrator sets cannot be raised by the person at the keyboard.

Two capability switches are worth setting explicitly on a clinical image: screen capture and agent network tools. Each is enforced at the single point every call path converges on, so a policy field is not merely reported in a panel - it refuses.

Questions

Keep reading

• Does Fluent have a VPAT or accessibility conformance report?A self-assessment that names its gaps is more useful than an audit-shaped document that claims everything. This is the first kind.
• What does Fluent collect, and what leaves my computer?Fluent listens to your microphone and reads your screen. That makes this policy unusually important, so it is written to be read rather than to be defensible.
• Can Fluent control Windows by voice?Say the outcome you want. Fluent prepares the Windows route and keeps its progress visible.

Sources

• HHS: HIPAA Security Rule — The regulation itself, and why product certification does not exist.
• Fluent accessibility conformance report — The self-assessment, including what it does not claim.
• Fluent privacy and safety disclosure — Input-by-input data flow and the current limits.
• Contact Fluent about a deployment — Security questionnaires, BAA, and the control matrix.