Agentic AI & Automation – Lessons from Scaling Real AI Systems

Then there is Relevance AI. So Relevance AI is like a new software and basically it's an engine. So you prompt an engine, let's say, hey, I need a lead generation bot which will research on the leads on LinkedIn, on Facebook, and then give me the qualified leads and engage with those leads as well. Make sure that those leads, those leads are catered. And about 50% of those leads would book a call. Okay, so you give that prompting to Relevance AI, it will create the complete landscape for you, for the agents. Okay, how cool it is. Like, uh, it, it will pick up the apps, uh, himself and then come up with that particular workflow as well. Then there is Zapier. Zapier, like everyone knows about it, it's, it's very old as well. It's like have the options for a lot of apps, like more than 10,000 applications, which you can just plug and play into it as well. Now there's Crew AI. Crew AI basically is sort of like creating enterprise-level AI agents with Crew AI. And what you can do is you can connect one agent to another agent to make a complete crew and then run your business. So basically, let's say if you're a web development agency, okay, you have different processes. So for different processes, you create one specialized agent for that particular process and then combine it into one workflow to get an output. So which should be a completely designed website which you can build through crew AI. One for one agent for the design, one agent for, uh, let's say ideation, second agent for deployment, third agent for research, fifth agent for the content creation, and sixth agent to purchase a domain on your behalf. Something like that. You can create it with, uh, Crew AI. Now, uh, for some of the reasons, if you're not looking to use no-code tools, you can move towards the custom toolings as well. There is LangChain, there is Python-based LLM agents, there's RAG applications that you can create, okay? There's several tools you can create the RAG applications from. There is AutoGen, which is new from Microsoft. Then there is LlamaIndex, HiStack, okay? So these are all custom tools which you can use as well. Another thing is why, why and when should we use no-code tool and custom So for example, if you get a client's query, then you should understand like what is the limitation of a, of a no-code tool, just to make sure you move towards a customized tool because the customized tool would be more costly. You have to hire more people. And for the no-code tools, like you can set it up for like just take 1 day or 2 days and set up the entire workflow. Okay. So there are some comparisons. So no-code tools are ideal for startup agencies, MVPs, minimum viable products, lean teams. Custom tools is enterprise regulated industries, complex systems. Let's say for accounting firms, for big law firms, for big hospitals. If you would like to deploy an agent over there, obviously they are very reluctant about the security. So you can do it with the, with, with a custom tool. As far as the speed is concerned, no-code tool is very fast to build and iterate. Okay, if you don't like anything, if you would like to scale it up, you can easily do it with the no-code tool. You can just go into it and just remove a module, you add new module as well. Nowadays, on everyday basis, there are not new modules coming in. Now they have launched MCPs as well, they have launched new agents. Uh, no-code tools are basically making it easier for people to create agents. You just have to prompt in a natural language, and then, uh, all the workflow would be completed. You just have to connect the accounts. Uh, slower due to dev cycles and testing. Obviously it's custom, so it could take time, uh, to build for custom tools. Control limited logic, customization limited, debugging. Although no-code tools have the scenarios for error handling as well, but It's limited. Okay, so you shift towards the, um, the custom tools for that particular reason. If there's no requirements for, uh, uh, like going towards a custom tool, you can easily build it with like no-code tools. I have companies who are paying like about $20,000 for a system and I'm building it with no-code tool. Uh, yeah. And I'm basically, I'll show you how I present it to the customer because customer doesn't want to know, is it done with no-code tool? Is it done with custom tooling or Python whatsoever? That's it. Okay. They want the work to be done. They want the problem to be solved. If it could be solved with no-code tool, that's fine. Um, and obviously it's like $20 membership. Per month, which the client have to pay. And for scalability, you cannot scale the no-code tool to a certain extent. Okay, so it's good if you can move towards the custom tools if there is a requirement for that. And at least for security and privacy, custom tools are like good with security. Now, Anaden is providing self-hosting as well. So you can just self-host the entire workflow on your own server. So you can do it with that particular tool as well. Now normally the thing is, sometimes the customers comes in and they say that, hey, I have this problem and I need an AI agent to be built around this problem. Okay. Now, Obviously, the first thing is, is that problem worth pursuing? Okay, if it's not saving money for you, if it's not increasing your productivity, if it's not cutting up your maybe employees or employees' work, then why is the reason you are looking to get this automation done? It's still— if someone is reluctant to do that automation, I can do it for them. Okay, like there is No hard feelings for that thing, but that customer would do just one automation or one agent and he would be vanished because he is not receiving any like productivity in their business with that particular automation. So for that reason, I created like a calculator. I'll show it to you. So it's a calculator which I usually present to the client. Okay, if he is looking for, let's say, a problem to be solved. Let's say if a client is— which kind of problem do you guys think we should solve? Like, any, any ideas? Finance. Finance. Okay, uh, like reviewing invoices and, uh So let's say this is, uh, a kind of workflow that someone's looking to create, or any AI agent. Let's say, what's the time efficiency? So you ask the client, um, what should be the time efficiency in minutes? Let's say let's do it 3,000. Execution frequency per week, like how much time your employees are, um, spending on this particular task. So maybe just do it 4,000. Labor cost, let's say you're paying one employee, uh, €30 per hour, and there are 2 employees, so that's €60. So maybe let's do €60. And for the priority, like, is it like on a highest priority for the, for the company? You asked to the, uh, to the person, um, or maybe you decided yourself. Let's say it's a standard. And low risk, like, is it, uh, is there a probability for high errors. You can decide, decide for yourself. Let's say low risk, fully compatible with AI, partially compatible, and non-compatible. Okay, these are the filters that I created for myself. Okay, you guys can create your own filters as well. Let's say it's partially compatible. Okay, now it says that the time recovery is about 20,000, and this is the cost impact. That maybe it's like weekly savings. Then the AI score— this is an AI score that I give to every automation. And then automation is recommended, like we should pursue this automation, okay? It's good for the business. If it's not recommended, it will be red and it will tell that it's not recommended. You can save this into a database And you can export the report and you can give it to the customer. Okay, you can email it to the customer, or maybe you can show it over the call. Simple, real quick. Okay, just to understand if it's worth pursuing. So this gives you a perspective. Let's say, uh, the, um, the automation is worth pursuing. Okay, it's providing you the company the productivity increase as well. It, uh, it's cutting the dollars as well. Giving a lot of labor work, like decreasing labor work as well. So you basically ask these questions on a call as well, just to understand the process. So normally what I do is I speak with a client and I just ask him for the problem. Like, what is the problem? Okay, it's a problem. He's an accounting firm. There are so many manual reviews that they have to do for the documents. Okay, that's a problem. Let me talk with your finance person. Okay, let me talk with your employee. Okay, talk with an employee once you win the project. Okay, once you get the project, then talk with anyone, because else you will be wasting time and client won't be purchasing it for you. Just, just close the deal first and then get on a discovery call and just ask these particular questions as well. As well. And what does the success look like? Like, could be faster response, could be higher productivity, could be, um, like managing more manual work for the company, something like that. Okay, so these kind of questions that you can ask to a person over a call as well. And obviously it's like $20 membership per month which the client have to pay. And for scalability Uh, you cannot scale the no-code tool, uh, to a, to, to a certain extent. Okay. Uh, so it's good if you can move towards the custom tools if there is a requirement for that. And at least for security and privacy, custom tools are like good with security. Now, AnetN is providing self-hosting as well. So you can just self-host the entire workflow on your own server. So you can do it with that particular tool as well. Uh, now normally the thing is, uh, sometimes the customers comes in and they say that, hey, I have this problem and I need an AI agent to be built around this problem. Okay, now, um, obviously the first thing is, is that problem worth pursuing? Okay, if it's not saving money for you, if it's not increasing your productivity, if it's not cutting up your maybe employees or employees' work, then why is the reason you are looking to get this automation done? It's still— if someone is reluctant to do that automation, I can do it for them. Okay, like there is no hard feelings for that thing, but that customer would do just one automation or one agent and he would be vanished because he is not receiving any, like, productivity in their business with that particular automation. So for that reason, I created, like, a calculator. I'll show it to you. So it's a calculator which I usually present to the client, okay, if he is looking for, uh, let's say, a problem to be solved. Let's say if a client is— which kind of problem do you guys think we should solve? Like, any, any ideas?

Bill approval, or especially on finance, it's a lot of manual checks.

On finance? Finance. Finance. Okay. Like reviewing invoices and, uh, So let's say this is, uh, a kind of workflow that someone is looking to create for any AI agent. Let's say, what's the time efficiency? So you ask the client, um, what should be the time efficiency in minutes? Let's say, let's do it 3 2,000. Execution frequency per week, how much time your employees are, um, spending on this particular task. So maybe just do it 4,000—

not 4,000, too much—

400. Labor cost, let's say you're paying one employee, uh, €30 per hour. And there are 2 employees, so that's 60. So maybe let's do 60. And for the priority, like, is it like on a highest priority for the, for the company? You asked it to be, uh, to the person, or maybe you decided yourself. Let's say it's a standard and low risk. Like, is it, uh, is there a probability for high errors You can decide, decide it for yourself. Let's say low risk, fully compatible with AI, partially compatible, and non-compatible. Okay, these are the filters that I created for myself. Okay, you guys can create your own filters as well. Let's say it's partially compatible. Okay, now it says that the time recovery is about 20,000, and this is the cost impact that maybe it's like weekly savings. Then the AI score— this is AI score that I give to every automation. And then automation is recommended, like we should pursue this automation, okay? It's good for the business. If it's not recommended, it will be red and it will tell that it's not recommended. You can save this into a database and you can export the report and you can give it to the customer, okay? You can email it to the customer, or maybe you can show it over the call. Simple, real quick, okay? Just to understand if it's worth pursuing. So this gives you a perspective. Let's say, uh, the, um, the automation is worth pursuing, okay? It's providing you, the company, the productivity increase as well. It, uh, it's cutting the dollars as well. It's giving a lot of labor work, like decreasing labor work as well. So you basically ask these questions on a call as well, just to understand the process. So normally what I do is I speak with a client and I just ask him for the problem. Like, what is the problem? Okay, it's a problem. He's an accounting firm. There are so many manual reviews that they have to do for the, for the documents. Okay, that's a problem. Let me talk with your finance person. Okay, let me talk with your employee. Okay, talk with an employee once you win the project. Okay, once you get the project, then talk with anyone, because else you will be wasting time and client won't be purchasing it for you. Just, just close the deal first and then, uh, get a— just on— get on a discovery call and just ask these particular questions as well. And what does the success looks like? Like, could be faster response, could be higher productivity, could be, um, like managing more manual work for the company, something like that. Okay, so these kind of questions that you can ask to a person over a call as well. So one is working. Obviously, it's for some of, uh, some of the queries, a person have to go in just to have a human-in-the-loop perspective over there. Another thing is the cloud. It doesn't matter you are doing it with the no-code or you're doing it with the custom coding. What you have to show to the client is very important. So you have to create a centralized system, okay, where you can show the client how the AI agent is increasing the productivity of the business. Okay, so this is a law firm. Okay, so there, there might be multiple agents working, but when the client comes in every day, he opens up a dashboard and he sees that total active cases, lead conversion rate, average response, monthly revenue, client satisfaction, productivity gain. There's graphs and everything. They can generate a report as well. And this lead management, client analytics, case management— you can come up with any kind of analytics, but you have to show the front-facing CRM. Doesn't matter what, what is at the back end. The customer should know that this is working for me so that he could invest in more agents with your company as well. And I created a recommendation system as well. So let's say when the AI agent is working, there is one agent which is monitoring all the processes. And let's say if you're doing it for documents, uh, for the law firm to read the documents, create a RAG application for that. There is a probability we can do it to get the leads for the lawyers, to convert the leads for the lawyers, to do the outbound or the inbound, to do the content management of the marketing for the lawyer as well. So that agent would research on that particular system and do that stuff. And then on the dashboard, whenever the owner or maybe the team member opens up, he could see some recommendations from the agent. So automate personal injury intakes process. Okay. And it gives the rating as well. The expected ROI, it calculates the ROI, time to implement 2 to 3 weeks, start implementing. If they click on it, start implementing, they have to give me the money for that thing. So I have to go over the call and discuss with them, hey, okay, we have to create it. This is very complex and We have to create it separately, so it's another project for us. Okay, then there are other recommendations over there as well. They can view more details for those recommendations. Just so you guys know, this system, this calculator, and this system is built with Bold. Like, I created the Bold, I don't custom code it. would like spend 1 hour for that reason, but the clients love it. They could like have their own rules, they could have their employees, uh, sign up on this particular portal as well. They could see real progress over there. Okay, so and purchase it, and then they flip the land. Okay, let's say they purchase it for $80,000, they flip it for $120,000, something like that. Okay, so they make about $40,000 on that particular deal. So what, what their current flow is, they get the responses of the clients on an SMS campaign And on the SMS campaign, they receive the responses of the sellers and they say, I would like to purchase this, like buy this land. Okay, now there are a team of virtual assistants who take that information and then structurize that information. So they have the full name of the client, the APN number of the property, the land details, maybe the location, something like that, okay? And they inject it into a Slack channel, and on the Slack channel they have about 3 to 5 acquisition managers. So what, what those acquisition managers do is they recheck the information, they check it on other websites, let's say on Zillow, on LandID, on Realtor, on ArcGIS. There are so So many websites for real estate. So they check it over there, the information, see if there's a— if it's like really a vacant land or there's a commercial property over there, if the land is landlocked, okay, if there's a wetland under it. So there are so many things they review from it. Then they see the prices for the properties on there at the— at that neighborhood, okay. It's what— so they have to decide like how much we should offer the client so that we can make the money. So obviously it should be under the market value, and then they offer the client and they purchase the, the vacant land and they just flip that. Okay, so this is the complete process. So basically once they get the lead on the Slack, the acquisition manager would take time, do all the research, evaluation, and then send the offer to the person. That this is my offer. Okay, so what that client is looking to replace is that acquisition manager. Okay, all the 5 people, they want the AI agent to be created in a way so that whenever the lead is on Slack, in a minute there will be an offer with, with the structured method. So basically, when you create a— when you come up with this problem or this problem statement, first thing what you can do for the client is create this type of graph. We call this workflow graph. So obviously the graph could be very big, it could be very complex, but the thing is you have to make it very easy for the client to understand. There's a Slack, parse the seller input, research comps and build valuation, generate and send Offer, confirm parcel details, track and follow up. Okay, that should be another automation which could come in future, or maybe for the VAs as well, uh, if you would like to replace those VAs. So the agent would have structured the information for that purpose as well. Uh, so we created this for the client and he, uh, um, like put it in their system and they replaced— like, they have 5 employees, they replaced with the AI. So one is working, obviously, if for some of, uh, some of the queries a person have to go in, uh, just to have a human-in-the-loop perspective over there. Another thing is the cloud. It doesn't matter you are doing it with the no-code or you're doing it with the custom coding. What you have to show to the client is very important. So You have to create a centralized system, okay, where you can show the client how the AI agent is increasing the productivity of the business, okay? So this is a law firm, okay? So there, there might be multiple agents working, but when the client comes in Every day he opens up a dashboard and he sees that total active cases, lead conversion rate, average response, monthly revenue, client satisfaction, productivity gain. There's graphs and everything. They can generate a report as well. And this lead management, client analytics, case management— you can come up with any kind of analytics, but you have to show the front-facing CRM. Doesn't matter what, what is at the back end. The customer should know that this is working for me so that he could invest in more agents with your company as well. And I created a recommendation system as well. So let's say when the AI agent is working, there is one agent which is monitoring all the processes. And let's say if we are doing it for documents, uh, for the law firm to read the documents, create a RAG application for that, uh, there is a probability we can do it to get the leads for the lawyers, to convert the leads for the lawyers, uh, to do the outbound or the inbound, to do the content management of the marketing for the lawyer as well. So that agent would research on that particular system and do that stuff. And then on the dashboard, whenever the owner or maybe the, the team member opens up, you could see some recommendations from the agent. So automate personal injury intakes process, okay? And it gives the rating as well, the expected ROI, it calculates the ROI, time to implement 2 to 3 weeks, start implementing. If they click on it, start implementing, they have to give me the money for that thing. So I have to go over the call and discuss with them, hey, uh, okay, we have to create it. This is very complex and, uh, we have to create it separately. So it's another project for us. Okay, then there are other recommendations over there as well. They can view more details for those recommendations. Just so you guys know, this system, this calculator And this system is built with Bold. Like, I created the Bold. I don't custom code it, uh, with like spend 1 hour for that reason. But the clients love it. They could like have their own rules. They could have their employees, uh, sign up on this particular portal as well. They could see real progress over there. Okay, so