AI Privacy and Security

CuRE is designed to use AI in two ways: the AI bar, which answers questions about the product, and features built into specific workflows, such as drafting a document or suggesting a code. This page explains what each one sends, where it goes, what is kept, and what a person still decides. Your contract, data processing agreement and subprocessor list remain the binding terms

Where this stands

CuRE's AI is built and tested in development and is moving to production. The rest of this page describes the design that work follows, so read it as how AI in CuRE works by design, not as a report of what is running today. To find out what is enabled for your organization, ask us

The first production release is deliberately narrow. It sends a model only requests that carry no patient data, such as product help and form design. Features that read records, documents or clinical text stay off in production until Principia has a business associate agreement with the model provider, or the model runs inside the deployment's own perimeter

The AI bar

Product help answers how-to questions from the published CuRE documentation and cites the sections it used. If the documentation does not cover a question, it says so rather than guessing. Some surfaces return the closest documentation sections instead of a generated answer

The model receives your question, the documentation sections found for it and your language. It has no access to your organization, study or subject data. Earlier questions in the conversation stay in your browser, and the page you are on is used only to find the right documentation. Principia's gateway records which organization and user made each request, for audit and billing, and that record does not go to the model

AI in CuRE's workflows

AI also supports specific tasks in the apps: drafting documents and narratives, suggesting codes and classifications, extracting structured fields from unstructured documents, and answering questions grounded in a chart or a document set. Each feature sends the content its task needs, which can be the document, report or note being worked on. This is the main difference from the AI bar

These features produce drafts and suggestions, and a person reviews them in the app that owns the record. Fields extracted from a safety report, for example, stay suggestions until an operator confirms them. AI does not apply electronic signatures

There is one opt-in exception. A Capture study can allow low-risk, high-confidence data queries to receive an automatic response, recorded in the audit trail with its AI provenance. High-risk queries are never answered this way. Closing a query stays with a person unless the study opts in to that as well. Both settings are off by default

How a request reaches a model

No CuRE feature calls a model from your browser. In production, model requests pass through Clarion, Principia's AI gateway, which holds the provider keys so the apps do not. Clarion applies an automated de-identification scrub before text leaves for a provider, and it refuses attachments rather than sending them. If part of that scrub is unavailable, the request is blocked unless your organization has chosen otherwise

Clarion also checks each request's type against a short list of tasks that carry no patient data, and refuses everything else, until the business associate agreement described above is in place. The scrub does not admit a task to that list, because a scrub can miss an identifier

For question surfaces such as the AI bar, Clarion also refuses content declared as identified protected health information, whichever provider would serve it. That check depends on the declaration, so it does not replace the guidance below

Protected health information

Do not enter protected health information, or any patient or subject data, in the AI bar. It answers questions about the product and does not need them. The gateway's checks are a backstop for mistakes, and they cannot recognize everything typed into a free-text question

Some workflow features are built to read clinical text, including safety report intake in Canary, clinical note extraction in Conduct, chart questions in Cue and data queries in Capture. This text passes through the de-identification scrub first, but an automated scrub can miss an identifier, so it does not stand in for a business associate agreement. Principia has no business associate agreement with an AI model provider, so Clarion refuses these features in production. Talk to us if your work depends on them

Who processes it

A standard deployment uses the processors below. Your subprocessor list is the binding record

• Fly.io hosts the Clarion gateway
• Neon stores the gateway's audit records
• Anthropic and OpenAI provide the AI models. OpenAI also provides the embeddings used for search in standard deployments
• Amazon Web Services (Amazon Bedrock) and Microsoft (Azure OpenAI) serve models only in private-cloud deployments configured to use them
• On-premises deployments can use a customer-hosted model instead

What is kept

• In your browser: the conversation stays in the open AI bar and is not saved to your account. Your browser also remembers that you dismissed the first-use notice
• With your records: a draft or suggestion you accept becomes part of the record it was accepted into. Calculate keeps prompts typed into its query bar, and keeps the prompt and output of regulated drafts, such as statistical analysis plan sections, for audit. Capture stores numeric embeddings of protocol documents for search
• At Principia: the gateway's audit record keeps a SHA-256 hash of the prompt and the result, and the model used, not the text itself. We also keep usage counts for allowances and billing
• At model providers: as read on October 30, 2026, Anthropic's standard API retention is up to 30 days, and longer for content its safety systems flag. OpenAI retains API abuse-monitoring logs for up to 30 days

Model training

Principia does not train AI models on your questions or your organization's data. As read on October 30, 2026, Anthropic's and OpenAI's API terms do not use this content to train their models by default

Your organization's controls

• Administrators decide whether product help may draw on purchased credits
• In Capture, each study sets which AI tasks run, how they are deployed and whether automatic query responses are allowed
• Your organization's model selection is set with Principia, from the models Principia has approved
• A read-only governance console shows your organization its own AI audit evidence

Connecting outside AI tools

Clarion also offers a public connector for outside AI tools. It is separate from the gateway, holds no customer data and makes no model calls. Its tools work from what the caller supplies and from public reference material

What comes next

We are adding task assistants to the AI bar. In that design each assistant stays off until your organization enables it, and it receives only the context it has been granted

We are also designing saved memory, so a person can deliberately keep working knowledge for the assistant to reuse. It is not part of the AI bar today. In that design a person saves every entry, subject-level data is refused, administrators can turn it off, and saved content never trains a model. This page will change before either one ships

Questions

See the Trust Center for how the platform is secured, or contact us