Tool Use / Function Calling
What it means in practice
A model on its own can only generate text based on what it has been trained on and what sits in its context window. Tool use gives it a defined set of callable functions, for example “look up a client record”, “check today’s exchange rate”, or “search this document store”, each with a description of what the function does and what parameters it expects.
When the model decides a function would help answer the prompt, it outputs a structured call to that function rather than a plain-text answer. The calling system executes the function, returns the result to the model, and the model continues its reasoning with that fresh information in hand. This can happen once, or in a loop of several calls, depending on how the task is designed.
In a professional services setting, tool use is what lets an AI assistant pull the actual current balance from a practice management system, check a live regulatory register, or query a case management tool, rather than guessing or working from stale training data.
How it differs from agentic AI
Tool use is a capability. Agentic AI is a design pattern that often depends on it. An agentic system observes, plans, acts and reassesses across multiple steps toward a goal; tool use is frequently the mechanism it relies on to act, letting it look things up, write to a system, or trigger a downstream process instead of being confined to producing text. A model can have tool use without being agentic (a single well-defined tool call answering one question), and in principle an agentic workflow could be built without external tools at all, though in practice most useful agentic systems in a business context depend on tool use to do anything beyond generating text.
Governance and professional services context
Every tool exposed to a model is effectively a permission granted to it. In a regulated environment, that permission needs the same scrutiny as any other system access: what data can this tool read, what can it write or change, and under what conditions should a human confirm the action before it executes. Read-only lookups (checking a record) carry a different risk profile to write actions (updating a record, sending a communication, submitting a filing), and governance design should reflect that distinction explicitly rather than treating all tool calls as equivalent.
Audit logging matters here specifically: because a tool call is a discrete, structured action rather than free text, it is usually easier to log, review and explain after the fact than a purely generative response, which is one of the practical advantages of well-designed tool use in a governed setting.
Insight
Misconceptions
Tool use is not the same as plugins or integrations in the generic software sense, though the terms get used loosely. A plugin is typically a fixed feature added to a product; tool use is the model's own decision, at inference time, about whether and how to call a given function based on the task at hand. Tool use also does not make a model inherently more accurate. It gives the model access to current or specific information it did not have, but the model can still call the wrong tool, misinterpret a result, or take an unnecessary action, so the same oversight principles that apply to any AI output still apply to the tool calls that produced it.