AI-assisted working has moved from conference talk to procurement question. Sponsors and CROs no longer ask whether their partners use AI, but where, under what controls, and with what accountability. The useful answer is not a technology list. It is a line, drawn precisely, between assistance and decision.
The line that matters. A GCP data workflow has a validated record at its centre. The eCRF data and its audit trail. The controlled documents in the trial master file. The statistical outputs that will support a submission. Everything that enters that record must be attributable, reviewed and reproducible. That is what the framework protects, and it is the right place to start when deciding where AI belongs. Drafting and
Where assistance earns its place. The appropriate uses share a shape. The output is a proposal, a person verifies it against source, and an error caught at review costs nothing but time. Four uses recur in practice. None of these writes to a validated system. All of them compress the mechanical part of work that a qualified person still owns and still signs.
Where AI must not sit. Some placements fail regardless of how capable the model is. The reasons are practical, not cautious. The audit trail must show an accountable person taking each action that touches the record; an AI write breaks that chain. And systems that make or record GCP-relevant decisions carry computerised-system validation expectations — reproducibility, change control, defined intended use — that non-det