AI Summaries Transform Customer Support
Key Points
- Customers frequently experience frustration with traditional call centers due to lengthy navigation menus and agents lacking context about prior interactions.
- The speakers propose leveraging generative AI (large language models) to improve the experience by automatically summarizing past call transcripts for agents.
- AI can also perform sentiment analysis on previous calls, giving agents insight into whether the customer’s prior experiences were positive or negative.
- Intent classification by the language model can identify the primary reason for a customer’s call (e.g., product inquiry, billing issue, promotion interest) before the agent speaks.
- Retrieval‑augmented generation (RAG) can be used to supply agents with relevant historical information during call transfers, ensuring continuity and better service.
Sections
- AI‑Powered Call Summaries - The speakers discuss using generative AI, like large language models, to summarize call transcripts, perform sentiment analysis, and classify intent, enabling agents to quickly grasp customer issues and improve the support experience.
- Using RAG for Real-Time Agent Support - The speaker explains how Retrieval‑Augmented Generation can turn speech‑to‑text inputs into LLM‑driven answers, instantly equipping call agents with expert knowledge and eliminating costly transfers and extensive training.
- Automating Customer Support with AI - The speakers explain how AI tools such as automated ticket generation, product recommendation, next‑best‑action guidance, summarization, sentiment analysis, and intent classification can dramatically increase call‑center agent productivity and improve the customer experience.
Full Transcript
# AI Summaries Transform Customer Support **Source:** [https://www.youtube.com/watch?v=Vipb458S-co](https://www.youtube.com/watch?v=Vipb458S-co) **Duration:** 00:07:40 ## Summary - Customers frequently experience frustration with traditional call centers due to lengthy navigation menus and agents lacking context about prior interactions. - The speakers propose leveraging generative AI (large language models) to improve the experience by automatically summarizing past call transcripts for agents. - AI can also perform sentiment analysis on previous calls, giving agents insight into whether the customer’s prior experiences were positive or negative. - Intent classification by the language model can identify the primary reason for a customer’s call (e.g., product inquiry, billing issue, promotion interest) before the agent speaks. - Retrieval‑augmented generation (RAG) can be used to supply agents with relevant historical information during call transfers, ensuring continuity and better service. ## Sections - [00:00:00](https://www.youtube.com/watch?v=Vipb458S-co&t=0s) **AI‑Powered Call Summaries** - The speakers discuss using generative AI, like large language models, to summarize call transcripts, perform sentiment analysis, and classify intent, enabling agents to quickly grasp customer issues and improve the support experience. - [00:03:06](https://www.youtube.com/watch?v=Vipb458S-co&t=186s) **Using RAG for Real-Time Agent Support** - The speaker explains how Retrieval‑Augmented Generation can turn speech‑to‑text inputs into LLM‑driven answers, instantly equipping call agents with expert knowledge and eliminating costly transfers and extensive training. - [00:06:16](https://www.youtube.com/watch?v=Vipb458S-co&t=376s) **Automating Customer Support with AI** - The speakers explain how AI tools such as automated ticket generation, product recommendation, next‑best‑action guidance, summarization, sentiment analysis, and intent classification can dramatically increase call‑center agent productivity and improve the customer experience. ## Full Transcript
Sai, have you ever called customer care
and ended up being completely frustrated?
Yes Sharath, all the time.
First of all, when I call the customer care,
getting to a real person is an impossible task because
I have to answer a whole bunch of questions,
need to press a lot of keys before even I get to a real person.
Even when I'm talking to a real person,
the agent wouldn't understand why I called,
my history or any of those details.
So, overall, it was a real poor experience.
Yeah, I've had the same poor experience.
Wouldn't it be really cool if we can use generative AI
to help the agent and make it a much better experience
for the end customer?
Generative AI, that will be really cool.
How can we use generative AI in such situations?
So we can use LLMs, or large language models,
to do a number of different things, such as summarization.
So let's say we take a previous transcript,
call transcript, between an agent and a customer.
We run that through a large language model,
and the large language model can then generate a short summary
of the entire long call transcript.
Okay, so the agent will be able to understand
why the customer called in the previous instances
without actually looking at the whole transcript,
but instead looking at just a summary transcript
that is provided by the LLMs.
That's right.
We can do a couple of other things with those previous transcripts.
One is sentiment analysis.
And the third thing is intent classification.
Okay, so the agent already knows in advance
what kind of experience the customer had in the previous instances,
whether it was negative or positive experience.
That is good information to have
before the agent picks up the call and talks with the customer.
But can you explain a little more about
how intent classification can be utilized here?
Sure, so we can look at this previous call transcript
and then we can classify it as what is the main reason
or intent the customer has called.
So this could be things like
maybe the customer's calling to ask about a particular product
or a billing issue.
Or, let's say there's a recent promotion
and wanted more information about that.
So the large language model is able to look at the transcript
and determine what is the main intent for that conversation.
Oh, that'll be really great because
even before the agent talks to the customer,
picks up the call and talks to the customer,
the agent already knows a lot about the customer.
Knows the summary of previous conversations,
why the customer call in the previous instances,
and also the kind of experience the customer had.
So that'll be good information to have when
the agent is talking to the customer, so that he can
tread carefully when talking to that specific customer.
That will be helpful.
That's right.
But, haven't you had a lot of times when
an agent has just switched over
or will have to transfer to another agent?
So that's where we can use another thing known as RAG,
or Retrieval Augmented Generation.
That is interesting because every time I call the agent,
I get transferred to a different agent and I have to end up
saying the thing, saying all the things over and over again.
But how does this RAG work?
As in, can the agent just type in a question and
get the responses back from the generative AI LLMs?
Sure, so instead of transferring to a number of different agents,
RAG can help any agent become an expert on any particular topic.
So that way you don't have to get transferred to another agent.
So instead of typing out the question,
imagine if AI could automatically be listening in to the conversation
so we could have the speech-to-text listening in to the conversation.
That text then sent to the large language model,
which can then bring up the relevant information
and present it to the agent
so that the agent is knowledgeable about any topic
that a particular customer is asking about.
Okay, that actually makes a lot of sense because
the agent doesn't need a lot of training on all of the things that are available
and requires a lot of lesser switching to different agents and
the agent will be able to help the customer on the call in real time.
That is good information for the agent to have.
But how does the RAG framework work?
Can we talk about that and how can it be applied in such scenarios?
Sure, so let's say there are a number of different data sources.
This could be things like product documentation.
You could have a FAQ information
as well as previous trouble tickets.
All of this is text information,
which can then be split up or chunked
and sent to an embedding model.
This embedding model can then convert all of this text
into embeddings or vectors.
Really it is just numerical information
which can then be stored into a vector database.
So now when a user is asking a question,
this vector database is able to understand the semantic information
and then bring up the most relevant content,
send that over to a large language model,
which can then generate an answer
and send that back over to the user.
That actually will help the agent in a lot of scenarios because,
as we've been talking about,
the agent doesn't know anything and everything.
So having the generative AI LLMs bringing up
the learned information real time
as the agent is talking to the customer will be really valuable.
That's right.
So we can also do a number of other things.
So let's say, you know, in some cases
it might still require that a particular agent
needs to send some information over to another system.
So think of trouble ticketing systems
where you can have a large language model
automatically pre-populate all of the different fields
in this trouble ticket form
when it makes it easy for the agent to then just review that information.
Yes, that will be that will be really helpful because
usually the agents spend a lot of time in taking notes,
creating those trouble tickets after the call is ended.
So it'll really help boost the agent's productivity and efficiency, and
all the agent has to do is just look at the automated trouble ticket
that is created and just review it and update it if needed,
and submit a trouble ticket.
That saves a lot of time.
Exactly, yeah, and then we can also do a couple of other things.
So we can do things like product recommendation,
where a large language model can automatically recommend
the product based on the particular customer.
So that can be also personalized.
We can also do things like next best action,
where we can tell the agent what is the next best thing that
the agent needs to be doing while on the call.
So this can really guide the entire conversation between
the agent and the customer.
Well, there you have it. So with all of these things -
summarization, sentiment analysis, intent classification, RAG,
and these kind of generation tasks -
the agent will now be able to talk to the customer,
and help the customer in a more productive fashion.
Right, so next time you call a customer care
hopefully you won't be as frustrated.
That'll be really helpful, thank you.
If you liked this video and want to see more like it,
please like and subscribe.
If you have any questions
or want to share your thoughts about this topic,
please leave a comment below.