How We Use AI

Navigation and explanation. trellOS has more than twenty screens, around a hundred exception reasons and a dozen business rules, and most questions people ask are “where is this” or “why did that happen”. The assistant answers those.

It is not a clinical decision support tool, it does not give medical advice, and it plays no part in prescribing, dispensing or any decision about a patient. Nothing it says overrides a prescriber’s judgement or a pharmacist’s.

The assistant answers only from a knowledge index compiled at build time from the platform’s own source: the navigation registry, the business rules as they are written in code, and the product documentation. That index is committed to our repository as a reviewable file.

This is the whole guarantee, and it is worth being precise about why it is stronger than a filter: a corpus assembled from source files cannot contain a patient. There is no query the assistant could run that would return one, because there is no database connection in the code path at all. A runtime filter can have a gap. An absent capability cannot.

The build step enforces it rather than assuming it: it refuses to publish any entry that names an encrypted column, a patient-bearing table, a date of birth or anything resembling a personal identifier. When we first ran it, it held back fifty sections of our own engineering documentation for exactly those reasons.

Beyond that fixed knowledge, it can look up non-clinical reference data — the medication catalogue, product categories, pharmacy names, the exception reason list. Those are global reference tables carrying no patient link.

• It is checked before it goes anywhere. If a question looks like it is about a specific person — a name after “patient”, a date of birth, a record number — it is refused and no model call is made. You are pointed at the patient’s chart instead, which is where that question belongs and where the access is logged.
• It is never stored. We keep no record of what you typed. We count how often the assistant is used and which topics it matched, and nothing else. A question could contain sensitive text even after the check, so the safe thing is to keep none of it.
• It is sent with zero data retention. Requests go through the Vercel AI Gateway with zero data retention set on every call, which restricts routing to providers operating under a zero-retention agreement. Prompts and outputs are deleted on completion.
• It is never used for training. Zero data retention includes a prohibition on training. No trellOS data of any kind trains any model.

The assistant runs on Anthropic’s Claude, accessed through the Vercel AI Gateway. We chose the gateway deliberately: Vercel already hosts the platform and is already a contracted subprocessor, so routing through it adds no new party to the data path.

We will change models when a better one exists. What will not change is the boundary — any model we use receives the question and passages from our own documentation, and nothing else.

Language models produce fluent text whether or not they are correct, and confident wrong answers are the failure mode that matters. The assistant is instructed to answer only from the passages it retrieved and to say when it does not know, but that instruction reduces the risk rather than eliminating it.

Every answer shows the sources it drew on so you can check. Where an answer concerns a clinical or regulatory question, treat it as a pointer to the authoritative source, not as the source. The rules themselves live in the platform and in your own policies; the assistant only describes them.

To be explicit about the negative, because it is the question a compliance officer will ask:

• No AI decides whether a prescription is approved, routed or rejected.
• No AI sets a price, a quantity limit or a shipping tier.
• No AI reads, summarises or writes to a patient chart.
• No AI decides eligibility, credentialing or clinic onboarding.

The exception queue carries fields for machine classification, and today they are filled by a deterministic rule-based classifier — plain code with a fixed set of rules, not a language model. If that ever changes, this page changes first, and a human confirms anything clinical or compliance-related regardless of how confident a classifier is.

A clinic that would rather not have the assistant available to its staff can have it disabled for its organisation. Write to compliance@trellos.org.