Not which tool to buy: AI agents across the pharma value chain

Two documents, opposite ends of the chain

BCG’s report makes the commercial case. Over ten years, it says, China’s biopharma innovation went through a “deep and systemic ‘efficiency revolution’”, built on scale, speed and cost. The next step is harder: turning that efficiency into “globally recognized innovation value — with the accompanying price tag”. AI appears as the lever on speed and cost. BCG’s framing is precise: the aim is to move the advantage “from lower point costs to a lower end-to-end cost of success”, by migrating trial and error upstream so that course corrections come earlier and cost less.

The policy direction comes from the Five-Year Plan (工信部联规〔2026〕210号). In its chapter on digital-intelligence transformation (数智转型), one of eight priority areas, it asks companies, hospitals and research institutes to build high-quality pharma datasets and a trusted health-data space, and to “accelerate the development of pharma-specific vertical models and agents” (加快研发医药垂类大模型与智能体). It names where they should be applied — drug discovery, clinical research, process optimisation, quality management, distribution and use, and smart regulation — and sets out five families of priority AI applications across that chain.

The plan does not promise funding. It sets out where the state intends to push, which in China’s pharmaceutical industry is usually where the regulatory and industrial effort goes next.

Two halves of one chain

The two documents weight different parts of the value chain.

BCG’s weight is upstream. Its AI section covers target identification, preclinical risk filtering and clinical development. It is candid about maturity: AI deployment at scale “remains in the early stages”, and the proven applications are still concentrated in small-molecule discovery. But it also says AI is expanding “from a point tool into a multimodal capability stack deployed across the R&D value chain.”

The plan’s weight is downstream and lateral. It adds full-process data capture and traceability in manufacturing, real-time control on critical quality attributes, computerised-system validation, prescription review, real-world evidence, and data infrastructure underneath all of it.

Put together, they run from the laboratory to the regulator. That is the core of our argument: an agent strategy in pharma is an organisation-wide change, not a tool purchase.

The regulator is building agents too

The line in the plan with the most practical consequences is easy to miss. One of the plan’s five families of priority AI applications is the regulation of production and distribution: review, approval and inspection “large models and agents” for drugs, cosmetics and medical devices. They would handle product classification, task allocation, dossier review, knowledge retrieval, issue identification and report generation.

For a company filing in China, this means that before the end of the decade, the first reader of a dossier may be a machine working to explicit rules. Inconsistencies between documents, versions that contradict each other, and data that cannot be traced back to its source are the things such a system is built to find. Good document discipline used to be a mark of quality. It is becoming a cost of entry.

Why quality is the place to start

Pharma has an advantage that most industries lack: in quality, the criteria are already written down, versioned and audited. An SOP is already a specification an agent can be tested against. A deviation-triage agent or a document-consistency check can be judged by the same standard a human reviewer is judged by, and the plan explicitly asks for validation services to support this.

That also explains why governance matters more than productivity for a healthcare buyer. In a regulated industry, work done by a machine cannot sit outside someone’s responsibility. Every agent output that reaches a batch record, a dossier or a safety report needs a named person who answers for it.

The question is organisational

China’s regulators had already said as much. The 2025–2030 implementation plan for the digital-intelligence transformation of the pharmaceutical industry, issued in April 2025 by seven departments, asked companies to make the top executive personally responsible for the transformation (“一把手”负责制), to set up a dedicated team, and to adapt their organisation structure and management systems to it.

So the useful question is not which tool to buy. It is who is accountable for the work an agent does, and what a job becomes when one person directs several agents. In our work we treat each seat as human, hybrid, agent-held or suspended. We write the protocol for returning it to human operation before the agent is deployed, not after something goes wrong.

What it means in each direction

For a European company operating in China, the plan describes the environment your Chinese peers, partners and regulator are being pushed towards. An organisation that runs on unconnected documents and personal know-how will find that the counterpart across the table, and eventually the reviewer, works differently.

For a Chinese company heading to Europe, BCG’s evidence points the same way. China’s share of originators in global BD deals rose from 17% in 2020 to 31% in 2025. Co-development accounted for nearly 40% of licensing deals between 2023 and 2025. Dual filings among leading companies have reached around 30%. BCG notes that multinational buyers are moving “from buying an opportunity point to plugging into a platform-based organizational capability.” A partner that can show how its agent-assisted work is controlled, traced and signed off offers a stronger platform to plug into.

Neither document tells a company how to organise itself. That part is still a management decision.