People do not state their business in a clean sentence. They begin mid-thought, change direction, bundle two requests together and leave the important detail until last. Intent recognition is the step that turns that into something a system can act on.
Why the old approach struggled
Traditional systems matched keywords and phrases. Every new way of expressing the same request needed a new rule, and the rules multiplied until nobody could safely change them. Anything unanticipated fell through to a default that helped nobody.
Language models changed this substantially, because they generalise. "Cancel my booking", "I can't make it tomorrow" and "something's come up, need to move Thursday" all resolve to the same intent without three separate rules.
Where it still gets hard
- Multiple intents in one breath. "Move my appointment and also do you deliver to Noida" is two requests. Systems that handle only the first quietly drop the second.
- Missing detail. The intent is clear, a required field is not. The system has to know what to ask for next, and ask once rather than interrogating.
- Code-switching. Mixed Hindi and English changes the input enough to affect accuracy, which is why it deserves testing in its own right.
- Confidence. Knowing that it does not know is more valuable than a marginal accuracy gain. A system that hands over when unsure beats one that guesses slightly more often.
How to evaluate it
Not on a vendor's accuracy percentage. On a sample of your own real conversations, scored on whether the outcome was right — the booking made, the query routed correctly. Outcome accuracy is the only number that translates into anything commercial.
Questions
- How many intents does a system need?
- Fewer than most people expect. A small set covering the majority of contacts, handled completely, outperforms a long list handled shallowly. Start with what actually fills your inbox and phone line.
Wondering whether this applies to your business? That is a thirty-minute call, not a research project.
Book a Consultation Call