Voice AI technology

Letting a Voice Agent Look Things Up: Tools and Function Calls

How function calling works for phone voice agents: lookups versus actions, what makes a good tool, the risks of speech input and how to keep calls natural.

Call editorial team · Published · 6 min read

Information that changes every hour cannot live in a prompt. An order status, a stock level or an appointment slot is only correct for a moment. To use such facts, a voice agent needs a way to ask another system, and that mechanism is usually called function calling or tool use.

A support agent checking information on a screen while on a call
Illustrative image

This guide explains how it works in plain terms, what makes a good tool, where the risks lie and how the idea relates to the way Call's AI agents work today. For background on the voice pipeline itself, see How Voice AI Works on a Phone Call.

How a tool call works

A language model writes text. It cannot open a database or press a button. Function calling is a convention: you describe a set of functions to the model, with a name, a purpose and the inputs each expects, and the model may respond with a request to use one.

The life of a tool call

  1. Caller asksFor example: where is my order?
  2. Model proposes a callIt returns a function name and arguments, such as an order number.
  3. Your software validatesIt checks the arguments, the caller's permission and any limits.
  4. The system runs itIt queries the order system and gets a result.
  5. Result returns to the modelThe model turns the data into a short spoken answer.

The key point is that your software, not the model, performs the action. That gives you a place to enforce rules, log what happened and refuse anything unreasonable.

Lookups versus actions

Not all tools carry the same risk. A lookup reads information. An action changes something in the world. Treat them differently.

Read-only lookups versus actions that change things

LookupAction
ExamplesOpening hours, order status, queue lengthCancel a booking, issue a refund, update an address
Risk if wrongA wrong answer, which can be correctedA change that may be hard to undo
Typical safeguardsIdentity check, limited fieldsConfirmation step, human approval, audit log
Good starting point?YesOnly after lookups work well

A cautious path is to launch with lookups only, watch real calls and add actions slowly, with a human confirming anything consequential. This is the same attitude that sits behind Guardrails for Voice AI Agents.

What makes a good tool

Names and descriptions matter

The model picks tools using the names and descriptions you give them. Vague wording leads to wrong choices. Write descriptions as instructions to a new colleague: when to use the tool, when not to, and what each input means.

Spoken input is messy

Arguments come from speech. An order number may arrive with a digit wrong, or a name with an unusual spelling. Read key values back to the caller before using them, and design the tool to return a helpful result for near misses. Getting Names, Numbers and Addresses Right by Voice covers confirmation techniques in detail.

The risks to design around

Where tool use goes wrong

  • Wrong tool or argumentsThe model misreads the request, so validate everything before running.
  • Over-sharingA lookup returns more than the caller should hear, so return only what is needed.
  • Unverified identityReading account details to the wrong person, so verify before disclosure.
  • Slow or failing systemsDead air or a confusing error, so set time limits and spoken fallbacks.
  • Injected instructionsText from a lookup result that tries to give the model orders, so treat results as data only.

Each risk has an engineering answer, but the safest design principle is to give the agent the least access that does the job. Do not hand it a general database connection when a single narrow lookup is enough.

Keeping the call natural

A tool call adds time: the model decides, your system responds, the model writes the answer, and speech is produced. The caller hears silence. Tactics that help include a very short acknowledgement before a lookup, caching values that rarely change, and limiting each turn to one tool call. Voice AI Latency describes where the delay comes from. If a lookup fails or times out, the agent should say so briefly and offer a message or a person rather than waiting.

A worked example, step by step

Imagine a hypothetical furniture shop whose agent answers questions about deliveries. The shop defines one read-only tool that takes an order number and returns the delivery status and the delivery window. The agent is told to use it only after the caller has given an order number and confirmed it by hearing it read back.

  1. The caller says they are waiting for a sofa and gives an order number.
  2. The agent reads the number back in groups and the caller confirms.
  3. The tool returns a status of out for delivery and a window of the afternoon.
  4. The agent says that the order is out for delivery and should arrive this afternoon, then asks whether anything else is needed.
  5. If the tool returns not found, the agent says it cannot find that order and offers to take a message for the team.

Notice what is missing: the agent cannot reschedule the delivery. If the caller wants a different day, the agent takes the request and the team handles it. Starting with a narrow, safe tool like this teaches you how callers actually speak before you widen the scope. Keep a record of every call the tool receives so you can spot patterns, such as a common mishearing of one digit, and fix them in the confirmation step.

How this relates to Call

In Call, an AI agent answers from its instructions, its persona, the script attached to the line and the passages of its knowledge base that match the question. It is not designed to take actions such as booking or confirming appointments, calling back, sending anything or offering discounts during the call. It says so and gives the route your sources describe, for example a booking link or a callback from your team. For facts that change often, keep them in the agent's sources and update them, and use the help desk for follow-up work that needs a person. If you want to connect Call to your own systems, the developer interface is described in Call Developers and API; check what it offers before designing around any particular behaviour.

When you evaluate any platform that advertises tool use, ask the practical questions in the checklist below.

Questions to ask about tools on a voice agent

  • Which systems can it reach, and with what permissions?
  • Are lookups separated from actions that change data?
  • Is there a log of every call it made, with arguments and results?
  • What happens when a lookup fails or is slow?
  • Can a person approve consequential actions?
  • How is the caller's identity checked before details are spoken?
  • Can you test tools with realistic spoken input before launch?

For the evaluation side, see How to Evaluate a Voice Agent's Conversations.

Frequently asked questions

What is function calling in simple terms?

It is a way for an AI model to ask your software to do something, like look up an order. The model writes the request, your system carries it out and returns the result for the model to explain.

Does the AI run the function itself?

No. The model only proposes the call. Your own code checks and runs it, which is where you can apply rules, limits and logging.

Is it safe to let a voice agent change records?

It can be done carefully, but start with read-only lookups. For anything that changes data, add confirmation, limits and human approval where the stakes are high.

Why does a lookup make the call feel slow?

Each step takes time: deciding on the call, running it, writing the reply and speaking it. Fast tools, a short acknowledgement and one lookup per turn keep the delay tolerable.

Can a caller trick the agent into using a tool wrongly?

They can try. Validate arguments, check permissions in your code and give the agent narrow tools, so a misguided request cannot do real harm.

Does Call's AI agent look things up in my systems?

Call's agent works from its instructions, script and knowledge sources and does not book, send or confirm things during the call. Keep changing facts in its sources and use the team for follow-up.

See how Call's AI agent answers from your sources

In the demo console you can try incoming calls in a simulation.

Related guides