Anthropic Connects Claude Science to 60+ Databases and Tools for Enterprise Work

Anthropic is shifting its enterprise pitch from what Claude can answer to how Claude can work inside a company. Its new Claude Science workbench and its financial-services integrations pair existing models with outside data, specialized tools, agents and controls designed for workflows where a convincing response is not enough.

The distinction is clearest in Claude Science. Anthropic says the product is not a new AI model or a more capable biology model. It is a research workbench that routes existing Claude models through more than 60 scientific databases and specialized tools for multistep work.

Anthropic life-sciences head Eric Kauderer-Abrams says better models and the products around them reinforce each other: stronger models can use more sophisticated tools. The company says Claude can coordinate specialized scientific systems, including through Nvidia’s BioNeMo ecosystem, for work such as protein-design workflows.

Claude Science adds a reviewer agent that checks citations, figures and numbers before a human evaluates the result. That design addresses a central failure mode in scientific work: plausible mistakes can be difficult to spot. But Anthropic says the system cannot do the physical laboratory work that produces experimental data, and scientists must verify results before consequential actions such as regulatory submissions.

The financial-services version of the strategy begins where professionals already work. Anthropic lists integrations with LSEG, FactSet, S&P Global, Morningstar, Microsoft applications, Excel and PowerPoint. Jonathan Pelosi, Anthropic’s financial-services head, says Claude is used for tasks including covenant-term analysis, credit memos, know-your-customer screening and actuarial-workbook review.

The controls Anthropic says institutions can set

• Which systems an agent can access.
• How much independence the agent has.
• Who must approve an action before it proceeds.

Pelosi says every output is traceable and every step logged. He also warns against “agent sprawl”: firms could lose the promised consistency and auditability if every desk deploys its own agent rather than using a small number of governed systems.

Anthropic reported roughly $47 billion in run-rate revenue by late May 2026, up from about $9 billion at the end of 2025, and said more than 1,000 enterprise customers spend at least $1 million annually. It also says some Claude Science users have cut project timelines by roughly tenfold; one University of California, San Francisco lab manually validated a Claude-generated glioma review.

Anthropic’s argument is that reliability matters more when an agent must complete a long sequence of consequential tasks, and that its operating layer must survive model upgrades without forcing customers to rebuild. The unresolved question is whether those integrations and controls become a lasting edge: they are valuable, but financial institutions also have incentives to keep control of systems holding their most sensitive information and decisions.