What is a intent recognition?

Intent recognition meaning

An intent is the goal behind a message. Typical intents for a business are: ask price, check opening hours, book, cancel, track an order, report a problem and speak to a person. Details inside the message, such as a date or order number, are called entities and are pulled out separately.

How it works

Older systems matched keywords, so "cancel" triggered the cancel flow even in "do not cancel". Later ones were trained on example sentences for each intent. Current systems use a large language model given a list of possible intents in a prompt, which copes better with unexpected wording and with mixed Nepali and English.

A good system also has an "I am not sure" outcome. When confidence is low it asks a clarifying question or hands over, instead of guessing.

Intent recognition vs IVR menus

An IVR asks the caller to do the sorting by pressing keys (DTMF tones). Intent recognition lets the caller just say what they need and does the sorting itself. This avoids long menus, and callers who press the wrong option. See IVR vs AI call agent.

Using intent to route calls

Once the intent is known, call routing can act on it:

• Billing questions go to the accounts team.
• Urgent problems go to the front of the call queue.
• Simple questions are answered directly from the knowledge base.
• Anything unclear goes to a person with a summary of what was said.

A worked example on a pharmacy line

Four callers say: "Do you have paracetamol?", "Is Dolo available?", "Paracetamol cha?", "I need fever medicine". All four have one intent, check stock, with different details. A fifth says "My father has chest pain and needs medicine now". This is a different intent, an emergency, that should not be answered from the stock list at all.

Good intent design includes high-risk intents with clear rules: emergencies go to a person at once or to emergency advice, and complaints go to a manager.

Designing the list of intents

A short list that covers most calls beats a long list that overlaps. Start by reading a week of real calls and counting the reasons for them. Merge intents that lead to the same action, and split those that need different answers. Add an "other" intent that always leads to a person.

Checking that it works

Take a sample of two hundred real messages or calls and label the correct intent for each yourself. Run the system on the same sample and compare. Look at which intents get confused with each other, because those often need to be merged or described more clearly. Repeat each month, since customers’ reasons for calling change with seasons, offers and festivals.

Intent in Nepali and mixed language

The same request can arrive as "refund chahiyo", "mero paisa pheri paunu paryo" or "I want my money back". A system that learned intents from English examples may handle only the third. Test each intent with Devanagari, Roman Nepali and English wording, and add the missing styles to your examples or instructions.

On a phone call the words first pass through speech-to-text. A misheard word can change the intent, so a good system looks at the whole sentence and asks for confirmation before acting on something important.

Intent recognition on AirFone

The AI call agent works out what each caller needs, answers from the business’s own information when it can, and hands the call to a person with a short summary when it should.