To check today’s weather, KTX schedules, exchange rates and internal company data separately, you need to switch between multiple apps and systems. AI that only uses information available up to the time its training was completed finds it difficult to reflect constantly changing circumstances.
Tool Calling enables AI to select external tools suited to the user’s request, retrieve the latest information and execute tasks. This article explains how Tool Calling works, its applications and what to consider when designing it.
What is Tool Calling?
Tool Calling is a technology that connects large language models to external functions, APIs, databases and business systems. After analysing the user’s request, AI determines the required tools and input values, then uses the results returned by those tools to provide an answer or perform the next task.
| Category | Training data-based processing | Tool Calling-based processing |
|---|---|---|
| Information timing | Uses information available up to the model’s training date | Uses external data available at the time of the call |
| Calculation | Generates results based on context | Calls a calculator or calculation function to verify the result |
| Internal company data | Uses only information included in the training process | Queries authorised databases and document systems |
| Task execution | Generates a response | Performs tasks such as making reservations, registering, sending and generating reports |
How a request is processed
What is possible with Tool Calling
Tool Calling connects information retrieval and task execution within a single conversational flow. Users can provide conditions in natural language without knowing how to search each service or where its menus are located. Enterprises can configure different callable tools based on user permissions and business rules.
| Use cases | Connected systems | Processing examples |
|---|---|---|
| Real-time information | Weather, traffic, foreign exchange, and stock price APIs | Retrieving current information and comparing results based on conditions |
| Accurate calculations | Calculators, statistical functions, and fare calculation APIs | Calculating taxes, distances, costs, and statistical values |
| Internal information | Document repositories, databases, and search systems | Searching for policies, contracts, and performance data within the scope of permissions |
| Schedule management | Calendars and booking systems | Checking available times and adding events to the calendar |
| Communication | Email, messenger, and notification systems | Drafting and sending messages after user approval |
| Task automation | Reporting, customer relationship management, and enterprise resource planning systems | Collecting data, filling out forms, and changing statuses |
Points to check during design
As connections to external systems increase, the scope of data access and execution permissions also expands. Service operators must define specific validation procedures before and after calls, error-handling methods, and cost limits. Users must be able to check what information is sent to which systems.
| Check item | Potential issue | Design approach |
|---|---|---|
| Personal data protection | Personal data may be sent to external systems during the call | Send only the necessary data and manage encryption, retention periods, and conditions for sharing with third parties |
| Access permissions | Users may access data or functions they are not authorised to use | Apply user authentication and role-based permissions, and recheck permissions before making a call |
| Data accuracy | An API may return outdated or incorrect values | Display the data timestamp and source, and apply validity checks and cross-validation |
| Execution accountability | Unintended actions may be confirmed during booking, payment, or sending | Apply user confirmation and approval steps to actions that are difficult to cancel |
| Cost management | Operating costs may increase due to repeated calls and bulk requests | Set limits on the number of calls, store results, and establish rules to prevent duplicate requests |
| Maintenance | External APIs and input formats may change | Operate version control, automated testing, failure detection, and fallback procedures |
- Limit sensitive information to the scope necessary for the purpose of the call.
- Require user approval for payments, bookings, transfers and deletion tasks.
- Continuously record the tool's response time, error rate and call cost.
- Display the source and retrieval time of external data in the results.
- When a tool failure occurs, apply retry limits and fallback handling criteria.
CLEVI's approach to applying AI to real-world work
The usefulness of AI in business depends on model performance, the tools that can be connected, execution accuracy and the level of security controls. By connecting multiple tools to a single request, you can handle everything from information retrieval to organising results and performing follow-up tasks on the same screen. Reflecting role-based permissions and approval procedures for each business process can reduce repetitive tasks while maintaining the organisation's security policies.
- Tool Calling
- Function Calling
- Generative AI
- AI agent
- Business process automation
- API integration