CAS's AI chief says trusted data and lab hardware, not model skill, now hold science AI back

Big promises, a different bottleneck

AI leaders have set high expectations for science. As R&D World reported, Anthropic's Dario Amodei has argued that powerful AI could compress 50 to 100 years of biological progress into five to 10, and Google DeepMind's Demis Hassabis has said AI could potentially help cure all disease.

Whether the money flowing into AI pays off will depend partly on labs turning AI-driven discoveries into products such as new drugs and materials. That is where the article's main interview subject comes in: Andrea Jacobs, senior director of AI at CAS, a scientific information provider. She trained in chemistry and computer science and has worked at CAS since 2010.

Her argument is that the models themselves are no longer the weak point. On GPQA Diamond, a set of 198 graduate-level physics, chemistry and biology questions, Google DeepMind's Gemini 3 Pro scored about 93% in Epoch AI's testing after its November 2025 release. Epoch's leaderboard now lists Gemini 4 Argon at 96.4%.

Putting a science reference inside Copilot

CAS's answer to the trust problem is to meet researchers where they already work. On Oct. 20 it announced an integration with Microsoft Copilot, the assistant built into software such as Word and PowerPoint. Users get CAS's curated scientific content, plus CAS Newton, an AI agent that uses CAS data and tools to handle complex science questions.

The link runs through the Model Context Protocol, or MCP, a standard way to connect AI systems to outside tools and data. Copilot can pull facts straight from CAS, such as a compound's properties, or hand harder questions to Newton, which can combine several kinds of CAS content in one analysis.

Jacobs pointed to Copilot's reach as the appeal. Many people already live in Microsoft products, so the science reference arrives inside tools they open every day.

Why names and characters matter

Chemistry is unforgiving about naming. The pain reliever acetaminophen is also called paracetamol, and the asthma drug albuterol is also called salbutamol. CAS gives each substance one Registry Number whatever it is called, so an agent has a single identifier to check.

Small errors can also be costly. Jacobs noted that a single character in chemical notation can separate two very different compounds, which has safety consequences for scientists and accuracy consequences for their work.

A related feature, CAS Connections, links parts of an AI-generated answer back to CAS source material when that material contributed. Jacobs said this helps people who handle science without being deep experts, such as business development or project management staff. Without it, she said, checking an answer takes a lot of effort and background knowledge.

CAS Connections launched in June 2026 with several initial collaborators. Beyond Copilot, it reaches Claude and ChatGPT Enterprise through MCP. It also plugs into electronic lab notebooks and lab information management systems, where the integrations are narrower, such as looking up a Registry Number or finding reference papers and reactions.

From software to the physical lab

Jacobs expects the scientist's role to shift as agents take on more work. Planning experiments and anticipating bench results give way to directing software. In her words, "It becomes much more of a management context."

The second problem she names is physical. Agents need to connect to the instruments that run experiments. She sees early signs of common standards, citing Anthropic's Model Hardware Standard, which Anthropic previewed on Aug. 27 so agents can operate equipment such as microscopes, liquid handlers and robotic arms. She compared its role to what MCP does for software connections.

Standards alone will not solve it. Instrument makers have to adapt equipment that can stay in service for years. Automation also has hard limits: a system that dispenses liquids may struggle with solids, and automated chemistry equipment is restricted in the reagents and reaction conditions it supports. An AI can plan an experiment, but the lab's machines decide whether it can be run.

Reasons for confidence, and caution

Jacobs finds encouragement in math and physics, where answers can be checked by computation. She said those results give her confidence that chemistry and biology, which usually need physical experiments, are achievable from the AI side. What remains is how fast people and the physical world can catch up.

Keep in mind that this is an interview with an executive at a company selling these integrations, so it reflects CAS's view. Still, the framing is useful beyond chemistry. Wherever an AI result will drive a decision, two questions matter: where did the underlying facts come from, and who confirms the output?