AI Employees Outperform Human Employees?! Build a real Sales Agent

Key Takeaways

thanks to hopspot for sponsoring this video this new AI can perform all coding jobs in seconds including blockchain development while this AI is already outperforming accounting firms meanwhile the new graphic design AI aims to automate graphic design and could minimize the need relax CU in a future where AI does most of the work there'll be one thing that humans will finally get to do all all day long nothing so it's time to learn the most essential skill of the future five star presents nothing University with the first ever diploma program and doing nothing an intensive course covering the most Future Ready skills out [Music] there but let me check with the and get back to in a few days with our state-of-the-art workshops campus placements and guest lectures the world can finally prepare for the Glorious future that awaits us eat five star do nothing I saw the snack company ads the other day on Twitter and I saw it this brilliant to Showcase what kind of impact AI might have on the job market so every time when there's a new technology emerg it Reise people's job responsibility and create a huge transition in how we work and AI agents has been one of the key Concepts that can deliver this level of automation the concept of AI agent became really popular mid last year with project like Auto GPT where people can just give it a high level go where Auto gbt will break down this big goal into subtasks prioritize what's the most important ones execute it and then reflect what other new task I should do based on information I collected and it was extremely exciting idea but due to a lot of limitations most of people just have this impression that it feels more like a tool instead of a poer product with the release of Devon it kind shift people's impression of AI agent it claimed to be the first AI software engineer who can have access to different tools like browser command line to Cod the entire purchase taking and complete freelancer software engineer work on upor and even automa find you an AI model that can be used for itself to be honest Deon was not the first AI software engineer attempts there were multiple purchase came before like chadf where they provide a framework for multiple different agent to work together to deliver a software or other open source project like GPD engineer and small AI but Deon is the only fully patched AI agents that make you feel like you are dealing with the real capable AI employees who can actually deliver results cuz when you look into it Deon actually packaged a lot of complexities into this AI agents it has a kind of building code interpreter where it can access the computer and generate code and save files and also have a browser to access information and test code that it implemented and all those Innovation I think really made people realize that the era of AI virtual employee might come earlier than we saw software engineer probably just the first of a Kind AI employees that we're going to create but there could be so many more and there are actually not that hard to build I was lucky enough to be able to work on a few AI employee project and deploy it for actual business so today I want to share my learnings about what does it take to actually build such AI employee and in the end show you an example of how can you create a real AI employee who can autonomously do marketing for you on Reddit so whenever when someone post something on Reddit that is relevant to your keywords the agent can be wake up post a comment and add a plug to your product and service but before we dive in I want to introduce you to a research the Hop spot release so I always like observing how people are hacking chat GPT to get certain job done because it is great way for me to find product and business ideas if I can find use case that even just 1% of people are hacking with chat gbt then I can productionize it and build a micro sess around it I already see a few people try and start getting some early attraction to their Niche AI product this why this research hopspot did is really useful they talk to people from Top firms about how they are hacking chity today at work from lead generation to support ticket resolution and there are some use case that I never expect before like assist with scheduling meetings and appointments or create project timeline and road map it definitely give me a lot of good idea and Inspirations of potential AI parb so if you are also like me and want to find inspiration for potential business ideas I definitely recommend go ahead and have a look you can click on the link in the description below to download this report for free now let's get back to AI employees I want to firstly share a few learnings I got from building and deploying AI employees in real business setup so for the past few months I've been building a sales bdr agent so in bigger software company normally the sales function are divided into multiple different roles where bdr represent for business development rep whose role is Outreach follow up and eventually book meeting they will hand over to the account executive which are the typical sales that we are referring to and the reason we start building the sales bdr agent is because the company I work in is quite small and we only have one sales person who are doing both bdr and account executive work that's why we thought why not just let's build a sales bdr agent who can handle the whole process from receiving a leads from hubs Salesforce to actually book a meeting we break down the bdr agent into different job responsibility that is supposed to handle and it has multiple different aspects every time when it receive a new leads it should try to do some research about this perspect to understand what does this company do and what a value hypothesis could be then draft a personalized message out to them and also follow up with perspect because normally they won't respond immediately but if the perspect did respond they should be able to respond message answer questions they have logging all the information in CRM and of course book the meaning but the problem we quickly face with is that this process is actually quite complicated even just for one part like inbox manager to build an agent that can autonomously reply and triage your email itself is not easy I actually made a video before about inbox agents where I talk about how you can build atom's inbox agent and even just this task itself can require a lot of promp engineer to actually get a large L model and agent follow the process that wanted to follow if just one step is so difficult how can we get bdr agent handle so many different complex tasks and this kind of led to one of the paradigm shift I got last year so initially when I think about agent in my mind I'm thinking about something like a super agent who is extremely powerful and can do multiple different task by itself more I build agents more I realize that the possible pass might be many different specialized agent work together to deliver the results instead it has a lot of benefits it's easier to update because you can just change a specific agent instead of changeing the whole system but also this system can compound as we build different AI employees many of the fundamental ability are similar for example like ability to answer question about our product service or the ability to schedule meetings all those capabilities can be used for multiple different job functions like sales marketing and support that means a lot of things with beautiful one agent can be rused for another so building cost will reduce significantly and that's exactly the path we choose to implement the sales bdr agent even though for users it feels like just one bdr agent is completing the task but behind the scenes there's a group of Agents working together so this manager agent who is responsible to categorize and dedicate task to different sub agents as well as completing key actions like send receive emails create a calendar event and loging things in CRM and then we have a list of specialized agent who are trained for specific jobs one person researcher who has access to different data source like Google Apollo LinkedIn to be able to really understand who this perspect is and figure out a value hypothesis then we will also have a inbox manager who is trained with our company data drop the response back to any questions that customer potentially will have and third we will also have follow-up agent who has been trained with photo op Cadence that has been proving work really well so whenever the manager agent receive a lead from CRM then it will start creating list of different task delegate to those sub agent to complete task and take actions based on the results and here's quick demo so the agent will wake up whenever we're receiveing new leads in our CRM and then the manager agent will start dedicating task different sub agent at the beginning it will dedicate research task to the researcher to understand what kind of value hypothesis about this perspect send out email and also log information in our hopspot then the customer will receive a hyper personalized message based on the research results and if they respond back with certain questions the agent will also be able to answer it so you can see after the email has been sent back to it it will trigger a new message to this agent about this latest email reply then this agent will start dedicating task to inbox manager who will be able to retrieve relevant knowledge and if the customer is interested and want to talk then the agent will be able to look at or sales team's calendar availability and proposed by the earliest time available in the end after time is confirmed then it will will create a calendar event and send to both sales and customer and this screenshot I just took from our sales team we be internally deployed as bdr agent and already deliver roughly 20% of meetings to our sales team and after deploy such AI employees in company for a while there few things I found particular interesting of adopting AI employees versus normal human employees one is it is radically transparent when a human employee make a decision or deliver certain works that didn't really meet your expectation it is not exactly clear why why he made that deis but with AI employee I know exactly why he made certain decision because I can always rewind back and figure out what are the steps it took that lead to this conclusion then I can take specific actions by adjusting prompt to align this agent better than what I want or best practice so this was extremely useful and secondly there also no communication barrier as we add more and more team members into this group unlike human employees when you add a new member they Nate communication barriers which create politics and misinformation with AI agent employees all of them can share the same kind of memory or brain and always on the same page so there's really no communication barrier and also make it very easy to compound and scale because we can just add more and more team members into the group and increase job responsibilities so I'm definitely a big believer for the future that company will start adopting AI employees and the best part is that it is actually not that hard to build you can spin up AI employees very quickly and I want to show you example of how can you create AI employe EMP in just 10 minutes so I often heard people saying that R is a great place to Market about your product or content but I never get a chance to do so I really want to create a AI redit marketer who basically can be wake up every time when someone create a redit post uh that is about AI agents and then I want to trigger this agent who can look at this Reddit content drop a message that is relevant to the audience provide some value then soft plug the product or service that I want to sell in my case will be my AI YouTube channel and to achieve this flow create a group of Agents one is a redit postf finder so this agent will be triggered like every hour and try to search what are the top five most relevant Reddit posts that is relevant to the keywords or topics of product service I want to plug in then for each Reddit post it will send to a comment draft who is the agent that we train specifically that can fat the rid post content and have knowledge about the way I drop message as well as the product service that we want to sell and in the end I have a thir agent called guy who is responsible to actually post this combat batter rdit and I also give a special function called weight cuz RIT API has pretty strict rate limit so if you just keep calling the API in point for post command they will block you pretty soon and to build this red marketing team we will need to use a multi-agent framework and there are quite a few of them two most popular one is autogen from Microsoft and CI while Auto provides most flexibility it is a little bit more technical to get started while crei gets really popular because it is very easy to set up so I will give you a quick go through about how I set up this team of agents in crei and in the end for people who are less technical don't worry I also show you how can you set up this exact multi-agent system without much coding on platform like random so let's get it and firstly to do RIT integration you will need to create a radit app and it is really easy to set up all you need to do just firstly vit red.com PRI slapps you should see a section called create application so you can give a name like marketer and then give a description and then you will need to have a redirect URL you can just put any website URL here select I'm not robot create app then you have this app created and the credential you will need to copy is one is a secret another is a client ID so these two things will allow you to authenticate with Reddit API and start fashing content from Reddit and now we can go back to visual studio code to set everything up so firstly let's create aemv file and put all the credential in which including the open AI API key the red client ID which is this value you have here and then the client secret which is this value and and you can define a user agent so this one is more like reference just put random things like your app Name by your user ID and in the end you'll put your Rand username and password cuz we will need those information to be able to be able to post Ric command on behalf of you so we can save this and then we'll start setting up the agent system so the way crew AI works is that you will Define a list of agents in our case where I will create class called R agents this one agent called RIT post finder whose role is go and find most relevant RIT post from past hour related to by topic and then I will have another agent called RIT command writer so this agent that is going to take the rant post to be able to fetch actual content and write a actual comment based on the information I was given and it will have access to tools like f r post content which we were Define later and verbos equal to True means that we will be able to see the agent SS and in the end we will create a red Command Post poster so this is the last agent who will be able to take the command that generated by the command writer and then post un rdit on behalf of you so that's pretty much the St agent we need and on the out side crew AI also allow you to define the task where you can give more details so I will create another file called task. py give a Reddit task class and then there will be one task called search recent Reddit posts where I will pass on the things to promote or which is a product and service that I want to plug into the redit post so it will be a task that has description passing the things we want to promote and also give the go and second task will be the draft RIT comment and third will be post RIT comment so this allow us to add in more specific instruction details if you remember when we create agent we actually give it them some special Custom Tool like search RAC and red post or post R comand and this is something we can Define by creating folder called tools inside tools you can create different python file to contain those specific functions I create one called redit tools. py so this one will include all theun functions that we need to be able to interact with Reddit so our first St import all the credentials that we saved in EMV file we're going to use Pro Library which is python library that allow us to interact with Reddit and then I'll create different tools one tool will be search R and Reddit posts I will pass on the keyword then it will try to search what are the relevant redit posts that included those keywords for the past hour and maximum five posts and then return the list like this and after that there will be different tool for me to F the content of that specific post based on the submit ID last to would be post command on Reddit post so this will take actual message we generated as well as submission ID and post on Reddit this is all the function we need to actually interact with I create another python file called utility tools. py this is where I want to create a utility tool called weight because RIT actually had pretty reject rate limit uh so if you keep posting it will ban your account pretty quickly I want the agent to mimic the human behavior by calling a weight function where it can randomly wait for a certain amount of minutes before it the next content so that's pretty much it in the end we bring everything together I will create another file called rit. py This is where we will connect everything together so I will import all those libraries then F agent and task I we Define in the agent. py and task. py I will create a variable called s that we want to promote in my case will be the AI YouTube channel that I'm running then I will set up agents and also set up tasks the end putting everything together kick off this team of Agents so that's pretty much it if I round this by calling python rdit py oh one thing I forgot to put in here is expect output so if we go back to task for each task we can actually Define the expected output format which is quite useful so for the search recent Reddit post task I can return exped output which is array that contain post title URL submission and reasoning and same thing for the draft redit command as well as post redit command so if I try to run again now you can see that it start triggering the agent the manager agent delegate to the redit postf finder with the specific task that we have in mind and you can see the agent start searching for certain keywords like AI Co calling and return list of different Reddit posts it will do multiple rounds of research until it find the most relevant uh RIT post and then dedicate those most relevant RIT post to the coworker which is RIT command writer who will compose the actual message in the end the Reddit command poster will be able to post this command on Reddit on behalf of me so this is how you can quickly set up a team of agent working together and deliver some useful work for you but one part I found quite difficult When developing a agents with cre AI is that I don't really have a good way to debug or figure out ways to improve the agent as it was not easy for me to track the input and output that's why I do often use Rance AI platform which is nood platform to build AI agents to set up the agent system and I can quickly show you how you can do that here is my setup I will still have three different agents one is Agent that actually going to post comment on Reddit second is Agent that are going to find what are the relevant Reddit posts sech is Agent that actually going to write a comment for me the reply agent almost is like a manager agent it will firstly dedicate a research task to RIT post finder to go and research RIT post from pass hour that can be relevant to the recent streight video that I just uploaded if I click on this button to go to the Reddit post finder you can see that the way it works is that it has two different tools one is search for redit posts another is reflect if you click on this button you getting into the tool building experience it's basically like a no code interface for you to Define customer tools and functions for agents and for this one we're just give a name called search for RIT posts and description F latest RIT posts where we can define a few different inputs like the Search keywords the client ID client secret for Reddit as well as user agent do you want to search for pass hour or day and also number of submissions that we wanted to return and in this tool we basically just have one python C- step that copy paste function that we Implement before for returning the RIT posts once I create this tool and gave it to this agent it will be able to use it I did create a tool here called reflect so this is one way I found quite useful that Force agent to do some reasoning and Logics this basic one tool that has two inputs uh one is called reflection one is called Next Step I would force this agent to use this reflect tool to reflect on what they have done so far how far they away from achieved goal and what action they should take next and then have another input called what is best next action to take based on reflection so what this would do is basically after each time agent try to search for certain red posts it will cause this reflectable to have a reflection about what are the posts that actually relevant to the topic that we want to promote then what will be the next action we should do and here it reflect and decide okay the next step I should do is try to search post related to reducing large language model costs then it will search for new keywords get new results and reflect again so by having a reflect function I found it's quite a f way to force agent have some logic and reasoning uh to get better results so those are the five most relevant red post that it found for the past hour then it will return this exact result back to the reply guide then it will dedicate a task to the Reddit comand writer by passing on submission ID Reddit post title as well as a reasoning why he think it is relevant and what's the best way to plug your content it will compost a comment based on all those information and in the end this reply guy it will try to post this comment on this specific red post and I can set this action as require human approval or not and in this case I actually ask it to require approval so it want send out before I click on approve in this case I'm pretty happy with content so I can just click on approve so you can see here it successfully left the command if I click on this link so this agent just post a red command for me to this specific post and after that it will record this weight function to try to wait for 10 minutes before it do the next thing but if I switch back to another session you can see after it waited for a few minutes then it will call the red command writer again for the next posts until it post all the message out again all those agent to are very similar to to what we have built before so for post comment on R post is basically just a python code step with that specific API and the coin we want to call and the weight function is also just running this python code to wait for certain minutes and if we dive into RIT comand writer agent this agent that has a tool to fetch RIT post content but if I click on edit this is where I will also give some additional system Pro where I would say you are a experienced RIT user who always seek to post interesting comments and provide value to audience your goal is to write a comment to WR a post that will provide value to the author and audience recommend the SC we want to insert in the content in extremely Natural Way share your personal experience you're not selling this product you just want to share this product because you think it is good fit for them message should be short and just go straight to the content I also provide example of what a good soft plug might look like inside the RIT post finder I also have a special system prompt says the goal is to find the top five most relevant Reddit posts there are specific rules you need to follow you are determined to find super relevant Reddit posts you will keep changing the search keyword until you find five really random RIT posts and and always use this reflect function to actually reflect and figure out what's the next step you should do here I also put a rules that do not use this search RIT post for more than five times to prevent it getting to infinite loop I also have a kind of special prompt here to almost achieve the exped output function from crew AI to tell agent what should be your message format if we go to RIT reply dude we'll add this two sub agents to the team section so this basically means what are the team members that this reply guide can interact with and we can define a specific message template about what's the initial message that the reply D should send to those sub agents so the message that it should send to the Reddit postf finder is something like this so the message always like s we want to promote and plug so that it can insert this variable which is s to promote and then the go and I will put a description to the things to promote and in terms of the other one r a command writer I will give a message templates please fch the content of this Reddit post and drop that comment to the post based on things I want to promote where I pass all the detail information about what's the Reddit post are what's the submission ID what are the potential relevance what are things that we want to promote and plug I can also Define some instruction as well as process workflow so this is kind of one feature that random a I provide to give you some control to force certain Behavior so in here I can create this flowchart if someone said let's make some noise on Reddit it will use the Reddit post finder to find the five most relevant Reddit posts then for each post it should drive message then use post command Reddit function to post command and use weight function to wait for 10 minutes and repeat this process for all five posts if I click on this five links in the final message and you can see the agent has been combining on different redit posts that relate to AI agents or people learning python to provide value to them and also plug my AI YouTube channel some of them actually get some interaction and the best thing is once we set up this this agent we can actually run on schedule so I can just set up as this agent to be run hourly so every hour the agent can try to find is there any relevant redit post from the P hour so this is a quick example of how can you build an AI agent in just 10 minutes and I'm sure there are so many different use cases that you will know better and can start automating I hope you enjoy this content I will continue posting interesting AI projects that have been experimenting so please like And subscribe if you want to keep update thank you and I see you next time