● The role of AI

AI is the interface and the analyst — not the data owner

The AI’s job is narrow and supervised: understand a plain-English question, choose one of our pre-built governed tools to fetch the data, then format and analyse what comes back. It never holds the data, changes it, or decides what’s valid.

Enterprise-licensed, contractual AI

📜

Licensed & contractual

We use an enterprise-licensed large-language-model API under a commercial agreement — not a consumer chatbot.

🚫

No training on our data

Under the enterprise terms, prompts and data are not used to train the model. It’s used for inference only.

🧱

No autonomous framework

No LangChain / AutoGPT-style agent. We wrote the orchestration ourselves, so every AI action is one of our validated, logged tools.

What the AI does — and does not — do

✓ The AI does

  • Interpret the question and design the study (cohorts, endpoints).
  • Choose a governed tool — browse schema, build a cohort, run a read-only query, draw a chart.
  • Format results into clear tables and charts.
  • Explain the answer in plain language, with caveats (confounding, missingness).
  • Show its work — the exact query it ran is visible.

✗ The AI does not

  • Touch raw data directly — only via governed, read-only tools.
  • Change, write, or delete any data.
  • Decide what is “valid” — a human reviewer does.
  • Move data outside the governed perimeter.
  • Get trained or fine-tuned on patient data.
One sentence

The AI formats and analyses the data the platform retrieves — it is an assistant on top of a governed data layer, not a pathway to the data itself.

Why this is safe

🧰

Tools, not free access

The AI can only call a fixed set of pre-approved tools. It cannot invent new ways to reach data.

🔒

Read-only guard underneath

Even the query tool is forced through a SELECT-only guard — the AI cannot mutate data even if asked.

📋

Every call logged

Each AI tool call, model, and prompt version is recorded — fully reproducible and reviewable.

Enterprise-licensed AI · inference only · no training on data · no autonomous framework