Function calling, also called tool use, is the ability of an AI model to request that a piece of software carries out a specific action, such as checking stock, looking up an order or booking an appointment, and then use the result in its answer. The model does not run the action itself; it produces a structured request, and your system decides whether to run it.
How function calling works
A developer describes the available tools to the model: each one’s name, what it does and what information it needs. A booking tool might be described as “check_availability: returns free appointment slots for a given date and service”.
When a customer writes “Have you got anything for a men’s haircut on Saturday morning?”, the model recognises that the tool is relevant and replies, not with prose, but with a structured request: call check_availability with the date of the coming Saturday and the service “men’s haircut”. The application runs that request against the real booking system through its API, sends the results back to the model, and the model turns them into a natural reply: “There are slots at 9.30 and 11.15. Shall I book one?”
Chaining several of these steps together, with the model deciding what to do next, is what turns a chatbot into an AI agent. The Model Context Protocol (MCP) is an open standard, released by Anthropic and since adopted by other AI companies, that makes it easier to connect AI applications to tools and data sources in a consistent way.
Why it matters
Function calling is what lets AI move from talking about things to doing them. For a business, the practical uses include website assistants that check real availability rather than guessing, internal tools that pull figures from Google Ads or a CRM on request, and support assistants that look up an order status instead of telling the customer to email.
It also shapes how customers will increasingly find and buy. At the time of writing (October 2026), AI assistants and browsers are starting to complete tasks for users, such as comparing products, filling baskets and making reservations, through tool use. This is often called agentic commerce. Businesses whose stock, prices and booking systems can be read reliably by software are better placed for that shift than those whose details sit only in PDFs or images.
Common mistakes
- Giving the model too much power. A tool that can issue refunds or delete records should not run without limits or approval.
- Vague tool descriptions. The model chooses tools based on their descriptions. Unclear wording leads to the wrong tool or wrong inputs.
- Trusting inputs blindly. The model may pass a malformed date or an invented order number. Validate every request before acting on it.
- Exposing personal data. A tool that returns full customer records to answer a simple question shares more than needed.
- No logging. Without a record of what the model called and why, mistakes are hard to trace.
How to act on it
If you are commissioning an AI assistant or agent, start with read-only tools, such as checking availability, opening hours or order status, before allowing anything that changes data or spends money. For actions with consequences, keep a human in the loop who confirms before the action runs. Ask your developer to log each tool call and to return only the fields the task needs.
On the marketing side, make sure the facts AI systems look for (services, prices, areas covered, availability) are published clearly on your site. How AI tools find, read and act on information about a business is the subject of my AI search optimisation work.