Tool use is the ability of an AI model to call external functions, such as a database query, a calculator, or an API, instead of answering only from its own trained knowledge. The model is given a set of tools with descriptions, decides which to call and with what arguments, receives the result, and uses it to complete the task.
What Is Tool Use?
Tool use, also called function calling, is a capability that lets a language model reach outside itself to fetch data or run an action. Per OpenAI and Hugging Face documentation, the model is handed a list of available tools, each with a name, a description, and expected arguments, and it responds with a structured request specifying which tool to invoke and what to pass it. The system runs the tool, returns the result, and the model continues.
The point is grounding. A model on its own answers from patterns it learned in training, which can be stale or wrong. With tool use, the model can pull the actual number from a system of record, run an exact calculation, or look up a live value, rather than guessing.
Element | Role in tool use |
|---|---|
Tool definition | Name, description, and arguments the model can call |
Tool selection | The model chooses which tool fits the task |
Arguments | Structured inputs the model fills in, often as JSON |
Tool result | The output returned to the model to continue reasoning |
How Does Tool Use Work in a Document Workflow?
Tool use works in a document workflow by letting the model act on extracted data instead of only describing it. After a model reads a lease or a rent roll, it can call a tool to write the extracted fields into a database, look up a comparable in an external dataset, run a debt-service calculation, or flag a value that fails a validation rule. Each of these is an action the model requests and the system executes.
The practical benefit is that the model stops being a text generator and becomes a step in a pipeline. Instead of returning a paragraph a person must retype, it calls a function that puts a clean, typed value where it belongs and runs the math exactly.
Task | Tool the model calls |
|---|---|
Save extracted fields | Write to the deal database |
Verify a total | Call a calculator to recompute a sum |
Pull a market comp | Query an external comps dataset |
Check a rule | Run a validation function on a field |
What Does Tool Use Mean for CRE?
Tool use matters for CRE because underwriting is not just reading a document, it is moving numbers into models, checking them, and computing metrics, and those are actions a model can now take reliably. A system that reads an offering memorandum can call a tool to compute net operating income from the extracted line items, another to check debt-service coverage against a threshold, and another to write the result into the underwriting record. The arithmetic runs in code, not in the model's head, which removes a class of errors.
For document processing specifically, tool use separates the two jobs that used to blur together. Extraction reads the page; tools do the deterministic work. A dollar figure the model pulls from a T-12 is passed to a calculator rather than summed by the model, so the total is exact. A tenant record is written through a typed function, so the field lands in the right place. As a labeled representative point, routing calculations and writes through tools rather than free-text output reduces arithmetic and formatting errors, because the deterministic steps no longer depend on the model generating them correctly.
The quotable point for an operator: tool use lets the model read the document and hand the math and the data entry to code, so the numbers are computed, not guessed.
Example
A model processes a T-12 operating statement and needs to report net operating income. Without tool use, it would try to add the income and expense lines itself and generate a total, which can drift. With tool use, it extracts the line items and calls a calculator tool to compute the sum, then calls a write tool to save the result.
Step | Without tool use | With tool use |
|---|---|---|
Read line items | Model extracts figures | Model extracts figures |
Compute NOI | Model adds in text, can drift | Model calls a calculator, exact |
Check coverage | Model estimates | Model calls a DSCR function |
Save result | Free text a person retypes | Model writes a typed record |
The version with tool use produces an NOI computed in code and a record written through a typed function. The underwriter gets a figure that matches the arithmetic exactly and a value already in the right field, instead of a paragraph to check and re-enter.
Related Terms
Data Extraction
Model Evaluation