Who reads the dossier now? Regulators are building AI agents

The line in the plan

Published on 18 September by ten Chinese government departments, the 15th Five-Year Plan for the pharmaceutical industry lists five families of priority AI applications. One of them concerns the regulation of production and distribution. It calls for review, approval and inspection “large models and agents” for drugs, cosmetics and medical devices, to handle product classification, task allocation, dossier review, knowledge retrieval, issue identification and report generation.

The plan did not invent this. China’s drug regulator had already written the rules six months earlier.

What the NMPA has already decided

The regulator’s implementation opinions on “AI + drug regulation” (国药监综〔2026〕6号), published on 2 April 2026, set two targets: AI applied across review and approval, inspection and testing by 2030, and a mature “smart” regulatory system by 2035. Two passages matter directly to a company that files in China.

The first concerns the dossier itself. The NMPA wants electronic submissions to become “standardised and structured” (申报资料电子提交标准化、结构化). A file that a machine can parse is becoming the expectation, not an option.

The second concerns how the regulator’s own machines are governed. Human–machine work is to follow three principles: AI support, human re-check, and a record of every step (数智赋能、人工复核、全程留痕). Keep that sentence in mind. It is also the best description we know of how a company should govern its own agents.

The other side of the table

The firms that write dossiers are moving too. On 8 July 2026, ProPharma announced an AI capability that generates structured first drafts of US generic-drug applications and helps manage regulators’ questions and deficiency responses. Freyr’s freya.RTQ drafts replies to health-authority questions from a firm’s past submissions. At the scale of a contract research organisation, IQVIA reports more than 150 agents deployed across internal teams and client environments.

So a dossier drafted with the help of an agent may soon be read first by another agent. Both ends of the exchange are changing at once.

What the numbers say, and what they do not

The figures in this field come from the organisations themselves. The FDA reports that more than 70% of its staff have used its generative AI tool; its first AI-assisted review pilot described some tasks going from three days to minutes. IQVIA’s 150 agents are its own count. None of these numbers has been independently audited, and the regulatory service firms publish no figures at all.

Taken one by one, each claim is weak. Taken together, they point in the same direction, and one detail stands out: the regulators publish more numbers than the firms that sell to them.

What changes for a regulatory affairs team

If the first reader is a machine working to explicit rules, the faults it is built to find become expensive: documents that contradict each other, versions that do not match, data that cannot be traced back to its source. Good document discipline used to be a mark of quality. It is becoming a cost of entry.

And the NMPA’s three principles apply in both directions. A company using agents to draft its submissions will be held, sooner or later, to the standard the regulator sets for itself: the machine supports, a named person re-checks, and every step leaves a record.

In our own practice

We run this model for one of our clients, a European medical-device company. Agents draft the answers to regulatory questions; a regulatory specialist reviews and signs off every answer before it leaves. The client renewed the arrangement. What made it work was not the model. It was deciding in advance who signs, on what criteria, and what happens when the agent is wrong.

Three questions to ask your team

• When an agent drafts part of a dossier or a query response, who signs it off, and is that written down?
• Is your dossier structured and consistent enough to be read by a machine working to explicit rules?
• Are people in your team already using AI on submissions outside any rule?

If one of these answers is unclear, the question is not which tool to buy. It is how the work is organised.