Assistant (AI)

The Assistant is the first X7 Insight platform module: a context-aware AI helper over your study data. Permissions and audit are the same as for human users. The inference perimeter is set at deployment time: it can be fully local (the model is deployed with you — then data never leaves the system) or external, if so decided during rollout.

The golden rule

The Assistant suggests — you decide and act. It has no ability to enter or change clinical data, mark SDV, sign, unblind or close queries. Every change is a proposal that a human confirms and applies.

What it can (and cannot) do

It can:

• explain — what is left to fill in a form, why a check fired, what an open query means, how a version’s design is built (read-only, within your own rights);
• prepare drafts — prefill a form from an uploaded source document (OCR), draft a query/answer text, assemble a study skeleton from a protocol;
• execute with confirmation — on your explicit command, open a query or build a draft design version (only where level 2 is enabled).

It cannot (and never will — these are red lines):

• enter, change, auto-verify (SDV) or sign clinical data;
• work with unblinded data (it sees the ***BLINDED*** mask, like you without rights);
• advise on managing a specific patient (diagnosis, dose, treatment) — that is outside an EDC;
• fabricate data “to pass a check” — with no source it stops.

Autonomy levels

What the Assistant may do in your study is set by the administrator (level + feature toggles). The level cannot be “jumped” from the UI — it is part of change control.

1. Level 0 — Explain. Read-only chat: it answers from data you can see (the intersection of your rights and the module’s rights).

2. Level 1 — Prepare. Drafts are added: OCR form prefill, query text drafts, parsing a protocol into a design structure. You review and confirm every draft.

3. Level 2 — Execute with confirmation. On your explicit confirmation button the Assistant performs a prepared action (open a query, build a draft design version). Without confirmation nothing happens.

Where you meet it

• Chat dock (Assistant button) — questions about the study, subject or form. Answers link to objects (fields, queries, forms, subjects, visits) as chips. A conversation survives a page reload; History lists your earlier conversations, “+” starts a new one.
• Form panel — “ask about this form” and “fill from a document”: upload a scan/photo/PDF → extracted values appear as a suggestion card with confidence and a source snippet. You can correct a value right in the card (the decision is recorded as “edited”). You choose what to accept; accepted values are saved by your action (marked AI-assisted in the audit trail). “Reject all” is a decision too — it is journalled.
• Design builder panel — explain the design, generate a draft version from a protocol, and phrase an edit-check in words. The flow is: upload the protocol → the Assistant extracts its structure → you review and edit it → preview with warnings → build. The result is an ordinary draft that goes through the normal path: submit → review → approve by another person → publish.
• Settings (Admin) — level, enabled features, inference health and a quality dashboard (for the module administrator).

Verify-by-criticality

For critical fields (e.g. those flagged as requiring SDV) the suggestion card will not let you accept a value until you open the source snippet. “Accept all” does not apply to them. Low-confidence values arrive as “not sure” — with no value filled in.

AI literacy: working with AI responsibly

AI speeds up routine, but responsibility for the data stays with the human. Remember:

• Verify against the source. A suggestion is a hypothesis, not a fact. Check the value against the source document, especially for critical fields.
• Beware automation bias. The temptation to accept everything without looking is the main risk. Open the evidence; read drafts before sending.
• AI can make mistakes. The model is deterministic and works from the source, but recognition and extraction are not flawless; low legibility → lower confidence.
• It is AI, not a person. All AI surfaces are labelled; every interaction is logged (which model, which prompt) — part of the audit trail.

Your responsibilities by activity

• Data entry (CRC): accept prefill suggestions deliberately, checking against the source; open evidence for critical fields; the final record is yours.
• Monitoring (CRA): source-vs-data discrepancy hints are candidates only; SDV remains your manual act.
• Design (Builder): a generated design is a draft; you review forms/fields/visits and checks on the preview; the version is approved by another person (separation of duties).
• Data management (DM): you edit query-text drafts and send them under your own name; opening a query by the Assistant requires your confirmation.

Privacy and confidentiality

• Your conversations, uploaded documents and suggestions are visible only to you — even to colleagues who use the Assistant on the same study. An inspector reconstructs the document → suggestion → your decisions chain from the audit trail, not from someone’s chat.
• Uploaded documents may contain personal data — only upload sources you are authorised to process; they are retained per policy and then deleted.
• The Assistant does not see blinded data (***BLINDED*** mask), just like you without rights.
• The model does not train on your data under the recommended (local) deployment; with an external inference provider this is guaranteed by the contract with that provider.

Note

AI can make mistakes. Verify data against the source. The “this is AI” disclosure and logging of AI interactions are non-optional (EU AI Act, Art. 50).

What modules areHow modules are enabled and configured; capability consent.

Roles and accessHow grants and scope define what you and the Assistant can see.