Tool calling (function calling)

What it is

A tool is a function the platform exposes to the language model with a name, a description and a list of parameters: find_availability(provider, date_range), create_booking(patient_id, slot), transfer(queue). When the model decides a tool is needed, it emits a structured call instead of a reply. The platform runs it against your API, returns the result, and the model then speaks. Tool calling is what turns a chatty model into an agent that can act.

Each tool call adds a round trip: the model's decision, your system's response time, and a second model pass to phrase the answer. A turn that checks availability is therefore slower than a greeting, sometimes by a second or more.

Why it matters when buying

Tool calling is the boundary between the vendor's product and your systems, so it is where integration cost, latency and risk concentrate. Models can call tools with wrong parameters, call them twice, or call them when they should have asked a clarifying question. Anything that writes (a booking, a payment, an address change) needs a confirmation guard: read the details back and get a "yes" before the call fires.

What to ask

Ask which tools ship pre-built for your systems and which you must build. Ask for voice-to-voice latency on tool-backed turns separately from greetings. Ask what the agent says while a slow tool runs. Ask how a failed or timed-out tool call is handled and reported. In the demo, watch a booking land in a sandbox of your own system.