- il y a 2 jours
Live agent revolution: build your generative AI assistant in a few minutes
Catégorie
🤖
TechnologieTranscription
00:01Bonjour à tous, bonjour à toutes. Est-ce que vous nous entendez bien déjà ?
00:06C'est bon ? Super.
00:08Let's switch in English.
00:10Hello everyone, super super nice to meet you
00:14and thanks a lot for being here on this session.
00:16We are glad to have you here with James from Salesforce.
00:21So this session is going to be focused on, in the real life,
00:24how we're going to build an AI agent.
00:26And so Michael was mentioning 30 minutes.
00:29I'm actually going to take 5 minutes to introduce why
00:32we're talking about AI agents, what is actually an AI agent
00:36and then we'll take like 15 minutes, 20 minutes maximum
00:40to build these agents in front of you.
00:42So the first thing.
00:47Once again, thank you for being here on this session.
00:52So the clicker is not working.
00:54Yeah.
00:57So I'm Paul Medoville,
01:00ARVP for AgentForce, Salesforce.
01:03The first thing is, this is it.
01:05We are in the third wave of AI.
01:09And that's agentic AI.
01:11You've seen it and that's crazy at VivaTech.
01:14You have like autonomous and AI technology everywhere.
01:17And it's actually have an impact on the businesses, on the companies.
01:21And at Salesforce, we are working with a lot of companies that have launched their transformation already
01:27with agentic AI.
01:29Companies like Buick Telecom, like ADECO in the rocketing industry, but also scale-ups and start-ups.
01:37Everyone is starting now.
01:43And now, I'm able to manage this speaker. Thanks for that.
01:48So what we see on the market is actually people in the companies are overwhelmed by like useless tasks and
01:58silos in the tools they are using on a daily basis.
02:02At the same time, their clients, our clients, your clients as company have new expectations.
02:09They want and they need real time.
02:12They need contextualization.
02:14They are using on their daily life generative AI themselves.
02:18And so they want this kind of experience for themselves in the company.
02:21What it means for a company is that if we are not able to manage this new expectation and deploy
02:28this kind of technology,
02:29you have missed opportunities because you have limited capacity, but also you have a lack of ability to innovate.
02:35And that's a burden for a lot of companies.
02:37And that's actually why we launched Agent Force at Salesforce.
02:42Let's make it super simple.
02:43We deployed on top of our CRM platform, our Salesforce platform, an agentic layer to give the ability to every
02:52company, every business,
02:54leveraging our platform to leverage and deploy agents for sale, service, marketing, commerce directly in the flow of work on
03:07their platform.
03:09And those agents are able to leverage the right data and the proper data, thanks to data cloud.
03:15When we talk about data, we talk about structured data, unstructured data, data coming from Salesforce obviously,
03:21but also from external sources like Snowflake, Databricks and the technology you are already using.
03:31And what we see, and that's super important when we are talking with our clients is the time to market,
03:37time to value.
03:38Like we are not in the pilot or POC phase anymore.
03:41We are in the time where we need to see value and we need to deploy at scale.
03:48And that's really the challenge that we see at the moment on the market.
03:51And we have a good study saying that in average, when you launch AI agents thanks to AgentForce, this is
03:57going 60 times faster than when you create your own technology or are trying to build everything from scratch.
04:05So super important to mention.
04:08And a little bit of explanation for everyone to be aligned on that.
04:12What is AgentForce when you talk about agents, AI agents, etc.
04:17Basically, it's two pillars. The first one is giving the ability to a company to have the right set of
04:23tools.
04:24And when we say the right set of tools, it's basically the ability to manage the whole agent lifecycle.
04:30From the ideation, the creation, the configuration, the test, the deployment, the supervision.
04:37Because when we talk about, once again, with our companies, our partners, our clients,
04:41they want to monitor the full cycle and be able to adapt and iterate super fast.
04:48Because once again, the time to market, the time to value is the key here.
04:52Second thing is our ability in the set of tools to leverage the existing technology.
04:57The reasoning technology, making sure that agents are able to sync, plan, reason and execute.
05:04It's nothing linked to what we've known about chatbots, executing simple tasks they were asked to do.
05:13But also the ability to leverage the rag to make sure they are using the right data, contextualized and real
05:20-time data
05:21to make sure that they are contextualized, relevant and can better activate the right action in the flow of their
05:27work.
05:28And the second thing is the suites of agents.
05:31When we say the suite of agents, basically, we can leverage and our customers and partners can leverage out of
05:39the box and templates agents.
05:41Because a lot of agents we are seeing on the market are the same for different kinds of companies, different
05:50kinds of industries.
05:51For instance, when we talk about autonomous service agents, it's the ability super fastly to deploy an agent, for instance,
05:59on your website
06:00to handle questions that your customers can have to make sure that you are a support for them,
06:06you help them to find the right project, etc.
06:08So that's a kind of agent like you can deploy because that's a template, so that's super fast.
06:13Exactly the same with SDR agent, for instance.
06:16It's the ability to do prospection on your behalf, to qualify leads and to make sure that you escalate to
06:23the human once the lead is ready
06:26and the human can focus on the real value of the work.
06:31And what we see on the market, once again, is both customer facing agents and employee facing agents.
06:36Employee facing agents like to assist and augment your employees and customer facing agents to augment your teams
06:44and work on your behalf once again in the floor of work to meet your customer expectations.
06:51And I was mentioning it earlier, but we are not at the pilot and POC phase anymore.
06:56That's real.
06:57And a lot of industries, clients in France, obviously, but also worldwide
07:01are winning thanks to this kind of technology and thanks to this kind of strategy they are deploying,
07:08leveraging what we call digital label and extending their team thanks to agents.
07:12to give you some examples.
07:14I was talking about capacity limitation just earlier in our introduction.
07:20We are working with ADECO and ADECO, so you know, it's obviously the recruiting firm.
07:26They have around 300,000 million of candidates interaction a year.
07:32And obviously, they want to be able to tackle every request they have and to meet their client's expectation
07:39to find the right profile in the right proper of time.
07:43So, once again, as fast as possible.
07:47Thanks to AgentForce, they will be able to handle every candidate in a personalized way,
07:54to prescreen them, make sure that they can match the desk job they have on the market
08:00and execute super fast to find and position the right profile.
08:04And actually, when they launch it and after several months of usage,
08:08they realized that almost 70% of the candidates interaction was outside of working hours.
08:17Which means at night, during the weekends, because when you are looking for a job,
08:22that's when you are looking for information and when you are looking for a job position.
08:26So, basically, thanks to AgentForce, they were able to extend their team
08:31to make sure they could handle those interactions faster than their competition,
08:37even when their employees are not working.
08:39And that's the whole goal of this.
08:42Another example is Fairmob.
08:45You know Fairmob pretty well.
08:47You have them in most of the Paris garden here.
08:53Fairmob, they decided to handle every request they have around spare parts,
08:59looking for and buying for spare parts directly with an agent
09:03because they want to focus on the new sales.
09:05They want to focus on the acquisition.
09:08And so, they deploy an agent directly available on their website
09:12to be an advisor, one-to-one advisor with the client
09:17to find the spare parts linked to the projects they bought before.
09:19Another example.
09:21And the last example, maybe last but not least,
09:24I was talking about employee experience also
09:27and the way you can augment your employees.
09:30A company like Royal Bank of Canada, for instance,
09:34has deployed agents to support their 2,000 advisors
09:39to prepare the meeting.
09:40Because before, to prepare the meeting and be super accurate and sharp
09:45to position the best offer to the clients,
09:47they needed two, three hours to gather all the data,
09:50to gather all the information they need.
09:52Now with Agent4, they're doing it in less than two minutes
09:55because they are going to leverage all the data they have
10:00and support their advisors to make the best decision
10:04and better prepare their meeting.
10:06So, that's just three examples.
10:08but to give you some real-life example about what we love at Salesforce,
10:12it means customer stories, real customer stories deployed at scale.
10:20I'm talking about, and that's the last slide before actually seeing how you build that,
10:25how you build an agent for your company.
10:27but we're talking about agent for about 10 minutes now,
10:32but what is actually an agent?
10:34Let's think about it.
10:36An agent is like a digital employee for your firm.
10:39So, it's just thinking about you're going to hire a new employee for your firm.
10:44And for that, we have five key attributes.
10:47The first one is, what is the role of this agent?
10:51What is the purpose? What is mission?
10:53What are the topics they are going to tackle?
10:56It's obviously super important.
10:58The second pillar is the data.
11:00If I want my agent to be relevant,
11:02if I want my agent to be actionable,
11:04what kind of knowledge do I need to give him and to enrich him with?
11:11That's the second pillar.
11:12The third one is the action.
11:14I want this agent to be able to execute tasks on behalf of my team
11:19when I'm asking it but also autonomously.
11:22And that's the third attribute.
11:24The fourth one is the guardrails.
11:26We all talk about the guardrails.
11:28What are the limits we want this agent to have?
11:30Because obviously, we want to deploy,
11:32we want them to execute,
11:33but we want to put some guardrails
11:35because at some point, we need a human escalation
11:37or maybe a supervision
11:40or talk with other agents before to launch a second action.
11:45And that's basically the guardrails.
11:46And the last one is the channel.
11:47Where do you want this agent to be deployed?
11:50Is it going to be in Slack?
11:52Is it in Teams?
11:52Is it on your website?
11:54Is it in the background?
11:55Because you want this agent to work for your workforce
11:57without a human conversational interface.
12:01So that's the channel.
12:03And obviously, it's super important, the trust layer.
12:05And that's where everything is built on Salesforce
12:07to make sure we have the right setup for data privacy
12:10and data security.
12:13And now, my suggestion is to go directly to how we build
12:18in the real life these agents.
12:20And James, please show us how.
12:23Okay, thanks Paul.
12:25Hi everyone, my name is James.
12:26Welcome to today's Agent Force demonstration.
12:30So today we're going to walk through the process of creating
12:35a smart, autonomous agent within Salesforce
12:40that's capable of answering employee questions
12:43around company travel policy.
12:46So we'll cover four main steps.
12:49The first step will be to define the scope of intervention
12:53of the agent.
12:54The second step will be to define the roles
12:57and responsibilities of that agent.
13:00Thirdly, we'll create what we call the topics and the actions,
13:03including a RAG, so a Retrieval Augmented Generation powered
13:09lookup action that will be capable of basing requests
13:13and answers on the company's knowledge base.
13:17We'll then deploy the agent and then we'll give it a test run.
13:21So today we're going to use this Agent Creator tool
13:25to create the agent.
13:27And I had planned on actually giving the instructions verbally,
13:32but I think with the ambient noise around,
13:35it might be simpler to do it in the written way.
13:42So we're going to come into the Agent Creator.
13:44And as I mentioned earlier, the first step is to define
13:47the scope of intervention of the agent.
13:49So, you know, what the agent is going to do.
13:52So we're going to start that process here.
13:59So here I'm interacting with an agent.
14:01So basically there's an agent that's going to create an agent.
14:04That's what we're going to see here on the screen.
14:06And I've put some instructions here for the agent.
14:08So I'd like that this agent be for internal employees only.
14:15I'm going to define a descriptor of the agent to explain that it's to answer travel policy questions.
14:24And I'm also going to mention to the agent that if the agent doesn't know the answer to the question,
14:29then it shouldn't invent anything.
14:31It should escalate that to a real-life human agent.
14:36Okay.
14:37So here we can see that the agent has taken into consideration the different instructions that I've given.
14:46And the next step, once that's done, is to actually create what we call the topics and actions.
14:55So the topics are a collection of different actions that the agent can execute based on the request.
15:06So it might be that you've got various different topics.
15:10It might be that you've got various different actions.
15:12And we're going to group those actions together under the different topics and categories.
15:18So we can type in the next.
15:20I'm going to just say that we're okay with what's being proposed here in terms of roles and responsibilities
15:25and scope of intervention.
15:36So again, we're interacting with the agents in the background.
15:39He's going to execute various different actions in order to take into account the exact instructions that I'm mentioning.
15:49Okay.
15:50So now we're going to move on to the topics, as I mentioned.
16:03So these are the roles and responsibilities.
16:05So this is the actual scope of intervention.
16:07And again, this is really important for Salesforce in order to understand that the agent should only intervene in specific
16:16circumstances.
16:18So in the background, Paul talked about it earlier, we have the Atlas layer.
16:23So the reasoning engine, which is capable of determining when an agent should be called and under what circumstances.
16:29So here is the agent has described the different roles and responsibilities, as I mentioned earlier.
16:33And he's suggesting to create the topics and associated actions.
16:37So we're going to go ahead and do that.
16:43Okay.
16:44So basically, I'm saying, okay, let's define, okay, so we've done that.
16:49Sorry, that's the wrong one.
16:56There we go.
16:58So here we're going to create the topics as I mentioned.
16:59So basically, we're going to create a topic called travel policy.
17:02So we'll cover all questions around travel policies.
17:06We're also going to request that the agent creates an action that will take the input.
17:12So the question that the end user is requesting to the agent.
17:18And then it will go and look against the company's knowledge base in order to provide an answer.
17:22So in this case, we're looking at travel policies.
17:24But the important thing to remember is that here we're using travel policies.
17:28But this use case could be applied to literally any kind of use case, any kind of scenario and situation.
17:38So here I'm saying, let's go ahead, create the topics and actions.
17:45Again, in the background, the agent is determining what exactly should be done.
17:49So it's going to create the topic.
17:50It's going to create the RAG lookup in the background.
17:56And this is all being done through Salesforce's AI layer called the reasoning engine.
18:05So here we can see the different instructions given to this particular topic.
18:12So if an employee asked about the travel allowance, provide them with specific details.
18:16If the employee's questions cannot be answered, would the available information escalate to a real human agent?
18:24And here we've included some sample utterances.
18:27So these are the kind of sample questions to give the AI an idea of the kind of questions that
18:33we would potentially be requesting.
18:38So now we've created the scope of intervention, the roles and responsibilities and the topics.
18:44The last step is to deploy the agent.
18:47So we can say to the agent to go ahead and deploy what we've created.
18:56So here is a step to ask whether I'm happy with what we've created.
19:01And here, because all the data is stored as metadata, I can deploy to any Salesforce environment.
19:10So here, for example, I've got my test environment and I can start the deployment process.
19:16So this will process in the background, deploy the metadata and the agent will become available within Salesforce.
19:22So now I'm going to give it a quick test run.
19:24One thing to keep in mind is that the agents can be deployed almost to any channel.
19:29It could be deployed to Salesforce directly.
19:32We could deploy it to an Experience Cloud website.
19:34It could be deployed to your company portal.
19:37And here we're going to, we've deployed it to WhatsApp.
19:40So we're going to use WhatsApp as the channel for interaction.
19:47Okay.
19:48So I'm just going to initiate a conversation with the agent.
19:53Again, in the background, Salesforce's AI is taking over and deciding which is the best course of action to follow.
20:00And if the demo gods are with me, we should get a response from the agent.
20:06Okay.
20:07So the agent has been evoked.
20:10He's asking if he can answer any questions.
20:12So here I'm going to ask a question about the travel policy for, let's say, London.
20:38So in the background, again, thanks to the scope of intervention that we define the roles and responsibilities, the topics
20:44and actions,
20:45and Salesforce's AI is determining which is the best course of action to take.
20:49And in a second, we should have a reply from the agent based on the company's knowledge base around travel
20:56policies.
20:58So here we can see that the travel policy for London has been displayed on the screen.
21:04I'm going to ask one more question around travel allowance for taxis.
21:24So this is a much more specific question.
21:28And again, the agent is going to determine that the correct course of action to take is to execute the
21:34travel approval, the travel policy action.
21:37And here we can see that we can claim up to 25 pounds for the taxis.
21:44And because of all of this information is in Salesforce, you can actually build dashboards out on usage and how
21:51users are interacting with your agents.
21:56So that's the end of the demonstration. I'm going to hand back to Paul to do a quick wrap up.
22:04Thank you very much, James. Can you hear me well? Yeah.
22:07Thank you very much, James, for the demonstration. And guys, we hope you enjoyed it.
22:11If you have three points to have in mind as a wrap up for this session, the first one is
22:17it's super important to define the right use case.
22:21What is the use case that's going to bring the most value in your organization?
22:25What is the simple use case you can deploy because you have access to the data, because you have the
22:28processes?
22:30And that's the first one. Find the right use case to prove the value and accelerate the time to market.
22:35The second topic is you have everything with the platform to deploy and monitor your agents.
22:43And let's be honest here, we are all learning on this new path.
22:46So the goal, once again, is to start, to learn, to iterate.
22:50And that's why it's important to have a partner that gives you all the tools to create, configure,
22:57supervise, test and monitor your agents.
23:00And the third point, super important, is to start now, obviously.
23:03As I was saying, a lot of companies are on this path already.
23:07A lot of companies are deploying at scale.
23:10We are not in the pilot and POC phase anymore.
23:12So let's start.
23:15And if you have any question, we'll be more than happy to answer.
23:17Thank you, guys.
23:22Thank you. Thank you very much, guys.
23:24A couple of minutes for questions, if anyone has a question.
23:28For the guys. No? One question here.
23:30Thank you.
23:35One question.
23:38Yes, a quick question. Thank you for the demo.
23:43So, but I assume if you're training, if you're creating your agent based on the company data,
23:48anytime your policy is changing, or there is a tweak, you have to make an adjustment.
23:53Can you talk a little bit about that and how not to create a bottleneck so that that could be
23:58also done automatically and in real time?
24:01Sure.
24:01Maybe I can talk about that.
24:03Would you want to take it?
24:04Okay.
24:06So it's a very good question, first of all.
24:09So the first thing that I'd like to mention is that Salesforce will integrate with your preferred LLM.
24:17So whether it be OpenAI, Anthropic, you can even bring your own LLM.
24:24And that's what we're using in the background in order to provide context.
24:30So business context, that's where there's a real added value to Salesforce's Agent 4 solution.
24:35We're providing business context to your preferred LLM.
24:39And that's what we're using in the background.
24:41And we have that, as I mentioned earlier, that layer called the reasoning engine, which is Salesforce's AI,
24:47which is determining the correct path in order to answer a question or perform an action.
24:54And an important thing to remember is that the data is not persisted.
24:59So your business data is not being used to train the LLMs.
25:04That's the first thing.
25:06Let's come back to your question around the RAG and the vectorized database.
25:10So Salesforce through data cloud, because data cloud, we haven't mentioned it,
25:15but data cloud and agent force are very tightly coupled.
25:18Through data cloud, we have access to the retrieval augmented generation and a vectorized database.
25:25So you can ingest data, whether it be structured, semi-structured or unstructured data,
25:32from whatever the data source might be.
25:35It might be SharePoint.
25:37It could be any kind of repository.
25:39And then when that data is ingested into data cloud, we then perform the process of vectorizing that data.
25:48So any updates can be pushed into data cloud.
25:52All right. Thank you.
25:53That's all the time that we have, ladies and gentlemen.
25:56Thank you, guys.
25:56Salesforce.
25:57First pause.
25:57.
Commentaires