AnthropicScaling enterprise AI: Fireside chat with Eli Lilly’s Diogo Rau and Dario Amodei
CHAPTERS
- 0:00 – 1:00
Why enterprises need AI optimized for truth, not engagement
Dario contrasts consumer AI incentives (engagement and growth) with enterprise needs (accuracy, reliability, and truth). He argues that enterprise deployments—especially in regulated, high-stakes domains—require models that resist flattering users and instead prioritize correct, grounded outputs.
- •Consumer AI competition can reward engagement over correctness
- •Enterprise settings demand accuracy, reliability, and truth-seeking behavior
- •Anthropic’s model design choices aim to align with enterprise incentives
- •High-stakes decisions amplify the cost of incorrect or overly agreeable outputs
- 1:00 – 1:58
Sycophancy as a real enterprise risk (and why it matters in life sciences)
They discuss “model sycophancy,” where a model validates a user’s idea even when it’s wrong. Dario highlights how this is merely annoying in consumer use but potentially dangerous in enterprise contexts like drug R&D where decisions can cost millions and affect patient outcomes.
- •Definition and examples of sycophancy (model agrees regardless of truth)
- •Consumer anecdotes vs. enterprise consequences
- •In drug discovery, false confidence can drive costly, incorrect investment
- •Enterprise AI should challenge and verify, not just affirm
- 1:58 – 2:25
What enterprises value: deeper domain knowledge and economically valuable capability
Dario explains that improving specialized knowledge (e.g., biochemistry from undergrad to graduate level) is largely irrelevant to most consumers but highly valuable to enterprises like Eli Lilly. This illustrates why enterprise AI roadmaps should emphasize domain depth and task performance, not only general-purpose features.
- •Domain depth can be a decisive differentiator for enterprises
- •Consumers may not perceive value from specialized knowledge gains
- •Life sciences organizations directly benefit from higher scientific competence
- •Focus on economically valuable tasks rather than engagement metrics
- 2:25 – 3:00
From general models to “skills” and specialized Claudes
Diogo asks about Anthropic’s “skills” concept and whether it’s part of the future roadmap. Dario describes a direction that includes specialized variants of Claude, sometimes via fine-tuning and sometimes via product “wrappers” that provide structured access to relevant information sources.
- •“Skills” as a way to shape models for specific domains and workflows
- •Specialized Claudes can be built through fine-tuning and/or tool+data integration
- •Enterprise needs often require more than a generic chatbot
- •Productization matters: making domain functionality accessible and usable
- 3:00 – 3:07
Example: Claude for Financial Services—value of connecting the right data
Dario uses Claude for financial services to illustrate how connecting a model to domain indices, ratings, and reference sources can create immediate enterprise value. The point is that usability and integration with trusted data can be as impactful as raw model capability.
- •Financial services Claude integrates common indices/ratings sources
- •Integration can unlock value without changing the base model dramatically
- •Making connections easy is itself a product advantage
- •Domain awareness depends on both intelligence and access to context
- 3:07 – 3:36
Toward Claude for Life Sciences: model intelligence plus scientific databases
They discuss building a life sciences-focused Claude that combines improved underlying capability with access to specialized resources. Dario notes the abundance of biology data—proteins, compounds, assays—and suggests these should be “at the model’s fingertips” to support real R&D work.
- •Life sciences specialization likely blends smarter models with data/tool wrappers
- •Biology has many critical databases (proteins, compounds, assays)
- •Effective R&D AI needs fast access to high-quality scientific context
- •The goal is practical utility in real discovery and development workflows
- 3:36 – 3:50
Parting advice: start small, but plan for end-to-end transformation
Asked for advice to drug discovery teams, Dario acknowledges that most organizations begin with small AI use cases. He argues teams should still stay ambitious and prepare for models that can eventually run an entire multi-step process end-to-end rather than only optimizing isolated steps.
- •Small pilots are often necessary, but shouldn’t cap ambition
- •Replacing only a few steps in a long process can create integration friction
- •End-to-end capability may arrive sooner than expected
- •Strategic planning should assume rapid model improvement
- 3:50 – 4:28
Avoid “hill-climbing” that locks you into today’s workflows
Dario warns that swapping AI into a couple of sub-steps can be harder than expected because non-AI steps remain tightly coupled to adjacent stages. This can cause organizations to over-invest in partial automation that becomes obsolete when models can soon handle far more of the workflow.
- •Partial automation can create bottlenecks with non-AI neighboring steps
- •Legacy process coupling makes incremental upgrades complex
- •Organizations risk optimizing for a soon-to-be outdated state
- •Better to redesign around where the technology is heading
- 4:28 – 5:09
Parallelize readiness: prepare now to avoid multi-year deployment delays
Dario argues organizations should prepare for large workflow changes while models improve, rather than waiting until models are “perfect.” Waiting can introduce a further multi-year implementation lag—time that directly delays patient benefit in healthcare and pharma contexts.
- •Waiting for full capability can trigger an additional 2-year enterprise delay
- •Preparation can happen in parallel with model progress
- •Early readiness can translate into years saved
- •Faster adoption can accelerate patient impact
- 5:09 – 5:25
Planning principle: two-year projects must assume a different AI landscape
Diogo summarizes the implication: don’t run long projects assuming today’s AI constraints will persist. Dario agrees that while it sounds obvious, it takes courage and foresight to plan for a rapidly shifting frontier and build adaptable strategies.
- •Long timelines demand assumptions about future, not current, AI capability
- •Roadmaps should be robust to rapid model advances
- •Organizational courage is required to bet on change
- •Strategic flexibility reduces the risk of rework
- 5:25 – 5:34
Closing remarks and thanks
They wrap up the conversation with mutual thanks and an intention to reconnect. The close reinforces the collaborative enterprise focus between Anthropic and Eli Lilly.
- •Conversation concludes with appreciation and sign-off
- •Reinforces partnership-oriented tone
- •Signals ongoing dialogue about enterprise AI adoption
