When Anthropic introduced structured tool-use capabilities for Claude in 2023, the primary use case in most developers' imaginations was function calling: giving the model a defined set of operations it could invoke to retrieve information or perform discrete actions. The reality of how enterprises have adopted the capability, two years on, is considerably more architecturally ambitious.

The companies that have moved beyond experimental Claude deployments are, in the majority, building systems in which the model operates as an orchestration layer — receiving high-level objectives, decomposing them into sequences of tool calls, synthesising the results, and producing outputs that feed downstream systems rather than human readers. The conversational interface, in this pattern, is a configuration artefact rather than the primary interaction mode. The model is talking to databases, APIs, and other AI systems far more than it is talking to people.

This shift matters for how enterprises think about the value and risk profile of their AI investments. A model that answers questions has a limited blast radius; a model that orchestrates actions has a fundamentally different risk surface, and the governance frameworks that enterprises have built around conversational AI are proving inadequate for the agentic pattern.