What is intent detection?

Intent detection is the step in which a support system works out what a customer wants from their message, such as "dispute a charge", so the request can be routed or handled correctly. AI agents do it far better than keyword rules. But knowing what someone wants is not the same as being allowed to give it to them.

Why does intent detection matter once agents act?

In older chatbots, intent detection decided which scripted flow to run. A wrong guess sent the customer down the wrong path, which was annoying but limited. An AI agent that acts on the intent it detects can do something: reset a credential, waive a fee, change account details. A wrong or manipulated intent then leads to a wrong action.

The bigger risk is treating a clear intent as authority. A fraudster who says "I've lost my phone, please move my number to a new SIM" expresses the intent perfectly. The agent understanding the request is the easy part. Whether the person is entitled to it is a separate question, answered by verification and the agent's bounds, which is why handling SIM swap requests safely with an AI agent is mostly about what happens after the intent is clear.

How does intent detection work in practice?

A modern agent reads the whole message, and often the conversation so far, and maps it to one or more intents from a defined list. It may also recognise that a message holds several intents ("I want to change my address and I think I was overcharged") or none it is allowed to handle.

The defined list matters. Each intent the agent may act on should map to a capability in its brief, with its own depth and conditions. Intents outside the list should lead to a handover, not a best guess. And intents with high risk, such as anything touching credentials, contact details or payments, should trigger verification before any action.

An example from a lender's support chat

Take a lender whose support agent receives: "I missed my payment because I was in hospital, can you take the late fee off?" The agent detects two things: a fee waiver request, and a possible vulnerability indicator.

Its bounds allow it to waive a first late fee under a set amount where the account is otherwise up to date. But the vulnerability indicator takes priority: the firm's policy says such cases go to a trained person. The agent acknowledges the request, does not waive the fee itself and passes the conversation on with a note. The question of whether an AI agent can waive a late fee depends on context the intent alone does not carry.

How is intent detection different from entity extraction?

Intent detection identifies what the customer wants to do. Entity extraction pulls out the specific details needed to do it: the account number, the transaction date, the amount, the new address.

Both can go wrong independently. An agent can correctly detect "dispute a charge" and extract the wrong transaction. In agents that act, entity extraction errors are often the more dangerous, because the action happens to the wrong item. Each extracted value that drives an action should be confirmed against a sealed or signed record, not taken from the customer's message alone.

Which terms sit next to intent detection?

• Routing: sending a conversation to the right agent, team or queue based on its intent.
• Verification: confirming the customer is who they claim before a sensitive intent is acted on. Whether an AI agent can reset a customer's password turns on it.
• An AI support agent that knows when to say no treats unknown intents as a reason to hand over.