Goldman Sachs had options for where to deploy AI agents first. Customer service. Content generation. Internal knowledge management.

They went straight to compliance and accounting.

Compliance work has historically broken automation attempts because it combines characteristics that resist simplification. You're processing millions of transactions annually against regulatory frameworks that require interpretation, not just pattern matching. Screening sanctions lists. Determining beneficial ownership structures. Parsing global databases to flag exceptions. The rules are clear until they're not, then you need judgment about how frameworks apply to ambiguous situations.

The stakes make this harder. Regulatory penalties run into hundreds of millions. Audit trails need to satisfy regulators who will scrutinize decision logic. You can't deploy something that works 95% of the time when the 5% creates compliance failures. The error tolerance is effectively zero, which means the system needs to handle edge cases reliably, processing standard scenarios and exceptions with equal reliability.

Goldman's deployment becomes interesting when you see compliance as a place to test whether agents work at production scale in high-stakes environments. If agents can handle this domain—with its regulatory scrutiny, audit requirements, and complex reasoning demands—then other domains with similar characteristics become viable targets.

The progression Goldman followed reveals this logic. They started with coding, where Claude demonstrated strong capability. Then they asked whether that capability was specific to software development or whether it represented something more general.

That discovery shifts the question from "what can AI do?" to "where else does this reasoning capability apply?"

The answer points to domains that share compliance's profile: scaled, complex, process-intensive work where high volumes of structured and unstructured data need parsing, layered rules need application, and judgment needs exercising where rules run out.

Argenti mentioned employee surveillance, investment banking pitchbooks, vendor management as potential next targets. They share a pattern:

• Employee surveillance requires parsing communications data against policy frameworks, flagging patterns that might indicate violations, exercising judgment about context and intent
• Pitchbooks combine market data, financial analysis, and regulatory requirements into documents that need to be accurate, compliant, and persuasive
• Vendor management involves evaluating third-party relationships against risk criteria, tracking compliance obligations, making decisions about which relationships require enhanced scrutiny

They all fit: clear frameworks that require complex application, high stakes that justify deployment investment, massive scale that makes automation valuable, work that combines data processing with interpretation and judgment calls.

The next 12-24 months will likely see agents deploying in domains with similar characteristics to compliance. Financial operations beyond Goldman. Supply chain compliance. Regulatory reporting across industries. Contract review and analysis. These domains share the profile where passing the compliance threshold test matters—where demonstrating reliable reasoning under regulatory scrutiny signals readiness for production deployment.

The domain choice reveals where agents are on the maturity curve. When you choose compliance as your first production deployment, you're signaling confidence that the technology can handle environments where automation has historically failed and where mistakes create real business problems.

Domain selection carries weight beyond the technical achievement. It shows where the technology has reached production-grade reliability in environments that matter.