Security & Trust, proven — not promised.

Customer content never trains Anthropic’s models — and we can tell you exactly what leaves our boundary.

AI features run on Anthropic's Claude under an executed Business Associate Addendum, through a HIPAA-configured API organization. Model training: customer content is contractually excluded from training Anthropic’s models, and RxAiAnalytics does not provide customer data to any third party to train models. Inside the platform (AWS), RxAiAnalytics runs its own statistical models on your claims: spend forecasts and member risk flags built from your own uploads, and — unless you turn it off for a client or an upload — cross-client learning, which stores only aggregate drug-to-drug transition counts with no member identifiers and uses them to calibrate the drug-progression risk flags of every participating upload. Both can be turned off per client or per upload; turning them off removes that upload’s results and its contribution. What we transmit to Anthropic is scoped per pathway and enforced in code. Analytics requests — the assistant, the claims context for contract analysis, and the plan-audit narrative — carry aggregates only (apart from the contract analysis and the optional member-level detail described below): counts, distinct-member counts, spend, and cost and utilization measures, grouped by drug and drug attributes, condition, tier, pharmacy channel and type, network, days supply, plan (as “Plan 1”, “Plan 2” …), prescriber specialty, age band, gender and month. Where the assistant’s detailed breakdowns are enabled, each breakdown group is reported with its exact counts, however small; high-cost members are referenced by positional rank (“Member #1”), and prescribers appear only as specialty with counts, never by NPI or name. No member identifier or tokenized member ID is included, and your organization's name, like client, plan and contract-file names, travels as a placeholder we substitute back locally. High-cost claimants listed in carrier reports are likewise referred to by position (“Claimant #1”), never by name or the carrier’s printed label. When you select contracts in an assistant conversation, their stored analysis is sent too, with names as placeholders; because it derives from your uploaded document, we treat it as a limited PHI disclosure under the BAA. When you continue a saved conversation, up to four of its earlier questions and answers are sent again as context. Member-level detail is off by default. It is available only in an assistant conversation, and only when it is allowed for the upload and, where the upload is filed under a client, for that client, and chosen by the user when starting the conversation — each re-checked on every question. The assistant can then read individual pharmacy fill rows for members of that one upload: the masked MBR ID the platform shows on its member screens, exact fill date, drug, NDC, GPI, therapeutic class, brand or generic, quantity, days supply, cost fields (minus any a field broker’s console masks), pharmacy channel, age band and gender — at most 10 members and 200 rows per lookup, 40 rows per member, and 3 lookups per question. That is a limited PHI disclosure under the BAA, not de-identified data. Those rows never include member names, the member or subscriber IDs in your files, Social Security numbers, dates of birth, addresses or contact details. Rows in restricted drug classes are withheld (and the assistant will not build a member list defined by one of those classes), web search is off for the whole conversation, every member whose rows are sent is recorded in the accounting of disclosures, and member-level conversation turns are deleted after 90 days. Schema detection is another pathway where that isn't the whole story, and we state it precisely: to map an uploaded file's columns we send a five-row sample in which a fail-closed classifier masks the values of every column it doesn't recognize as clinical, financial or operational — but real dates of service and provider NPIs are transmitted, because a format can only be detected from real values. Dates tied to an individual are identifiers under 45 CFR 164.514(b)(2)(i)(C), so we classify that pathway as an identifier-masked limited PHI disclosure under the BAA, not as Safe Harbor de-identified data. Documents you upload for analysis are transmitted in full so they can be analyzed securely, and may contain your organization's name. Free text is pattern-scrubbed for direct identifiers first, and PHI in free text is prohibited by our Terms. Outside member-level conversations, the assistant can also use Anthropic's web search, an in-scope covered service under the Addendum, where query content may be shared with third-party search providers. Every disclosure to the AI subprocessor is written to our §164.528 accounting of disclosures.