When the Scientist Is an Agent: Governing AI Before It Reaches the IND

Ono Pharmaceutical, founded in 1717 and one of Japan’s oldest drug makers, just handed every discovery scientist an AI agent. The company’s new collaboration with Phylo embeds Biomni Lab, an agentic AI platform, directly into its research workflows, extending automated hypothesis generation and experimental design across the discovery organization rather than piloting it in a single team (BioSpace).

That is a scale decision, not a pilot decision, and it lands in a part of the pharma value chain that no current regulatory instrument is built to reach.

The gap is upstream of the device rules

FDA’s device-side AI guidance and the EU AI Act’s high-risk tiering both govern AI once it is embedded in a product or a deployed system with defined users and defined risk. Discovery-stage agentic AI sits before that boundary. It is not a diagnostic algorithm and not a consumer wearable. It is a research tool that generates hypotheses, designs experiments, and increasingly makes autonomous decisions about which lines of inquiry to pursue. None of that activity currently triggers a formal regulatory submission. But its outputs eventually will, once a candidate compound derived from an agent’s hypothesis reaches an IND filing, a patent claim, or a clinical trial design.

This is the exposure regulated life sciences leaders need to underwrite now, before the tooling scales further. If an agentic platform selected or deprioritized a compound based on a model version that later gets updated or retired, can the company reconstruct that decision path for a regulator, an auditor, or opposing counsel in an IP dispute? Biomni Lab and platforms like it are not yet required to answer that question. That will not last.

Why ISO 42001 is the right scaffolding, now

The EU AI Act’s Article 50 labeling requirements, which take effect August 2 and require disclosure when people interact with AI systems (HR Executive), and the bloc’s parallel investment of $11.4 billion in AI gigafactory capacity (AP News), both signal that European regulators are building infrastructure and disclosure regimes for AI at a pace that will eventually catch up with R&D tooling, even if it has not yet named agentic discovery platforms specifically.

Life sciences companies deploying agentic AI at the bench should not wait for that catch-up. ISO 42001’s AI management system requirements, built for exactly this kind of organizational-scale AI deployment, give a defensible structure now: model versioning logs, documented human oversight checkpoints, and provenance records tying a research output back to the specific model version and prompt chain that produced it. That structure is far cheaper to build during rollout than to reconstruct retroactively when a regulator or licensing partner asks how a candidate compound was actually selected.

The decision in front of compliance leaders

Ono’s move is a template other large pharma organizations will follow this year. The question for compliance and R&D leadership is not whether to adopt agentic discovery tools. It is whether the provenance and audit infrastructure goes in at the same time as the agents, or gets bolted on after the first compound derived from an AI hypothesis reaches a regulatory filing. One of those paths is a governance program. The other is a discovery request.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. Core argument that discovery-stage AI creates audit/provenance obligations before regulatory triggers is coherent and defensible, but the claim that ISO 42001 is ‘the right scaffolding’ is asserted ra
Source & Claim Verification	Qwen · local	cleared. Most factual claims are supported by citations, but some lines, particularly those discussing the implications and future regulatory landscape, lack direct citations.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately identifies the regulatory gap for discovery-stage agentic AI and correctly positions ISO 42001 as a proactive governance framework, though it slightly overstates the immediate
Technical Accuracy	Llama	cleared. The article accurately describes the current regulatory gap for AI in pharmaceutical discovery and correctly identifies ISO 42001 as a relevant standard for managing AI risks.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies a gap in current regulatory frameworks and proposes a proactive solution, but could benefit from explicitly addressing potential counterarguments regarding the imme
Novelty & Non-Duplication	Grok	held. Hooks a commodity Ono-Phylo press release to a familiar pre-IND/regulatory-gap + ISO 42001 governance pitch that overlaps existing AI-in-life-sciences commentary and adjacent catalogue framing rather
Validation	DeepSeek	cleared. The central claim that discovery-stage agentic AI operates in a current regulatory gap is validated by the absence of specific FDA or EU AI Act rules for autonomous research tools, as evidenced by the

Sources cited: 12. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.