One of the biggest misconceptions I still see around NLU in CX is this idea that it works like a keyword detector.
Customer says one word, bot finds the same word, match made.
But that is not really what is happening.
In Genesys Cloud, NLU is built to classify a customer’s utterance into a set of predefined intents. In other words, it is not just looking for a word. It is trying to understand what the customer is trying to do
That difference matters a lot.
If NLU were only word matching, then every variation would need to be manually accounted for:
-
“I want to pay my bill”
-
“make a payment”
-
“pay invoice”
-
“settle my balance”
A simple keyword model would struggle unless you explicitly taught every possible phrase.
But Genesys Cloud NLU does more than that.
When bot authors create an NLU model, they define intents and train each one with example utterances that represent how a user might ask for that thing. Then, when a new input comes in, the model computes the probability of that utterance belonging to each intent and returns the most likely matches with confidence scores
That means the system is not asking:
“Did I find the exact same words?”
It is asking:
“Which intent does this sound most like?”
And that is a much better way to think about conversational AI in Genesys Cloud.
Under the hood, the model is using techniques like word embeddings and TF-IDF vectors to improve detection, which helps it go beyond exact token matching and better represent meaning and similarity between words and phrases
So when someone says NLU is just matching words, I think that misses the real point.
The real work of NLU is in recognizing patterns of meaning across different ways customers express the same need.
Of course, that does not mean magic.
The quality of the model still depends heavily on how well the intents are designed and how representative the training utterances are. In fact, Genesys is very clear that the customer or NLU author is responsible for defining the intents and utterances used to build the classifier
That is why good NLU design is not only a technical task. It is also a CX task.
You need to understand how customers actually ask for help, how they vary their language, and where intents are too broad, too narrow, or too overlapping.
So for me, one of the best ways to describe NLU in Genesys Cloud is this:
It is not trying to catch exact words.
It is trying to identify likely intent.
And that is exactly why training data, testing, and iteration matter so much.
Curious how others explain this: when you talk about NLU with customers or newer users, what is the biggest misconception you usually have to correct?
#GenesysCloud#Training-------------------------------------------
------------------------------
Rodrigo Romao
NALA Team Lead - Genesys - Employees
------------------------------