Building AI Agents for Regulated Industries

Brendan Geils shares insights on building and deploying AI agents for highly regulated industries including finance, legal, defense, and healthcare. The conversation covers the Forward Deployed Engineer model, why selling to business functions outperforms selling to IT, and how the future of enterprise software may look radically different than today.

About the Guest

Brendan Geils is the Founder of Athena Intelligence, a company building AI agents for regulated industries. With a background that includes TS clearance with the US government and experience at Palantir, Brendan has deep expertise in deploying technology solutions for enterprises with stringent security and compliance requirements.

Key Topics Discussed

AI Agents in Regulated Industries

Athena Intelligence focuses on finance, legal, defense, healthcare, and Fortune 500 companies with elevated brand risk. These organizations want AI but need vendors who understand their unique security, permissioning, and audit requirements.

The Forward Deployed Engineer Model

Borrowing from Palantir's approach, Athena uses FDEs to shepherd customers through AI adoption. This model enables product exploration with real customers before deciding what to build at scale—and increasingly, the FDE work itself is being automated by AI agents.

Selling to Business vs. IT

A pivotal lesson: stop selling to IT and engineering. Instead, sell to business functions who care about ROI on real tasks, not the latest RAG architecture or benchmark scores. One customer signed a pilot in two weeks through business functions versus six weeks of procurement limbo through IT.

The Future of Work Interfaces

People work in Microsoft Office all day—Word, PowerPoint, Excel, Outlook. Rather than forcing users into new AI products, push capabilities into their native environments. But as agents handle more end-to-end work, users may eventually leave traditional applications entirely, just as developers are moving from IDEs to autonomous coding agents like Devin.

AI Pricing Models

Enterprise buyers are used to user-based pricing for predictability. But with token consumption growing exponentially (10 billion tokens through OpenAI in one month, then 10 billion through Anthropic the next month), consumption-based models are becoming necessary. The future may include "overtime" pricing similar to human workers.

Expanding Work, Not Replacing It

Due diligence teams don't do less work with AI—they do more. First-year analysts who spent 80-120 hours on document review now get promoted to higher-value work like onsite visits. Companies using AI are hiring more, not less.

Key Quotes

"We don't want to be doing repeatable workflows. That's probably not where we're gonna be best suited... We're probably gonna be better at the creative, more unique, more open-ended, more edge cases style work."

"The more you want to control the path by which this thing gets to an answer, you're probably gonna start degrading some of its capability."

"If your fear of eliminating work is too big, then your scale of ambition is too small—because there's so much more you could do in your business."

"Move fast, try things. But brand risk and cyber risk are probably the more important parts to take hold of first."

Notable Insights

• Athena's company name was generated by GPT-3.5—the goddess of wisdom and war felt fitting for AI in defense
• Every internal meeting at Athena has an AI recorder that's been collecting data for two years; soon it will attend customer calls autonomously
• Users discovered they could parallelize agent work through Excel's familiar interface—dragging formulas down to spawn hundreds of agents
• The "explainability trade-off": demanding too much control over how AI reaches answers degrades its capability