# OpenAI DevDay 2024 | Community Spotlight | Parloa

## Метаданные

- **Канал:** OpenAI
- **YouTube:** https://www.youtube.com/watch?v=xZc0YQbIyWE
- **Дата:** 17.12.2024
- **Длительность:** 14:58
- **Просмотры:** 14,515
- **Источник:** https://ekstraktznaniy.ru/video/11367

## Описание

Transforming Contact Centers with GPT-4o Multi-Agent Crews and Human-in-the-Loop: Building agents with OpenAI o1 and GPT-4o for automation, quality assurance, and human-in-the-loop solutions.

Presenter: Maik Hummel, Principal AI Evangelist, Parloa

## Транскрипт

### Segment 1 (00:00 - 05:00) []

so my name is Mike um with AI and I work for a company called paloa um we are doing contact center Automation and uh just imagine you're calling into a telephone hotline um how it sounds nowadays is usually press one for insurance press two for contract um and that's not a great experience right um so today I want to talk about how the contact center industry is uh being transformed how it is being uh shaped in the near-term future by using um open AI technology especially GPT 40 with multi-agent cruise and uh the human in the loop integration to make sure we can use it safely we at po we believe that every customer interaction should be as easy as talking to a friend so just imagine you're calling in you want to have it natural you want it to be trustable you wanted to be safe um and this is what we call a personal AI agent so every conversation is unique it's not just deflecting the calls it's actually making sure that we resolve the issues um if you look in nowadays world we do have contact centers all around the globe all around the world there are human agents and customer care Representatives at is being called working on different levels these are all human agents who work day in day out in a call center and that's actually not an easy job I've been visiting call centers in Malaga in Cairo talking to the agents there it's a really hard job um to do um if we look into the future we want to support that kind of work by introducing um AI agents to the workforce of human agents that are working in contact centers and this does not mean that we are replacing human agents doing that job we will complement it because you saw it in the demo from Roma he was ordering pies but it wasn't him who was doing the call he instructed his assistant to do it on his behalf now imagine your personal assistant that you have in your pocket your series your Alex say they all doing phone calls not only to one place but maybe negotiating a p price with 10 different vendors who's picking up the call on the other side it can be human agents it won't be possible we cannot cover the whole Market we won't be able to manage that inflow of call volume and this is where AI agents can be a solution to also be able to pick up these calls on the other side the challenge that comes with it is we do have actual customer interactions we are talking to our customers and we care about our customers so we need to come up with new Solutions in how to put this technology safely and responsible into production and this is where the AI agent life cycle comes into play it's not just about building a hello world example or putting it into production without carrying about the performance it's actually making sure the whole life cycle is supported and this is where we earlier this year in September launched um our AI agent management platform um which we were working on for one and a half years um it was quite not to talk about it um so we needed to keep our mouth shut and we then launched it in September and finally we are able to talk about it we tried to cover all the parts that you need to have to properly put it into production this comes with design and integration means I'm building an AI agent I prompted with uh natural language briefing and I also integrated to thirdparty tools because I need to interact with the outside world fetching information and putting data back to the systems where they belong to and making sure that we test it properly testing changed so we were coming from an ivr world this is like intent classification natural language processing we are trying to take a word put it into an intent and then try to navigate along the path very fixed very easy to test a happy path but now we are dealing with a world that is non-deterministic so we need to change the approach and the approach that we are choosing is simulation and evaluation making sure that we can actually test it and the first two parts will be the key focus of the session today but for sure we also need to make sure that we deploy and scale it why in contact centers and in customer service we usually see large spikes of call volume so one day you could have thousands of calls more than on the other day and you need to make sure that using large language models can deal with the amount of calls that are coming in and on the other side we want to derive kpis and data quantifiable data of what's working how many people are calling in and this is the last step Monitor and improve and then the whole process starts again in order to make sure we increase the autonomy crate of the AI agents that we are using um coming to the first part the design and integrate part um this is how it looks like when we are building this AI agent it's a natural language briefing you can compare it to the way um how a human agent starts their job when they are on the first day like they are being told what company are you working for um what products do you support how do you handle the cases when somebody is calling in and you're prompting that with natural language so the whole way on how we build these kind

### Segment 2 (05:00 - 10:00) [5:00]

of experiences changes as well we are using natural language we are no longer programming it becomes much easier also for subject matter experts to actually um work on AI agents the thing is it's hard you all know prompt engineering um is a science for itself um and it won't become easier in the newterm future so at paloa we set out to support especially the initial stage of getting from zero to one when you are having a case that you want to support and this is where a multi-agent prompt engineering crew comes into play um where we are trying to support our conversational designers um how that works is we are replicating the roles that we usually have in our teams um and are trying to do the whole process with autonomous gbt 4 based agents so we have a vice president customer success with a lot of experience making sure to navigate at different work streams and we have two people collaborating the solution engineer which is focus focused on the integration part how do I get systems connected and a CX designer who is focused on how to make the best conversational strategy overall This Crew is working collaboratively and the manager makes sure that it meets the goals that we set out with best practices with prompt examples and all the best practices that we see in the field that we always reincorporate back and in the end we are generating a Jason configuration that is being used by our M Studio the thing is this is a recursive Lube in itself so the agent designer at the conversational designer also needs to potentially instruct back about improvements about things that have been missed out or about a use case that we didn't uh that we didn't consider in the beginning and then in the end if we are happy with the first draft with the first design that our um crew came up with we need to simulate and evaluate it to go a bit into technical details um our configuration structure is a bit different to the traditional prompt that you're seeing when you're plainly prompting why we need to support specific parts for telepon a classical use cases for example handovers telepon handovers where we want to hand over to human agents for specific use cases and this is where a zip tool for example comes into play we're also structuring the prom between the Persona um the different kind of descriptions that we need to pass for the parameters where we want to pass meter information to third party tools and all of that needs to be considered so it's much more complex than just um the normal texture prompt um and then we have the rols the RO descriptions where we opted for the yl way of describing it to make sure that also our solution Engineers themselves can easily collaborate all of this is based on the crew AI framework crew AI primarily because of the usability and the developer experience so that we can Loop in more people from CX or from our own solution engineering to help with that process and now we end up with a prompt the challenge is how do we know that a briefing that we generated and the briefing that we came up with actually does what it is supposed to do and this is where simulation and evaluation comes into play where we need to test um the reliability of the prompt so we have the conversational designer um we have a configured AI agent configuration and on the other side we have actual customer conversations so we're working with call centers um together so we have sample data from actual use cases in the field and we're taking these to derive customer Persona so just imagine we have an angry caller we have a child that is calling in an elderly person it's completely different personas that are calling in and we need to make sure that we simulate all the different Varian in order to make sure that our AI agent is simulated with the different customers and can handle these cases according to the briefing that we provided in the agent configuration all of these conversations are being evaluated with contact center specific evaluation criteria and in the end this generates insights for the conversational designer to make sure um that we improve over time the whole configuration and we make sure we are compliant with the goals that have been set out um by the company this is how it looks like in the UI so we have a we abstracted all the complexity away for our um customers themselves um they have an easy way to run simulations to configure the personas that they want to test how many iterations you want to do how many simulations you want to run so you can think about it like running thousands and thousands of conversations so that we can properly quantify the reliability of the large language model and make sure that it's safe to use also for production use cases and the evaluation criteria in itself is a mixture of the technical accuracy it's essentially like an endtoend integration test like we know it from software engineering um where we can also test for API errors because sometimes it's not even the AI agent who is at fault sometimes it's the API from the customer who is sending a bad response or that is not reacting properly and uh on the other side we have the language Behavior

### Segment 3 (10:00 - 14:00) [10:00]

itself how do we handle the conversation do we properly act for example with the corporate wording do we recommend a competitor if you're asking uh specifically all of the things that we need to have for compliance we can make sure that we either pass or fail that criteria the thing is in real life not every customer interaction can be handled autonomously this is how it is nowadays so we need to come up with a solution on how we handle conversations that cannot be fully autom at it um and at paloa we build a solution for that where we incorporate the concept of an AI agent with a human agent how that works is we hand over that call to a human agent but the customer and the human agent they don't hear each other the human agent sees this interface you're seeing right now the whole conversation is being transcribed in real time and the AI agent proposes proper responses to the human agent you saw these little blue buttons the human agent can just select a proposed response if the human agent is not happy with the proposed response they can just still speak to the phone and everything the human agent says is also being transcribed and then put back into the jml context that we use for the AI agent so the human can chime in control the conversation if need be and the AI agent will adopt that kind of guidance or how the conversation flows in the next turn and propose for example a better suggestion next time um this way the agent is um supported in their job um and the good side effect that we have by making sure that the human agent and the consumer they are not connecting directly to each other we can even Bridge the language barrier at the same time so this is completely language agnostic so a Swedish caller can call in and a German human agent can support them and we can make sure that we use large language model to accelerate the whole process and still make sure that the latency is as low as possible and this is also Al super interesting for customers who are a bit objective against generative AI usage or who have concerns this way we can make sure it's safe to use and it's always being supervised by human agents who have the experience on how to handle these kind of PES why is this important um if you look into the past um this is how the world looked like we have a lot of ivr systems press number one press number two and a lot of human agents Junior agents senior agents working on the different levels um with the emerge of AI we are entering we entered the AI transition phase where we came up with rule based AI agents mostly based on natural language processing intent classification in Parts but we still had a lot of ivr and Junior agents and Senior agents who have to handle that right now we are entering an era where we have autonomous AI agents um who are taking over more and more chobs who can complement for now the rule based systems and who will make sure that we are slowly but steadily getting rid of the ivr solutions with the bad customer journey in itself if you look into the future we will enter an era where it's called the AI first contact center and um funnily enough it was in one of the podcasts where Sam mman was uh talking to Lex Freedman and he estimated round about a year ago that potentially contact centers might be one of the first Industries being completely disrupted he was right um this is currently happening so we are rolling out that solution with call centers and contact centers around the world and if you look into the future when we have more capable models more autonomy um and more ways to connect to third party system more instruction following um we will have autonomous agents dealing with the cases where the customers are calling in and fully autonomously resolving them end to end even integrating to the systems and automating the whole process the important part is we still will have human agents and the differ to how they work nowadays and how they work in the future is they will be promoted essentially to become senior agents and AI coaches or supervisors so the future that we envision is that contact centers of the future will help AI agents to do a good job and the human agents will become the supervisors for all of these AI agents that's a future that we envision and um at paloa we believe that the leaders of this era the companies who try to do good contact uh and customer service um they will provide a personal AI agent for every customer um and um my last take for that round here let's prove to the world that uh large language models can have a good benefit um to society to humanity and also consider um the organizational changes that are about to happen in the past and help everyone adopt to it um thank you for taking the time and see you