How to Build & Sell AI Agents in 2026: Ultimate Beginner’s Guide

0:003 years ago, I taught myself how to

0:01build AI agents without any prior

0:03experience in AI, and since then, I've

0:06built multiple AI businesses that have

0:07generated over $10 million in revenue,

0:09grown from zero to 700,000 subscribers

0:11here on YouTube, and I've built AI

0:12agents for NBA teams, for Fortune 500

0:15companies, and even publicly traded ones

0:17as well. So, it's safe to say that

0:18learning how to build AI agents has

0:20completely changed my life. So, in this

0:22full course, I'm going to teach you

0:24everything that I've learned over the

0:25past 3 years about building and making

0:27money with AI agents, even if you don't

0:29know how to code. My goal here is pretty

0:31simple. I want to give you the skills to

0:32build the life that you want before

0:34these AI agents start taking jobs away

0:35from us. Now, as you can probably tell

0:36by the length of this video, I'm not

0:38going to be holding anything back here,

0:39so the whole thing is split into three

0:41different chapters. Firstly, we're going

0:43to be building your foundational

0:44understanding of AI agents, what they

0:46are, how they work under the hood, and

0:48the key concepts that you need to know

0:49before starting to build them. And no,

0:51there's no technical background required

0:52for any of this. Secondly, we're going

0:54to dive into four end-to-end AI agent

0:56tutorials where I take you over my

0:58shoulder step-by-step as we build some

1:00of the most in-demand AI agent types on

1:02the market right now. And finally, I'm

1:04going to be giving you my proven

1:05blueprint for turning these AI agent

1:07building skills into real income while

1:09this technology explodes. So, let's get

1:11into it. But, if you are new to the

1:12channel and don't know who I am, my name

1:13is Liam Otley, and about 3 years ago, I

1:15knew absolutely nothing about generative

1:17AI, and now I run Morning Side AI where

1:19we build AI systems for publicly traded

1:20companies and even NBA teams. My last AI

1:23agent course like this has over 2

1:24million views, and through those videos

1:26and my community and my accelerator,

1:28I've helped tens of thousands of people

1:29to learn AI skills and build a real

1:31business around it. Right. And so, if

1:32you look at how long this video is,

1:34you'll realize that there is a lot to

1:35cover, but I do not want you to give up

1:37halfway through. So, before we jump into

1:38the technical stuff, let's first get

1:40clear on why this actually matters, why

1:42AI agents are genuinely one of the most

1:44valuable skills that anyone can learn

1:46right now. So, stick with me on this

1:47part. But, here's the hard truth about

1:48AI and jobs. McKinsey released a report

1:51in November of 2025 with a pretty

1:52shocking finding. 57% of all the work

1:55that Americans do today can be automated

1:57with the technology that already exists.

2:00Not technology coming in 5 years,

2:01technology that we have right now. And

2:03the World Economic Forum is projecting

2:0592 million jobs displaced just by 2030

2:08alone. Now, I know that sounds

2:09terrifying, and honestly, it should be,

2:11and I don't like using this kind of

2:12stuff on my channel if I'm honest, but

2:13if you're planning to do nothing about

2:15it, then there's an issue. But,

2:16interestingly, here's what those same

2:17reports found when they looked at the

2:19other side of the data. They found that

2:21workers who have AI skills are earning

2:2356% more than people who don't, and that

2:25gap in pay has doubled in just 1 year.

2:2766% of employers are actively trying to

2:30hire people who understand AI. So, on

2:32one hand, we've got this massive

2:33automation coming and job loss, while on

2:35the other hand, companies are desperate

2:37for people who actually understand this

2:39technology and can help them use it. So,

2:41the difference between getting replaced

2:42and getting ahead comes down to one

2:44thing, actually just learning the stuff.

2:46And here's the gap that creates such a

2:47huge opportunity right now. Only 12% of

2:50workers have taken any AI training at

2:52all. That's not just a little gap,

2:53that's a canyon. And if you don't

2:55believe me when I say that a little bit

2:56of self-study on the stuff goes a long

2:58way, here is Naval Ravikant, one of the

3:00world's most respected investors and

3:01technologists, on the All-In podcast.

3:03>> Again, I would say the easiest way to

3:05see that AI is not taking jobs or

3:07creating opportunities is [music] go

3:09brush up on your AI, learn a little bit,

3:11watch a few videos, use the AI, tinker

3:14with it, and then go reapply for that

3:15job that rejected you, and watch how

3:17they pull you in. This video is exactly

3:19what Naval is talking about. So, whether

3:21you're an entrepreneur wanting to learn

3:22valuable skills to launch an AI business

3:24like I have, or you're a business owner

3:25wanting to understand what agents can

3:27actually do for your company, or you're

3:28just an employee who wants to make sure

3:30you're the last person the boss thinks

3:32about letting go because you're the only

3:33one who actually gets this stuff, well,

3:34I made this video for you. So, here's

3:36what I want you to do. I need you to

3:37close out of all your other tabs, grab a

3:39notebook, get your coffee or tea or

3:40whatever kids you locked in, and make a

3:42commitment to yourself right now to

3:44actually [music] finish this training.

3:45Right. So, if you've done all that,

3:47let's get stuck into it.

3:54All right. So, step one in learning to

3:56build AI agents is actually

3:57understanding what an AI agent is in the

3:59first place. Because the term gets

4:01thrown around everywhere these days, AI

4:02agents this, AI agents that, and every

4:04tech company is talking about them, but

4:06what actually is an AI agent? When I

4:08first started learning about this stuff

4:09a few years back, I found it pretty

4:11confusing, too. So, let me give you the

4:12clearest definition that I found that

4:14really helped it kind of click for me.

4:16An AI agent is a digital worker that can

4:18understand instructions [music] and take

4:20actions to complete tasks. So, in a

4:22simple way, just like businesses have

4:24employees who handle different

4:25responsibilities, AI agent is kind of

4:27like having a digital employee. But, the

4:29interesting part is that you can build

4:30them yourself and make them do pretty

4:32much whatever you want. It's like being

4:33able to construct a human worker from

4:34scratch. And unlike human employees,

4:36these digital workers cost a fraction of

4:38what a person does. They can work 24

4:40hours a day, 7 days a week. They never

4:42get tired. They never call in sick, and

4:44you can duplicate them instantly

4:45whenever you need more capacity. So, I'm

4:47sure you can see the appeal for

4:48businesses when they look at this versus

4:50humans. Now, here's why this matters

4:51right now specifically. Because all

4:53major tech leaders are aligned that 2026

4:55are the year that AI engines go

4:57mainstream. You've got Satya Nadella,

4:58the CEO of Microsoft, who said at the

5:00start of this year, "2026 will be a

5:02pivotal year for AI. We've moved past

5:04the initial phase of discovery and are

5:06entering a phase of widespread

5:08diffusion." Jensen Huang from Nvidia is

5:10saying the same thing. Sundar Pichai at

5:11Google said 2026 is when people will

5:13start using agentic experiences more

5:15broadly. The point is that this isn't

5:17just some far-off future technology.

5:19This is happening right now, and

5:21understanding how these things work puts

5:22you ahead of the vast majority of

5:24people.

5:27Now, to understand why AI agents are

5:28such a big deal, we need to look at

5:30where we are coming from, because you've

5:32probably encountered all sorts of

5:33chatbots before, right? You go to a

5:34website, there's a little chat widget

5:36that pops up in the bottom corner

5:37saying, "Hi, how can I help you today?"

5:39And so, you type something in, it gives

5:41you some kind of response. These kinds

5:42of basic chatbots have been around for

5:44years, but here's the thing, they are

5:45pretty limited, and they can really only

5:47do one thing, respond with some

5:49information. The way I like to think

5:50about it is this, a chatbot is like a

5:52waiter who can only recite the menu to

5:53you. You ask them what's on the menu,

5:55and they tell you, but they can't

5:56actually take your order. They can't

5:58bring you your food, they can't process

5:59your payment, either. They just give you

6:01information back. AI agents, on the

6:03other hand, are fundamentally different.

6:04But, let me give you a concrete example

6:06to show you what I mean. Say you go to a

6:07website and you want to book in an

6:08appointment with that business. If you

6:10ask a regular chatbot about booking an

6:12appointment, it might say something

6:13like, "Our business hours are 9:00 to

6:155:00, Monday to Friday. Please call us

6:17to book." And that's it. Just some

6:19information, and you still have to go

6:20and do all of the work yourself. But, if

6:22you ask an AI agent the same question,

6:24it could actually check the calendar in

6:25real time, find available slots, ask you

6:28which time works best, and then book the

6:29appointment for you, and send you a

6:31confirmation email, and then update the

6:32business's scheduling system, all

6:34automatically in a matter of seconds.

6:36So, that's the key difference here.

6:38Agents don't just respond with

6:39information, they actually take action

6:41and get things done. And this ability to

6:42take action is what makes agents so

6:44powerful and why everyone's so excited

6:46about them. They're not just fancy

6:47chatbots that sound a bit smarter,

6:48they're genuine digital workers that can

6:50do real tasks like searching through

6:52databases, updating spreadsheets,

6:54sending emails, booking appointments,

6:56generating documents, and honestly, so

6:57much more that we're going to be getting

6:59into in the build section of this video.

7:00Building and deploying an AI agent is

7:02actually a lot like hiring a new

7:03employee, because when you bring someone

7:05new into the business, you explain their

7:06role and responsibilities, you give them

7:08access to the systems that they're going

7:09to need, and then you trust them to

7:11handle tasks on their own. It's the

7:13exact same process with AI agents. You

7:14tell them what they're supposed to do,

7:16you give them access to the tools they

7:17need, and then they can work on their

7:19own independently. The difference is

7:20that they work around the clock, they

7:22can be copied instantly, and they cost a

7:23tiny fraction of what a human does. This

7:25is exactly why understanding how to

7:26build and deploy these digital workers

7:28is becoming such a crucial skill.

7:30Whether you're an entrepreneur looking

7:31to scale your business, or an employee

7:33wanting to become irreplaceable at your

7:34job, knowing how to create these things

7:36is genuinely one of the biggest

7:37advantages that you can have right now.

7:42All right. So, now that you understand

7:43what AI agents are and how they're

7:44different from basic chatbots, let's

7:46look under the hood and figure out how

7:47they actually work. Because just like

7:49humans need certain things to do their

7:50job, a brain to think, some memory to

7:52remember stuff, and tools to be able to

7:54work with, AI agents need specific

7:55components to function as well. There

7:57are five key parts that make up an AI

7:59agent. Let me walk you through on each

8:00one. First, every agent needs a brain.

8:02In the AI world, these are called large

8:04language models, or LLMs for short.

8:06You've probably heard of these already.

8:07We've got GPT-5 from OpenAI, Claude from

8:09Anthropic, and Gemini from Google. These

8:11are all LLMs, and you can think of the

8:13LLM as a super intelligent assistant who

8:15can understand your instructions in

8:17plain English and figure out how to get

8:18things done from there. It's the

8:20intelligence that powers everything

8:21else. And without the brain, all the

8:23other parts of this will be useless,

8:24like having a desk full of office

8:26supplies, but no one needed to actually

8:27use them. Secondly, the brain needs

8:29instructions on how to actually behave.

8:30This is what we call writing a prompt.

8:32So, writing a prompt is essentially how

8:34you program the behavior of your agent,

8:36but instead of writing complicated code,

8:37you're just writing clear instructions

8:39in normal [music] language to them. This

8:41is what makes building with AI so

8:42accessible to people who aren't actually

8:44programmers, because the way you tell an

8:45agent what to do isn't through code,

8:47it's through these well-crafted

8:48instructions that anyone can write.

8:50We'll get much deeper into this when we

8:51start building. Thirdly, agents need

8:52memory. Imagine trying to have a

8:54conversation with someone who completely

8:55forgets everything you said 30 seconds

8:57ago. It'd be impossible to get anything

8:58done. Memory allows your agent to

9:00remember what you just talked about,

9:01keep track of tasks that they're working

9:03on, and build on previous conversations

9:05as well. The good news here is that most

9:06agent platforms handle this memory

9:08automatically for you in the background,

9:10so you don't really need to worry about

9:11setting it up for yourself. Fourth, we

9:12have knowledge. Now, the AI models like

9:14GPT and Claude are trained on massive

9:16amounts of data from the internet, but

9:18that training data has a cutoff point,

9:19and there's a lot of stuff that they

9:20just don't know. More importantly, they

9:22don't know anything specific about the

9:23business that you're trying to help.

9:25They don't know about their products,

9:26they don't know about their services,

9:27they don't know about the policies or

9:28your pricing. So, just like you're

9:30training your employee with

9:31company-specific materials, you can give

9:33your AI agent additional knowledge

9:34through things like PDFs of documents,

9:36spreadsheets with product information,

9:37customer service transcripts, really any

9:39text-based information that you want it

9:41to have access to. Without this added

9:43knowledge, your agents will be limited

9:44to just general information and couldn't

9:46handle the specific tasks a business

9:48actually needs. And fifth, and this is

9:49the most exciting part, we have tools.

9:51Tools are what transform an AI agent

9:53from something that could just chat into

9:54something that can actually get things

9:55done in the real world. Think of tools

9:57like giving your digital employee access

9:59to different software and systems. Just

10:00like you might give a new hire access to

10:02your email, your calendar, and a CRM,

10:04you give an AI agent access to digital

10:06tools so they can take real actions.

10:08These tools let your agent do things

10:09like check real-time data, update

10:11databases, send messages and

10:12notifications, create documents, make

10:14phone calls, anything you can do on a

10:16computer, you can potentially give an

10:18agent the ability to do that as well.

10:19The really powerful part is when agents

10:21use multiple tools to solve complex

10:22problems, just like how we use multiple

10:25different websites and apps for when

10:26we're doing our own work. Now, here's

10:27the practical takeaway from all of this.

10:29While an agent has these five

10:30components, the brain, the prompt, the

10:32memory, knowledge, and the tools, you

10:34don't actually have to worry about all

10:35of these five things equally when you're

10:37building. The brain is handled by

10:39choosing which AI model to use, and

10:40honestly, any of the top models work

10:42great, and the memory is handled

10:43automatically by the platforms that you

10:45build on in most cases, which leaves you

10:47with what I call the three ingredients,

10:48the three things that you actually need

10:50to focus on as an AI agent builder. And

10:51that is prompting, how you instruct the

10:53agent to behave, kind of the glue that

10:55sticks it all together, knowledge, the

10:57external information that you give it

10:58access to, and three, the tools, the

11:01actions that you want it to be able to

11:02take. So, write that down, prompting,

11:04knowledge, and tools. That framework is

11:05going to guide everything we do in the

11:07build section of this video.

11:11All right. Now, we need to go even

11:12deeper on tools because they're honestly

11:13the most powerful part of AI agents.

11:15That's what separates a basic chatbot

11:17from a genuinely useful digital worker.

11:19But to really understand how tools work

11:21and to be able to build your own

11:22powerful agents later, we need to take a

11:23few steps back and cover some of the

11:25basics of how software and the internet

11:27actually work. Now, I know that might

11:29sound intimidating, but this is probably

11:30the most technical section of the entire

11:32video. But I promise you, once you

11:33understand this stuff, it's like having

11:35a superpower. It is the foundation of

11:36being able to build any kind of software

11:38and build apps for yourself and for

11:39other people. Everything else in this

11:41video is going to make so much more

11:42sense, so stick with me here. So,

11:43remember how we said that tools are what

11:44allow agents to take action, to actually

11:46do things rather than just chat. Well,

11:48the way agents actually do work online

11:50is pretty similar to how you and I do

11:52things online, too. But there's one key

11:53difference. Instead of clicking buttons

11:55and typing into forms and navigating

11:56around websites, agents use something

11:58called API. Now, before your eyes glaze

12:00over, let me show you what I mean.

12:01Because you already use APIs every

12:03single day, you just don't realize it.

12:05So, every time you do anything on the

12:06internet, you're actually making dozens

12:07of what we call requests to APIs and

12:10getting responses back. It's all

12:12happening behind the scenes for you,

12:13though. So, let me give you a real

12:14example. When you click on this video to

12:16watch it, here's what actually happened

12:17behind the scenes. Your browser sends a

12:19request to YouTube servers, essentially

12:21saying, "Hey, I want to watch this

12:22video." YouTube servers receive the

12:24request, found all of the data needed

12:26for this video, and then sent that back

12:28to your browser, to you, as a response.

12:30Then your browser took all of that

12:31information, unpacked that data, loaded

12:34everything up in front of you, and then

12:35started playing the video for you. And

12:37that whole request-response pattern

12:38happens with almost everything [music]

12:40you do online. If you open up Instagram,

12:42you're sending a request for your feed

12:43to be loaded, and the servers respond

12:45with all of those posts and stories.

12:46[music] If you're sending a tweet,

12:47you're sending a request that contains

12:49your tweet data, then Twitter servers

12:50save it and respond confirming that it

12:52worked. If you're checking your email,

12:54you're requesting your latest messages,

12:55and Gmail servers are responding with

12:57all of your emails in your browser as

12:58loading them. It's just like computers

13:00talking back and forth, requesting

13:01things and responding to those requests.

13:03We get nice-looking websites and apps

13:04that make it all feel simple, but under

13:06the hood, it's all just this

13:07request-response conversation that's

13:09happening constantly. [music]

13:10These requests and responses happen

13:11through what we call APIs, or

13:13application programming interfaces. And

13:15I know that sounds technical, but think

13:16about APIs as like waiters at a

13:18restaurant. You tell the waiter what you

13:20want, the waiter takes your order to the

13:21kitchen, the kitchen prepares it, and

13:23the waiter brings your food back to you.

13:25An API works in the same way. They take

13:27your request to the server, the server

13:29does whatever work is needed, and the

13:30API brings back that response to you.

13:32Now, there are two main types of

13:33requests you can make through APIs, and

13:35this is really worth noting down. You

13:37have get requests, which is when you're

13:38asking to get information from

13:40somewhere, like [music] checking the

13:41weather or looking up a stock price or

13:44loading this video. You're requesting to

13:46receive data. Post requests, on the

13:48other hand, is when you are sending

13:49information for the server to do

13:51something with, like posting a tweet or

13:53sending an email or uploading a photo.

13:55You are posting data for someone else to

13:57store and process for you. So, you've

13:59got get and you have post, [music]

14:00requesting data and sending data. Write

14:02those down because you're going to be

14:03seeing them a lot. All right, so here's

14:04where this becomes super relevant to AI

14:06agents. AI agents use the exact same

14:08APIs as their buttons to do things. When

14:11we talk about giving AI agents tools,

14:12[music] what we're really doing is

14:14giving it the ability to make these API

14:16calls themselves when they think it's

14:17necessary. Each tool an agent has access

14:19to is essentially an API that it can

14:21call to either get information or

14:23[music] send information somewhere else.

14:24Now, these kinds of tools come in two

14:26different flavors. Firstly, you've got

14:27pre-made integrations. These are tools

14:29that someone else has already built and

14:30packaged up nicely for you into an API.

14:32These are things like Google Calendar

14:34integrations or Gmail [music] or Slack.

14:36The hard work of building and figuring

14:37out the API is already done. You just

14:39kind of plug into it and your agent can

14:40use it. You can think of these as like

14:42buying a ready-made meal. Everything has

14:43already been prepped for you, you just

14:44kind of heat it up and eat it. On the

14:46other hand, you've got custom tools that

14:47you build yourself. This is when you

14:49need your agent to do something specific

14:50that it doesn't have a pre-made

14:51integration for. In this case, you need

14:53to build the tool yourself, as we're

14:54going to go into later, and set up the

14:56API connection to your agent manually.

14:58You can think of this as like cooking

14:59from scratch. It's more work, but you

15:00can make exactly what you want. Both

15:02approaches work great, but when we get

15:03to the build section, I'll show you how

15:04to use pre-made integrations and how to

15:06create custom tools from scratch, and

15:08it's much easier than it sounds,

15:09especially with the no-code platforms

15:11that we're going to be using. The

15:11important thing to understand right now

15:13is that any action an agent takes,

15:14whether it's sending an email, checking

15:16a calendar, updating a spreadsheet,

15:18whatever, it's doing it by calling an

15:19API, by making one of these requests and

15:21getting a response back. And once you

15:23understand that, you start to see the

15:24internet completely differently, where

15:26every action online is just requests and

15:28responses, which means that if you can

15:30identify the right API, then you can

15:32build a tool that allows your agent to

15:34automate almost anything.

15:38Okay, so now you understand that APIs

15:40are how the internet works. The next

15:41question is, how does an AI agent

15:43actually know how to use a tool? Because

15:45if you think about it, you can't just

15:46point an agent at an API and expect it

15:48to figure everything out. It needs to

15:49know what the tool does, how to use it

15:51properly, and what information to send

15:53to it in order to get it to work. This

15:55is actually a lot simpler than it

15:56sounds, and it's one of the things that

15:57makes AI agents so powerful. So, for a

15:59tool to work with an agent, it only

16:01really needs to know three things.

16:03Firstly, what the tool does, which is a

16:05clear description of the tool's purpose,

16:06like, "This tool sends an email." Or,

16:09"This tool checks calendar

16:10availability." Secondly, what input the

16:12tool needs, what information does it

16:14actually require to work? An email tool

16:16might require the recipient's address, a

16:18subject line, and the message body. A

16:20calendar tool might require a date range

16:21to check. And thirdly, what does the

16:23tool actually return? So, what does the

16:25tool give back to the agent after it

16:27runs? The email tool might return a

16:29confirmation that the message was

16:30actually [music] sent. The calendar tool

16:32might return a list of available time

16:34slots. And that's it, really. You've got

16:35the description of the tool, the inputs

16:37it expects, and the outputs. [music]

16:38Now, here's the cool part. The way you

16:39provide information to an agent isn't

16:41through complicated code, you're

16:42essentially just explaining how to use

16:44the tool in plain language, just like

16:46you'd explain it to your new employee.

16:47You write something like, "This tool

16:49sends an email. It needs three inputs,

16:50the recipient's email address as text,

16:52the subject line as text, and the

16:54message body as text as well, and it

16:56returns a confirmation that the email

16:57was sent successfully." The agent can

16:59read that description, and it genuinely

17:01understands it. It knows what this tool

17:03is for, what it needs to make it work,

17:05and what to expect back from it. There's

17:06a standard format for writing these

17:08descriptions known as a schema, [music]

17:09and I know schema sounds technical and

17:11scary, but it really is just a

17:12structured way of writing out those

17:14three things, what the tool does, what

17:15it needs, and what it returns. And the

17:17great news is that most of the no-code

17:18platforms we'll be using actually

17:19generate these schemas automatically for

17:21us, so you won't have to write them from

17:22scratch, but it helps to know that they

17:24exist [music]

17:25and what they're doing because if an

17:26agent is using a tool incorrectly, it's

17:28usually because the description isn't

17:30clear enough. Now, here's where the

17:31magic actually happens. Modern AI models

17:33like GPT-5 and Claude are able to just

17:35read these tool descriptions, they can

17:37understand them well enough to figure

17:38out both how to use the tool and when to

17:41use it, when it should actually be

17:42triggered in a conversation. So, let me

17:44show you what I mean with this simple

17:45example. Say you've given your AI agent

17:47a tool with the description that says,

17:48"This tool checks today's weather for

17:50any city." The tool needs one input, the

17:51name of the city as text, and it returns

17:54the current weather conditions. Now,

17:55someone chats to your agent says, "Hey,

17:56what's the weather like in Tokyo right

17:58now?" Here's exactly what happens inside

18:00the agent's brain. First, it looks at

18:01the message and understands that this

18:03person wants to know about weather. Then

18:04it looks through all the tools it has,

18:06and it might have 10 of them, and only

18:07one of them is the weather one. And it

18:08sees one tool description mentioning

18:10weather. It thinks, "Okay, this is

18:11probably the one I need." Then it looks

18:13further, and it checks what inputs that

18:14tool requires. It sees that it needs a

18:16city name. Then the agent looks back at

18:18the message, extracts Tokyo as the

18:20relevant city, and makes the API call to

18:22the weather API with Tokyo as the input.

18:25The API goes off, gets the weather data,

18:26and then sends back a response, probably

18:28a bunch of technical data about

18:30temperature and humidity conditions, all

18:32of that. And when the agent receives the

18:33response, the clever part is that it

18:35doesn't just dump all of that raw data

18:37on the user, it takes the information

18:38and writes a natural, conversational

18:40reply like, "It's currently 72° and

18:43sunny in Tokyo with a light breeze." So,

18:45what we see here is that the agent

18:46understood the intent, it found the

18:48right tool, figured out the inputs, made

18:50the call, and then translated that kind

18:52of technical database response into

18:54something actually useful for the user

18:55based on what they asked for, all

18:57automatically. And this is what makes AI

18:58agents so different from the software

19:00that we've had before. Traditional

19:02software needs you to click the exact

19:03buttons and fill in the exact forms,

19:05whereas agents can understand what

19:07you're trying to accomplish and figure

19:08out how to use their tools to get it

19:10done. When you really get this, you will

19:11never see technology the same way again.

19:13The combination of language

19:14understanding and tool use is genuinely

19:16a new kind of capability. And the better

19:18you get at setting up tools with clear

19:19descriptions, the more reliably your

19:21agents will be able to use them. In the

19:22builds later, you'll see exactly how to

19:24set up tools in practice, and you'll see

19:26the small changes to descriptions that

19:27can make a huge difference in how well

19:29the agent uses them. But for now, just

19:31understand the core concept. Clear

19:32descriptions are how agents know what

19:34their tools do and when to use them.

19:38All right, so now you understand how an

19:39agent uses a single tool. It reads the

19:41descriptions, figures out when to use

19:43it, grabs the right inputs from the

19:44conversation, makes the API call, and

19:46turns that response into a helpful

19:48answer. But obviously, having an AI

19:49agent that can only check the weather or

19:51do one thing isn't particularly

19:52impressive. The real power of AI agents

19:54comes when you give them multiple tools

19:56and the ability to use them together to

19:58accomplish their complex goals. Do you

20:00remember our definition from earlier? An

20:01AI agent is a digital worker that can

20:03understand instructions [music]

20:04and take actions to complete tasks. When

20:06you give an agent a task, it will try

20:08its best to complete it. But if it

20:09doesn't have the right tools available,

20:11it can't do much. It's like asking an

20:12employee to send a report to a client

20:14but not giving them access to their

20:15[music] email. Doesn't matter how smart

20:17they are, they can't do the job without

20:18the tools. The more tools you give to an

20:20agent, the more flexibility it has to

20:22solve problems. And this is where things

20:23start to get really interesting. So let

20:25me give you a real example from my own

20:26business to show you what I mean. Say I

20:28give an agent this task, which is find

20:29AI startups that have raised money

20:31recently, put them on a spreadsheet, add

20:32a summary of [music] each business, and

20:34email me the link when you're done. Now

20:35that's not a simple one-step task.

20:37There's actually a lot going on there.

20:38And when you give an AI agent a complex

20:40task like [music] this, and you provided

20:42it with multiple tools, it will actually

20:43break down the problem and plan out how

20:45to solve it. Kind of like how we you or

20:46I would approach it. The agent might

20:48think, "Okay, I first need to find AI

20:50startups who have raised money. So I'll

20:51use the web search tool, and I'll use

20:53that to run some searches about AI

20:54startups and [music] their recent

20:56funding rounds. Then I need to put them

20:57into a spreadsheet. I have a Google

20:59Sheets tool that can create new

21:00spreadsheets and add new rows. I'm going

21:02to use that. For each company I find,

21:03I'll need to add them as a new row in

21:05the spreadsheet with the information.

21:06Then I need to add summaries of each

21:08business. I can write those summaries

21:09myself based off what I found in my

21:11research. Finally, I need to email the

21:13link. I have a Gmail tool, so I'll use

21:15that to send the spreadsheet link." And

21:16then it just does all of that step by

21:18step using different tools in sequence,

21:20working towards the goal that you've

21:21gave. This is when you really start to

21:23see why they call these things digital

21:24workers. They can plan out multiple

21:26steps and execute them in order, use

21:28different tools as needed, basically

21:29approaching problems the way a human

21:31would. Now the really cutting edge stuff

21:32happening right now is with what are

21:33called reasoning models.

21:34>> [music]

21:35>> These are newer AI models that are

21:36specifically designed to be better at

21:38this kind of multi-step planning and

21:39problem solving. What makes them special

21:41is that they don't just plan once and

21:43execute. [music] They can actually

21:44reflect on the results as they go and

21:45adjust their approach if something isn't

21:47working. So let's say an agent runs that

21:48first web search for AI startups raising

21:50money, and it doesn't get very good

21:52results. A basic agent might just push

21:53forward with bad data, but a reasoning

21:55model might think, "Hmm, those results

21:57aren't that good. Let me try a different

21:58[music] search query. Or maybe I should

22:00search for recent funding announcements

22:02specifically." It's that ability to

22:03plan, execute, reflect, and then replan

22:06that makes these agents capable of

22:08complex tasks without constant

22:09hand-holding. Now I should be honest

22:11with you here, this technology isn't

22:12perfect yet. These multi-step tasks can

22:14be incredibly unreliable, and agents

22:16definitely still make mistakes on

22:18complex workflows. For anything

22:19important, you still typically want

22:20human oversight so that you can catch

22:22the errors before they actually cause

22:23problems. But things are moving

22:24incredibly fast, and the agents that we

22:26can build today are dramatically more

22:27capable than what we had even a year

22:29ago. And if we look a bit further ahead,

22:31we're already seeing the next evolution,

22:33which is multiple agents working

22:34together. Instead of one agent trying to

22:36do everything, you can have multiple

22:37specialized agents that each focus on

22:39one thing, kind of like having different

22:41employees with different job roles. So

22:43you might have a research agent that's

22:44really good at finding information, a

22:46writing agent that's optimized for

22:47creating summaries and content, and

22:49email agent that handles all the

22:51communication stuff. And the cool part

22:52is that one agent can actually use the

22:54other agents as a tool. So your main

22:56agent receives the task and delegates

22:57specific parts to the specialist agents

22:59like, "Hey, research agent, go find me

23:01these companies. Hey, writing agent,

23:03take this information and create

23:04summaries." Each agent focuses on what

23:06it does best, and together they

23:08accomplish more than any single agent

23:10could. This is exactly what companies

23:11like Microsoft, Salesforce, and Google

23:13are bidding heavily on right now. Entire

23:15systems [music] of AI agents working

23:17together to handle complex business

23:19processes. Now we're not going to build

23:20multi-agent systems in this video.

23:22That's a bit more advanced stuff that

23:23you can do explore once you've got these

23:24fundamentals down, but I want you to

23:26know that it exists and where all of

23:27this is heading. For now, let's keep

23:29building your foundation. Next up, we

23:31need to talk about the different ways

23:32people actually interact and use AI

23:34agents.

23:38All right, so we've covered what AI

23:39agents are, how they work, and how they

23:40use tools to get things done. Now we

23:42need to look at the different ways that

23:43AI agents are actually used in the real

23:45world, because not all agents work in

23:46the same way. There are basically two

23:48main categories that you need to

23:49understand. First, you've got

23:50conversational agents, and these are

23:52agents that humans interact with

23:53directly. So someone is actually there

23:55chatting with the agent, giving it

23:56instructions, asking questions, and

23:58getting responses back. You've probably

23:59seen these in a bunch of places already,

24:01like chat widgets on websites where you

24:03type messages back and forth, WhatsApp

24:04bots that you can text, Instagram DMs

24:06that respond automatically, or even

24:08voice agents that you can actually call

24:10and have a conversation on the phone. In

24:11all of these cases, there's what we call

24:13a human in the loop that's actually

24:15sending messages to the agent, and the

24:16agent is responding. It's a

24:17back-and-forth conversation, just like

24:19you'd have with another person, except

24:20the other side is an AI. Open AI's

24:22custom GPTs are great example of this.

24:24You can create an agent and then chat to

24:26it directly through the ChatGPT

24:27interface. Secondly, you've got

24:29automated agents, and this is where

24:30things get a little bit more

24:31interesting, because with automated

24:33agents, there's actually no human

24:34sitting there sending a messages

24:36directly. Instead, the agent is part of

24:38a larger system or a workflow, and it

24:40gets triggered automatically when

24:42certain things happen. If you think

24:43about it this way, conversational agents

24:45wait for a human to talk to them.

24:46Automated agents wait for an event or a

24:48condition to trigger them, and then they

24:50spring [music] into action on their own

24:51accord. For example, you could have an

24:53agent that triggers whenever a new lead

24:55fills out a form on your website. The

24:57moment that form submission comes

24:58through, the agent wakes up, looks at

25:00the lead's information, does some

25:01research on them, decides if they're

25:03qualified or not, and then updates the

25:04CRM with its findings. All of this

25:06without any human pressing a button or

25:08sending [music] a message directly to

25:09it. Or you could have an agent that runs

25:11every morning at 9:00 a.m., it checks

25:12your calendar for the day, looks up

25:14information about the people you're

25:15meeting with, and then sends you a

25:16briefing email before your first call.

25:18Again, completely automatic. You don't

25:20have to ask it to do anything. It just

25:22runs on a schedule or a different

25:24trigger. This is a really important

25:25distinction because it opens up a whole

25:27world of different use cases. Automated

25:29agents can handle processes [music] that

25:31happen constantly in a business. These

25:32are things that would be tedious or

25:34impossible for humans to monitor and

25:35respond to 24/7. In the building section

25:37of this video, you're going to get

25:39hands-on experience of building both

25:40types. We'll be building four different

25:41types of agents, three conversational

25:43and one automated, so you'll see exactly

25:45how each works in practice.

25:46Understanding this distinction is

25:47important because it shapes how you

25:49think about what's possible. A lot of

25:50people only think about chatbots and

25:51agents that you talk to, but some of the

25:53most valuable applications are agents

25:55that work completely in the background,

25:56handling tasks that you'd never want to

25:58do manually. All right, so we are almost

25:59done with the foundations. Let's just

26:01look at some real-world examples of how

26:03businesses are actually using these

26:04agents right now, and then we'll do a

26:06quick knowledge check to make sure

26:07you've got everything here before we

26:08jump into the building.

26:12Okay, so before we wrap up the

26:13foundations, let's quickly look at some

26:15real examples of how AI agents are being

26:17used by businesses right now. Seeing

26:18these concrete use cases helps all of

26:20this theory stuff to click into place.

26:21And these are all the kinds of things

26:22that we're going to be building in the

26:23build section. First use case is

26:25copilots, and these are AI agents that

26:27help someone in a specific role to do

26:29their job more effectively. For example,

26:31a customer support sales rep could have

26:32a copilot that would be able to

26:33instantly search the knowledge base of a

26:35company for answers mid-call, so it

26:37helps them to do their job. Same thing

26:38with the sales rep, they can have a

26:40copilot that gives them access to

26:41everything they need to know about the

26:42lead, and maybe can do research for

26:44them, and so on. In fact, in our final

26:45build, we'll be creating a sales copilot

26:47that was able to research leads and

26:48generate briefings so that the sales

26:50reps can walk into every call feeling

26:52fully prepared. Second major use case is

26:54lead generation and appointment setting

26:55agents. This is probably the most common

26:57business use case right now because it

26:58directly impacts revenue. When someone

27:00visits a website at 11:00 p.m. on a

27:02Saturday with questions, instead of

27:03having to wait until Monday and probably

27:05forgetting about it, an AI agent can

27:07immediately get back to them and answer

27:08the questions, capture their

27:10information, and even book appointments

27:11by checking calendar availability in

27:13real time. These days, speed to lead, or

27:15the speed at which you can reply to

27:17leads, is everything in sales, and the

27:19business who responds first wins the

27:20deal most of the time. These agents are

27:22starting to show up everywhere, website

27:24chat widgets, WhatsApp, Instagram DMs,

27:26even over the phone. And the point is

27:27that 24 hours a day, they are always

27:29ready to engage potential customers. In

27:30build two, we're creating a website chat

27:32agent with some very powerful tools that

27:34handles lead generation for a solar

27:35company. Third use case is research and

27:37qualification agents. These are agents

27:39that gather more information about leads

27:41so a business can prioritize the best

27:42opportunities. Sometimes a lead comes in

27:44with just basic details, a name, an

27:46email, a phone number, not too much to

27:47work with. A research agent can

27:49automatically dig deeper, finding

27:51LinkedIn profiles, company information,

27:52collect together context that helps the

27:54sales team know who they're going to be

27:55talking with. Or you can go even more

27:57direct and have an agent reach out by

27:58SMS

28:01or over the phone and ask them some

28:03qualifying questions. This is getting

28:05the information straight from the

28:06source. In build three, we're creating a

28:07voice agent that automatically calls new

28:09leads to qualify them, gathering that

28:10extra information to decide if they're

28:12good fit or not before human even needs

28:14to get involved. And fourth is voice

28:15agents more broadly. This is one of the

28:17fastest growing areas in the AI space

28:19right now, and the ROI with businesses

28:21has been proven. One of the startups

28:22Retail AI is processing over 40 million

28:24AI phone calls per month. Businesses are

28:26using these voice agents to answer

28:27inbound calls, qualify leads, book

28:29appointments, and even make outbound

28:31follow-up calls as well. The economics

28:32of this stuff is pretty compelling, with

28:34Client AI reporting that the AI agent

28:35handles work that required previously

28:37700 customer service reps with fast

28:39response times and higher customer

28:41satisfaction scores as well. We've got

28:43voice AI woven all through the builds

28:44that we're going to be doing, so you're

28:46going to get plenty of hands-on

28:47experience with it. So these aren't

28:48hypothetical future applications.

28:50Businesses are deploying agents like

28:51this right now and seeing incredible

28:53results. And you're about to learn how

28:54to build them. So let's do a quick

28:55knowledge check to make sure that

28:56everything in this foundation

29:00All right, we've covered a lot of ground

29:01in this foundations chapter, but before

29:02we jump into building these agents, I

29:04want to make sure that everything is

29:05landed for you. Because here's the

29:06thing, if you don't have these concepts

29:08down solid, the build section is going

29:09to feel very confusing. You'll be

29:11following along but just clicking

29:12buttons without really understanding

29:13why, and that's not going to help you

29:15when you're trying to build your own AI

29:16agents later. So let's do a quick

29:17knowledge check. I'm going to ask you a

29:18few questions, and you're going to

29:19genuinely try to answer them before we

29:21move on. Question one, what's the

29:22difference between a chatbot and an AI

29:24agent? Question two, what are the three

29:26ingredients you control when building an

29:28AI agent? Question three, what are the

29:29two main types of API requests? Question

29:31number four, what three things does an

29:33AI agent need to know about a tool to

29:35use it? Question five, what's the

29:36difference between a conversational

29:38agent and an automated agent? If you

29:40couldn't answer all of those

29:40confidently, I'd genuinely recommend

29:42going back and rewatching the sections

29:43that you are fuzzy on. So pause the

29:45video, go back, take some [music] notes.

29:47do not rush this. The foundations are

29:49what makes everything else click in this

29:50course. And trust me, once you get into

29:52the builds, it'll all make much more

29:53sense if you've got the stuff locked up.

29:55But if you're feeling solid on all of

29:56that, let's get into the fun part. Time

29:58to actually build some agents.

30:04All right, guys. So, jumping into build

30:05number one, we're starting off with a

30:07Telegram receipt analysis assistant.

30:09This is a pretty good general-purpose

30:11assistant that can be uh created quite

30:13easily on N8N, as I'm going to show you.

30:14And it shows you how to do a few things

30:16like extract information out of images,

30:18which is a really valuable skill to

30:19have. It shows you how to set up a chat

30:21loop that works on Telegram, so you can

30:22do any kind of Telegram deployments for

30:24your apps. And most importantly, what

30:25I'm going to teach you here, uh this is

30:26actually the second time I've shot this

30:28tutorial because I went through it the

30:29kind of traditional way of showing you

30:31how to build these nodes one by one and

30:33stack them together and connect the

30:34variables. I got to the end and I was

30:35like, that took a really long time to do

30:37a relatively basic system. And so, I've

30:40decided to take you more on the route of

30:42what the future of AI automation and

30:44building automations like this looks

30:46like. And that is using AI tools to

30:48research and plan out your automations.

30:50That's step one. I've created a GPT for

30:52you that you're going to be able to use

30:53for that. And then feeding it into a

30:55spec writing tool, which is basically

30:57taking in all the research and planning

30:58I've done, uh putting it into a few

31:00fields, and then it's going to print out

31:02a perfect N8N spec to actually pass to

31:04N8N's AI, and it will actually do a lot

31:07of the hard work for you. So, we're

31:08shifting in the AI automation space

31:10right now from needing to know every

31:11single setup and and how to connect each

31:13and every node to I know generally what

31:15I want and I can use AI to help me

31:17explore and research and figure out the

31:18best way to do that. And then I'm going

31:20to be heavily using the N8N AI feature

31:22within it to build this out for me. And

31:24when things go wrong, loop around and

31:26around and around until it gets to the

31:27point that I'm looking for. Now, this is

31:29going to be a much more future-proof and

31:31honestly faster way of learning the

31:32stuff. So, I'm excited to show you some

31:34of the tools that I've put together to

31:35help you guys learn this faster. Okay,

31:36the first handy AI tool that you're

31:38going to be using

31:38>> [music]

31:38>> in building out these automations is

31:40this GPT I've created called your AI

31:42automation CTO. This is as if you've got

31:44myself or someone else who's really

31:45familiar with N8N, um who can research

31:47the web and and help you to explore and

31:49sort of ask you the right questions so

31:50that you can get to the right kind of

31:52concept for your N8N automation. So, to

31:55save time, I've put this all in its

31:56place already. But this basically has

31:58all of the N8N documentation as a

31:59knowledge base, and it's been prompted

32:01to guide you through steps to extract

32:02the right information from you. So, you

32:04can click this button here, help me plan

32:05an N8N automation idea, and that will

32:07start the process that you're about to

32:08see that I've already gone through. The

32:09link to use this GPT along with the side

32:11the other tool we're going to talk about

32:12and all of the resources for this

32:14tutorial are going to be on my school

32:15community in the first link in the

32:16description. Head to the classroom, head

32:18to the AI foundation section, and you

32:20will see the full course for this video

32:21broken down there in chunks with the

32:23right resources at every step. So, make

32:25sure you go and grab that link now, and

32:27you can follow along. But here's how you

32:28use this GPT to plan out an automation.

32:30So, in this case, I said help me plan an

32:32idea. It asked me a few questions. I've

32:34said, this is an N8N automation for a

32:36Telegram-based assistant. Nice, that's a

32:37good front door. Uh it sort of runs

32:39through and just sort of verbal diarrhea

32:41of all the stuff that it knows about it,

32:42repeating what it's found in the

32:43documentation. And then it breaks down a

32:45practical blueprint for it. And then it

32:47asked me directly what questions it

32:48needs to move forward. So, I've gone

32:50down here and given it the feedback.

32:52It's triggered by a new Telegram

32:53message. It should be connected to a

32:54Google Sheet. Should be able to take in

32:56pictures as an input of receipts and

32:57then be able to log that info in a

32:59spreadsheet for company expenses. If

33:01over $500, then email my CFO via Gmail.

33:04I want to be able to chat to the data in

33:05the spreadsheet as well. Then it

33:06processed that, looked at its

33:07documentation, and then planned out

33:10what's going to be in the Google Sheet.

33:11This actually got a little bit more

33:12complex than I probably would have done

33:13this uh myself possible to get you

33:15through. But this is going to be

33:16interesting to see just how much more

33:17complex using AI as our assistant we can

33:20make this in even a a shorter amount of

33:21time than we would have before. So, it

33:23starts planning out the N8N

33:24architecture, plans out the

33:26conversational structure, and then it

33:28gets down to the bottom here when it

33:29asks me two tiny details, the company

33:32default currency and the CFO email

33:33address. I've given those here. And

33:35then, as you see on the side here, we

33:37have a different tool that I've created

33:38that takes in a bunch of fields about

33:40the automation purpose, the trigger

33:41details, the desired outcome, the

33:43specific providers, and the usage

33:45pattern, model preferences if you have

33:46any, and additional quotes. And what

33:48this does is runs it through a specific

33:50prompt that I've created in the

33:51background here that's going to take all

33:52that information and create the perfect

33:54prompt to pass into N8N. So, we've got a

33:56kind of like lot of prompt chaining here

33:58going on. But this is set up to be as

34:00usable as possible for you as a

34:01beginner. So, this GPT has been prompted

34:03to give you the right fields for this.

34:05So, as you can see here, I was just able

34:07to copy and paste in the automation

34:09purpose to here, trigger details,

34:11desired outcome, [music]

34:12specific providers, and so on, all over

34:13into this. Run the tool. And here we

34:15have our N8N spec, which is spat out by

34:18this tool, which again you'll be able to

34:19get it in the resources on school. And

34:20then I can just come in here and copy

34:22this, command C. And you can see down

34:23here there's actually a button that

34:24allows you to switch. It'll probably

34:25appear like this for you. Um it's

34:27actually better to click on formatted

34:29and grab the markdown formatted version.

34:31And then what we want to do is head over

34:32to N8N. You can come here to get

34:34started, put in your company email, and

34:36go through the sign-up process until

34:37eventually you see the home page and it

34:39looks like this, obviously without all

34:40of the workflows in it. Then you want to

34:42come up here and click on create a

34:43workflow. And this takes us into the

34:45editor, which we'll do a bit of an

34:46orientation in a second, but I just want

34:47to get this started as quickly as

34:48possible. So, you can click on build

34:50with AI. You'll see that we get monthly

34:51credits here. And now this does have a

34:53limit on the number of credits you can

34:54use. I have this because I have a a

34:56business plan going. But if you come

34:57over to the pricing page, you can see

34:59that you can start a free trial with no

35:00credit card required, and you can get up

35:02to 50 AI workflow builder credits that

35:04will allow you to get started as quickly

35:06as possible following the workflow that

35:07I'm walking you through here. Then all

35:08you need to do is paste in that spec

35:10from the tool that we had before. So,

35:12scrolling down and grabbing the spec,

35:14hitting back over and pasting it in

35:15here, and hitting submit. Now, that's

35:16going to work away in the background

35:18using N8N's AI to analyze what you've

35:20asked it for. It's going to figure out

35:21the nodes that it needs, and then it's

35:22going to throw them all on the canvas.

35:24Yours may vary, but I will provide the

35:25exact prompt that I gave it in the

35:27resources on school so you can try to

35:28follow it as closely as possible. But

35:30the thing about this is that you need to

35:31be a little bit more flexible with how

35:33you think about it. There's There's no

35:34one right answer necessarily. There's

35:36definitely simplified versions. This one

35:37that I'm looking at here that it's

35:38generated for me is definitely on the

35:39more complex side of how you could

35:40approach something like this. But often

35:42the AI will actually opt for a lot more

35:44strict and kind of non-AI features to

35:46make it a lot more deterministic. Now,

35:47this is important because AI models

35:49themselves are language models are

35:52non-deterministic. That means that you

35:53can give it the same input and there's

35:54no guarantee that it would generate the

35:56exact [music] same output, which is

35:57going to be a bit of an issue when

35:58you're trying to build reliable business

35:59systems. And so, the N8N AI leaning

36:02towards uh things that are much more

36:03deterministic where possible means that

36:05it tries to shrink the room for error

36:06down, which is actually a good thing and

36:08good practice in building automations.

36:10So, while this would be possible to do a

36:11lot simpler, I'm going to kind of go

36:12with the flow here. When you're using

36:14the N8N AI builder, it'll leave you with

36:16a few things here like kind of to-dos,

36:18which uh the last few things that you

36:19need to link up correctly in order to

36:21get it to work. What I like to do is

36:22just give this a a sort of skim through

36:24and try to understand each of the steps

36:25it's doing, maybe ask a few questions to

36:27help understand it better, and then I

36:29can move forward with actually going

36:30into testing. So, we start here with a

36:32Telegram trigger. That's going to be

36:33what we set up where it's sent a

36:34message. We're doing some workflow

36:35configurations here like setting the

36:37Google Sheet ID and the drive folder.

36:39That's okay. It's actually good practice

36:40to have these set first because then you

36:42can just change one node and everything

36:44that references those downstream will be

36:46updated in one go. Then we have a check

36:48of message type, so it's going to look

36:49to see if the Telegram message was a

36:50photo or a message, which makes sense

36:52because we've kind of got two different

36:54modes for this, which is chatting to the

36:55data or we're going to be sending a

36:56receipt photo for extraction and logging

36:58in the spreadsheet. So, here we have it

37:00going to the spreadsheet and trying to

37:01pull all the data from the spreadsheet,

37:02which is what we want. Then we have our

37:04AI agent, which is prompted to uh answer

37:06questions, which makes sense. It's

37:07already got the Google Sheet data here.

37:09And then it's going to send the response

37:10back to Telegram, so that'll make sense.

37:12Then we have the extract receipt data.

37:13Now, this is a bit more complex, but

37:14it's telling the agent that it needs to

37:16analyze a receipt, extract the following

37:17fields, return a confidence score, and

37:20some more in details about the

37:21formatting of those responses as well.

37:22So, that all seems to be pretty good.

37:24We're using the Gemini 2.5 Flash model,

37:26which is what I actually put

37:27specifically in this section here around

37:30the model preference because we needed a

37:31multimodal model out of the box so that

37:33we could just use Google Gemini 2.5

37:35Flash, which is really cheap and

37:36affordable. Um that requires a little

37:38bit of researching or just chatting with

37:39the uh the assistant here to figure out

37:41which model would be best for this. But

37:43you can see here it's actually not using

37:44the current model, so I want to come

37:45down and select the right one. Now, for

37:47you guys, you won't actually have your

37:48Google account set up correctly, which

37:50is something we're going to do in a

37:51second. You will see in the N8N AI down

37:53here it'll be saying, "Hey, you need to

37:54set up your Google. Need to set up your

37:55Telegram. Need to set up your sheets,"

37:57which are all things that we're going to

37:58do shortly. Then in order to be able to

38:00take the information in that receipt

38:01image and then pass it further down, we

38:03need a receipt data parser, which is

38:05just going to make sure that it's

38:06sending the data based off that receipt,

38:08turning that image into what's called

38:10structured data or JSON here. And this

38:12means that the language model is going

38:13to fit all of that information into

38:15something that we can easily interpret

38:17and then use downstream as variables.

38:19So, this is a very helpful skill to know

38:21in automation, and that's using an LLM

38:23to turn some sort of unstructured data

38:25into structured data and then passing

38:27that downstream into what's called a

38:28parser. And the parser is going to take

38:31I'll be able to read this JSON right

38:32here and then give us variables that we

38:34get to play with downstream. So, it's

38:35very important and we'll be using this

38:37quite a lot. And then we have a little

38:39format message, so it's just taking the

38:40variables. So, you can see here already,

38:42we've taken these vendors and expense

38:44date and currency, and all those things

38:46that we just extracted with Google

38:47Gemini 2.5 Flash, and then passed it

38:49through that parser, and out came these

38:51variables that we can now use. And then

38:52in this case, we're just turning this

38:53into a little message that can be sent

38:55back via Telegram. And it's asking them

38:56to please confirm or edit, reply with

38:58confirm to accept or vendor with the

39:00changes if we want to edit. So, that's a

39:03nice little feature that I didn't

39:03actually have in the original version of

39:05this build. And it's going to send that

39:07message off. Nothing too fancy [snorts]

39:08there. Now, this is where things get a

39:09little bit more hairy, if I'm honest. I

39:11saw this and I was like, "Hmm, this is

39:13waiting for a user confirmation. It's

39:15waiting for a webhook call." And I

39:17didn't see this webhook call actually

39:19get used anywhere else. And so, [music]

39:20in my little skim over before, I went to

39:22the N8N AI and so I said, "Basically,

39:24explain this to me. I don't understand.

39:26Can you explain how the wait for user

39:27confirmation works?" It goes, "Great

39:28question. It works by this, this, this."

39:30And then it spotted an issue in its own

39:32workflow. So, the tricky part is that

39:34when a user replies to Telegram, that

39:35reply comes as a new Telegram message.

39:37It doesn't automatically go to the wait

39:39node's webhook. You need to bridge this

39:40gap. Then it gave me two options, and

39:42the first one seemed great, so I'm going

39:44to say, let's implement option A. And

39:46the thing is, as you're going through

39:47this yourself, I obviously have a bit

39:48more information on how this all works,

39:50you get to go through here at every step

39:52and explain, hey, what is an output

39:53parser, and what does that mean? You

39:55need to really, really get used to

39:56relying on this n8n AI, because it is

39:58going to be your best friend in learning

40:00this platform as quickly as possible. So

40:01here we can go, it's planning out the

40:03steps to implement option A.

40:06Okay, so now we actually get to an

40:07important decision that you need to make

40:09as the automation builder here, the one

40:11steering the ship, and that is here it

40:13is actually asking us to use another

40:15external platform called Redis, which

40:17I've just looked up and appears to be

40:18some kind of lightweight database

40:20service for managing real-time data. And

40:23if I look at the pricing, they do have a

40:24free tier, but if I'm being completely

40:26honest, guys, this is supposed to be the

40:27introduction tutorial for this video,

40:29and doing all of these external

40:31integrations is actually a little bit

40:32too much for just easing you into it.

40:34So, I'm going to show you how to

40:35actually go back and rethink this, and

40:37[music] go back to the GPT and ask it to

40:39plan it in a slightly different way, cuz

40:41this is a really good example of what

40:43these different kind of, when you're

40:44heavily using AI for planning and

40:46building, there's so many like I've

40:48actually done this multiple times, and

40:50every single time it's given me kind of

40:51a different way of doing it, like I said

40:52before. And so it's an important skill

40:54for you to have is to be able to go back

40:55to this planning phase. So I'm going to

40:57ask the GPT and say, hey, is there any

40:58way we can simplify this, like a single

41:00agent with a few tools, and taking out

41:02the complex functionality like the

41:03confirmation message. So if we send

41:05this, so here it's giving me two

41:06different options, a simple sort of

41:08faster version, and one with a bit more

41:09cautionary around receipts. So I'm going

41:11to say, let's go with option A, give me

41:12the fields to put into the form,

41:14referencing the form that I've got over

41:15here. Then I can just grab these one by

41:17one. So I've got these all filled in

41:18now, I can just run the tool again. And

41:20actually I see a few things in this plan

41:21that I'd like to remove as well, so I'd

41:23say, so my last touch up here is let's

41:24remove the Google Drive saving feature.

41:26Let's make it as agentic as possible

41:28with the sheets and Gmail tools all

41:29connected to the agent for simplicity.

41:31That's going to make the automation look

41:32a lot smaller and simpler. And let's

41:34also use Gemini 2.5 Flash for the model

41:36as it's multimodal. That means it can

41:38take both image and even video and text

41:40as inputs without an issue. So with all

41:42these added in now, I can click run

41:44tool, wait a few seconds, and we'll have

41:46our new brief to pass into n8n. So I'm

41:47going to copy this.

41:53And we're going to go back and create a

41:54new workflow, open up the n8n AI, and

41:57then paste that in again.

42:02Nice. And now as I said before, it's

42:03actually easier to have a lot of these

42:05things handled within this. So when you

42:07change it here, it automatically updates

42:08all of these, for example, the Google

42:10Sheets. I only want to set the Google

42:12Sheet value once, and then it can be

42:13passed down through everything, nice and

42:15easy. So it's good practice to have. So

42:16I'm going to ask it, can you set up the

42:17workflow configuration steps so that I

42:19can add most of these in one go. So and

42:21I'm and I'm sure you can see how much

42:22simpler we got this down. This is kind

42:23of what I was trying to show you, that

42:25there are very advanced and complicated

42:26ways of doing things, and then if you

42:27can prompt n8n or prompt the sort of

42:30spec writer that we have to opt for

42:32simpler and often more agentic ones,

42:34it's going to look a lot less complex.

42:35Cuz as you can see here, we're getting

42:37the agent to do a lot of the heavy

42:38lifting, and it has to decide when to

42:40use the sheets, when to use the Gmail

42:41tool, when to send a message back. And

42:44so this AI agent node here is able to

42:46intelligently use the right tools when

42:48needed, rather than having these big

42:49long chains of like JSON parsing, and if

42:53this happens and do that, which yes, it

42:54is more deterministic and likely more

42:56reliable. For most use cases like this,

42:58you can rely more on using intelligence

43:00and tokens rather than trying to strip

43:02it all down to the most deterministic

43:03setup possible. Here we can see we've

43:05narrowed this down to just setting up

43:06this node now. I'm going to hide the AI

43:08assistant for now. And now we have

43:09things like the CFO email, when I want

43:11to escalate it, where do I want that

43:12email to go? I can just plug it in here.

43:13What's the expense threshold for sending

43:15an email to my CFO? That's $500. Default

43:18currency, USD. Google Sheet ID, which

43:20we'll need to set up in a second. Sheet

43:21name, which is here. Sheet headers, so

43:24on. Now to get this working, all we need

43:25to do is configure each of these nodes

43:27so that we can give it a test. So I'm

43:28going to start with Telegram here.

43:30You'll need to come to create a new

43:31credential here, and you're going to go

43:33to Telegram. So you want to download and

43:34open up Telegram. It's free to use, you

43:36can get it on your phone or your laptop.

43:38Then to create a new bot, you're going

43:39to want to search up here for BotFather.

43:41Click on the chat here, and then you can

43:43come in and type new bot.

43:45And it's going to take us through a

43:46setup process to get this new bot set

43:48up. So what's the name? We can call it

43:50Biz Receipts

43:52Bot.

43:54Let's also call it Biz Receipts Bot.

43:57And there we go, just like that we have

43:59our token. I'm going to copy that, hit

44:01back to n8n here, paste in the access

44:03token, and then click save. So that's

44:05super easy compared to some of these. If

44:07you do need help at any point when

44:08you're getting your integration set up,

44:10n8n AI here is actually really helpful.

44:12You can click here, and it's going to

44:13use AI to look at the documentation, get

44:15you the latest information on how to do

44:17this. And if you get stuck, you can chat

44:18here. So again, learning to use AI at

44:20every step and rely on it to help you

44:22get through roadblocks is really

44:24essential. As you can see, it's got the

44:25exact steps we just went through. So

44:27we'll click save there. It's going to

44:28test it for us, make sure that it's

44:29working correctly, and we have this

44:31green box, so we're ready to go. So

44:32that's our Telegram bot set up just like

44:34that. It's triggered on message, and

44:35it's going to automatically download the

44:36files and images, which is great, makes

44:38it a lot easier for us. And then we want

44:39to work through the rest of these

44:41configuration steps, like firstly the

44:42Google Sheet ID. So we can make this a

44:44lot easier for us by clicking into here,

44:47grabbing these sheet headers, and then

44:49going to Google Sheets.

44:51Now one thing to note about using Google

44:53Sheets and anything kind of Google

44:54related, that in order to integrate it

44:56with n8n as quickly as and easily as

44:58possible, you need to go to what's

45:00called a Google Workspace. This is

45:01basically a business account on Google

45:03that allows you to access Docs and

45:04Sheets and stuff through your through a

45:06business account rather than a personal

45:08one. And it's really important that you

45:09do this. And set it up, it does have a a

45:1114-day free trial. Um you can try it

45:13free for 14 days. You can come on, get

45:15on the starter one here, and start a

45:16trial. I highly recommend you do that,

45:17or it's going to take you a lot more

45:19effort to get these integrated in a

45:20second. But in this case, I'm on my

45:22Morning Side AI account, so I've got

45:23that all set up. I'm going to go to a

45:25blank spreadsheet here, and paste in

45:27these fields that we had before.

45:28Actually, put them down here. And then I

45:30can just start to grab these fields out,

45:32start popping them in. Then I have all

45:33of these columns set up correctly in the

45:35sheet. I'm going to just bold them to

45:37make it easier to look at. I'm going to

45:38call this company expenses tracker, and

45:40I'm going to name this sheet one

45:42expenses. [music]

45:43Then I'm going to grab the URL of this,

45:45and double-click this segment of it. As

45:47you can see, after the slash D, we want

45:49to copy that ID,

45:51hit back to n8n. And in case you got

45:53lost there, I was grabbing the sheet

45:54headers out of this section here. And so

45:55then what I need to do is paste in our

45:57Google Sheet ID that we just took from

45:59the webpage there. We've got the sheet

46:00name set up as expenses, which we had

46:02correctly here. And then we have our

46:03workflow configured correctly like that.

46:05Now we've got a few more configurations

46:06to set up. We have firstly the Google

46:08Sheets. You can go to create a new

46:10credential here, and here you can see as

46:12it says sign in with Google, that's

46:13connected to that Google Workspace. So

46:15if you're trying to connect with a

46:15personal account, it's going to walk you

46:17through a much more complex setup, so

46:18please, please get that Google Workspace

46:20set up. It's going to take you a few

46:21minutes. Do that. It's got a 14-day free

46:23trial. Come back, sign in here with that

46:25same Google account you created on your

46:26Workspace, and then you will have

46:28correctly set up your Google Sheets

46:30account here. And that means that n8n is

46:32going to be able to access that ID of a

46:34sheet that we gave it before. And here

46:35you can see it's inserting that ID that

46:37we set up just before into this

46:38document, and also the sheet names. You

46:40can start to see how when we set those

46:41variables earlier, they can be passed

46:43downstream a lot easier. And the key bit

46:45here to note that the n8n AI has

46:47actually correctly set up the

46:48descriptions of the tools. As I

46:50explained earlier in the video, tools

46:52have kind of descriptions, and they have

46:53information on the different variables

46:55that need to go in them. And then this

46:56tool description, which we are setting

46:57manually here, we are giving it all the

46:59information about what this tool does,

47:01what it takes in, how it's connected to

47:02the rest of the workflow, so that our AI

47:04agent here, with its system message,

47:06also understands what each and every one

47:08of these tools does. The Google Sheet

47:10for adding to the

47:12reading the whole spreadsheet for when

47:13we're asking questions over the data,

47:15and the Gmail tool for sending it, and

47:16the Telegram reply, and so on. So that's

47:18one of the reasons why the n8n AI can be

47:20so powerful when you prompt it

47:21correctly, because it would do all that

47:22context setting for you

47:24rapidly. So the same thing needs to be

47:25done for this Google here. So you can

47:27come to create a new credential, go

47:29through the same process, sign in with

47:30that new Workspace email, and then

47:32you're going to have your Gmail

47:33connected, and it's going to be sending

47:34from that email that you set up. And

47:35then finally down here, we have the

47:37Telegram set up, and we can connect that

47:38to the Telegram account that you just

47:39created. And now the final piece of the

47:41puzzle is to connect up our Google

47:43account correctly so that we can

47:44actually use the model that we want to.

47:46I'm going to change this model here

47:47because it struggles to use the exact

47:49one that you asked for, so we're going

47:50to use Gemini 2.5 Flash then to actually

47:52be able to use these APIs, you need to

47:53create a new credential here, and then

47:55you need to go to Google Cloud.

48:00You'll be able to log in here with your

48:01Google Workspace account. Then you go to

48:03the console.

48:04Then you want to come up here and click

48:06on a new project if you haven't got a

48:07project already. Then once you've set up

48:09the project, you want to head to billing

48:10account management. You'll likely need

48:12to create a new billing account. You'll

48:13want to go to payment method here. Then

48:15you will need to add some sort of debit

48:16or credit card in here, cuz that's how

48:17you're going to be charged for it. You

48:18can set all the limits you want, but for

48:20what we're going to be doing, it's not

48:21going to use much at all. You might be

48:22charged a dollar or two max. So get your

48:25card added in here. This is going to

48:26mean that you're able to create an API

48:27key, and then we can head back to our

48:29cloud console. You can see that we're in

48:31the project that you just created, and

48:32then we're going to want to head down to

48:34APIs and services.

48:37Enable APIs and services up the top

48:39here.

48:40We're going to search for the Gemini

48:42API.

48:44Click on this.

48:45Yours will say enable here, so you want

48:47to click enable, then go into manage.

48:51If you hit to credentials on the left

48:53here, then create credential, click API

48:55key.

48:56Now we have our API key. You can copy

48:58this, hit back over to n8n, paste it

49:00into this API key section, and then

49:02click save, and it'll do a little test

49:04on it, make sure it's working. You may

49:05need to go back and forth to make sure

49:06your billing is set up and it's able to

49:08charge you correctly. Once you've got

49:09that, you can click save, and you will

49:10have connected your Google Workspace and

49:12Cloud account to n8n to be able to use

49:14that in all future workflows. And one

49:16thing to go over here is this

49:17conversation memory. This suggests that

49:18we can have a bit more of a

49:19conversational assistant, and it's going

49:20to keep the last 10 messages in the chat

49:22in the context so the agent understands

49:24what's happened before. So that's just

49:25helpful for a conversational agent.

49:27Sometimes if it's non-conversational,

49:29then you can just leave that out. Now

49:30with all this set up, you can just pop

49:32into the system message here and make

49:33sure that it's all prompted correctly,

49:35cuz this is kind of the glue that holds

49:37all of this functionality together. You

49:38are an intelligent expense tracking

49:40assistant. Help users log expenses from

49:42receipt photos. Your capabilities

49:43extract expense data, log expenses,

49:45answer questions, send CFO alerts, reply

49:48to users. And it breaks down the

49:49workflow in full details there. Now, all

49:51that's left to do is the moment of

49:53truth. Uh we're going to save this. We

49:54are going to click execute workflow.

49:56Now, this means it is waiting for the

49:58trigger, so we can send a message on

49:59Telegram and it's going to ping this and

50:01it will receive it. This is just a way

50:02of testing it. It's not actually live.

50:04It's not going to quite every message

50:05yet. So, we can go into our chat here

50:07and say, "Hey man." Just to test it. And

50:09[music] yep, it has received our

50:10message, but we're getting an error

50:11here. So, this starts the process of

50:13using the AI agent and assistant in N8N

50:15to fix things. So, this is very

50:17important to see this live as we go. We

50:19can click open node here. No session ID

50:21found. Expected to find a session ID as

50:23an input called session ID. So, we can

50:24just click on the N8N AI here. It's

50:26going to analyze the error, give us a

50:28recommended fix, and then actually do it

50:29for us. So, again, I'm trying to show

50:31you the AI-based workflow for this. And

50:34then let's try to plan it out. Let's try

50:35to get the N8N AI to build it. Let's get

50:37it to explain it to us if we don't

50:39understand certain parts.

50:40So, to get this built out, I'm going to

50:41copy all of this. I'm going to go to the

50:43build section. Then I'm going to say,

50:45"Paste it in here. Fix this session ID

50:48error."

50:51So, the session is just an ID that's

50:53linked to our account. So, whenever we

50:55are messaging with the bot, it's like,

50:56"Okay, well, what conversation is this?"

50:58Cuz this can handle a thousands and

50:59thousands of users through this one

51:00automation. So, we need to be able to

51:02track which user we're talking to so

51:03that we add the messages to the right

51:05chain of messages that we're building

51:06for each user. So, here we can see it's

51:08fixed the conversation memory session

51:09ID,

51:10um changing it from input to custom tick

51:12key mode. Memory will now probably use a

51:14Telegram ID from message ID. So, we can

51:16execute and refine. Go back to Telegram

51:18here and say, "Hey man."

51:21There we go. It's working correctly. I

51:22couldn't understand a request. Please

51:24resend a receipt photo or text like,

51:26"So-and-so." Or ask a question like,

51:27"How much did I spend last month?" So,

51:29it's analyzed that and realized that

51:30it's not something that it's supposed to

51:31handle, and it's asked us for the

51:32correct input. And as you can see, to

51:34break down how this is displaying what's

51:36happened to us. Uh we have a green tick

51:38here, a green tick here, a green tick

51:40here, and then it's gone and used the

51:41Gemini Flash twice. [music] It's checked

51:44the conversational memory multiple

51:45times, which is a bit strange. And then

51:47it's used this tool once to send that

51:48response via Telegram. Now, all that's

51:50left to do is to test the rest of the

51:51functionality. So, I'm going to close

51:52this off. We can click execute workflow

51:54again. And this time I'm going to send

51:56it a receipt. So, I'm going to drag this

51:57in here. Send it.

52:01And we see it will should analyze it

52:02with this, extract the information, and

52:04then Oh, so we have an issue with the

52:07Google Sheets account. So, again, we

52:08just go through the process asking the

52:10N8N AI to help us with this. And as

52:12expected, it's just going to tell us to

52:13double-check our credentials. So, I'm

52:14going to go through this process again,

52:16create a new credential. We're going to

52:17rename this to latest.

52:20Save. Make sure both of these are on the

52:22latest.

52:24I'm going to save. Make sure you're

52:26saving the workflow constantly because

52:27it needs to save in order to upload

52:29those changes to the actual deployment.

52:49We should see over here. Boom. There we

52:52go. So, we have added in the information

52:54to the spreadsheet successfully. I'll

52:55just drag these out so we can see the

52:57receipt file ID, the vendor, expense

52:59date, currency, total tax, everything

53:02that we asked for here. Even the items

53:04that were on that receipt. It's got a

53:06high confidence as well, which is great.

53:07So, we have successfully extracted

53:09information from that first receipt and

53:11put it into a spreadsheet. So, that's a

53:12big win. Interestingly, it's chosen to

53:14read the sheet as well. I don't know why

53:15it's reading the sheet before it puts it

53:16in. If I was going to keep tweaking this

53:18around, I'd make sure that it wasn't

53:19doing that. But for now, it seems to be

53:20working fine. Maybe it's just getting a

53:22read of the columns so that it can

53:23better fill it out. At this point, I'm

53:25not too bothered. It seems to be

53:26working. Now, another thing we want to

53:27check is the ability for it to escalate

53:29it to my CFO if it's over $500. So, I've

53:31got this set up on a Google Workspace

53:33email as I recommended here. Then I'm

53:34going to execute the workflow again. Oh,

53:36and as you can see, we're getting a

53:37response back, which is great. Giving

53:38the user a bit of feedback on what's

53:40happening. And here I have an Apple

53:41Store receipt for a uh healthy 3,500

53:44euros.

53:46I'm going to send this here. And there

53:47you go. It looks like it didn't even

53:48need to read the sheet this time because

53:49it had already read it in the last

53:51session. It remembers the last message

53:52we've sent and the ones before it. So,

53:54maybe it already knew what the sheet

53:55looked like. And it seems we're getting

53:56more issues with authentication here.

53:58So, I'm just going to set this up again.

54:14Save that.

54:16Save the workflow. Execute it again.

54:22And send our receipt.

54:26Yep, and we've seen an email sent off

54:28and logged in the spreadsheet as well.

54:30Boom. There we go. Over 500, true. And

54:33this row here was from the failed ones,

54:34so we can probably delete that.

54:36But here we go. Our CFO email sent at

54:39It's officially sent an email and logged

54:40this in here. And it's detected that

54:42it's over $500. So, now if I go to

54:44email,

54:46then I've got an email from myself here

54:47saying high-value expense alert at Apple

54:49Store Grafton Street. High-value expense

54:51has been logged. So, this is going to be

54:52a helpful way to send off to your

54:53finance department or CFO whenever

54:55there's a big expenditure so that they

54:56can follow up on that and make sure that

54:58that's an approved purchase. So, now

55:00that we've got all this working, the

55:01last thing to do is ensure that we're

55:02able to chat to this and it's able to

55:03pull that information from the

55:04spreadsheet and make it ask questions

55:06over that data. What I'm going to do is

55:08just grab this here. I'm going to grab

55:09all the information in the spreadsheet,

55:11head back [music] to our handy-dandy

55:13assistant, and ask it, "Can you create a

55:17bunch of dummy data in a table and add

55:21to our sheets?"

55:28And I can turn off thinking here.

55:37So, I can copy those data out. I can

55:38head back to the spreadsheet and I just

55:39click in here, paste it in. Now, I've

55:41got a bit more to work with, a couple

55:43more categories and so on. And then we

55:44can head back to Telegram. I'll execute

55:46this workflow again. I'll say, "What is

55:49the total spending on this account?"

55:52Now, it should query the data, pull that

55:54all back to the agent, and boom. Total

55:56spending on your account is 5,000 USD or

55:59and a 3,000 total. So, everything is

56:02working as expected. [music] That's

56:03great. Now, we can click on this toggle

56:05here to make it live. That means that

56:06instead of having it executed every

56:07time, it's going to run with every

56:09message. And then we have our buddy here

56:11to chat away with, and we can say, "How

56:13many trans

56:18So, as you can see, we now have a

56:19handy-dandy expense assistant put onto

56:22Telegram uh that's able to add things to

56:23a spreadsheet from a image. Very

56:26powerful skill to have. We've heavily

56:27used the N8N AI and the AI tools that

56:29I've created for you guys to plan this

56:31out, to iterate, to go back and forth,

56:32to use the N8N AI to also debug and work

56:35through issues. You've just seen a

56:36realistic walk-through of what the stuff

56:38looks like when someone like myself or

56:40other AI automation engineers are

56:41building these systems. So, so that is

56:43build one under the belt. I will be

56:45giving you guys this template exactly so

56:46that you can import it yourself. If you

56:47don't want to do everything yourself,

56:49I'll also be providing those final

56:50prompts that we gave to the AI so that

56:52you guys can try to follow along as

56:53closely as possible with the experience

56:55that I've gone through here. This has

56:56been a great skills foundation for what

56:57we're going to be doing in the next

56:58three builds, which are kind of linked

57:00together as a a chain of different

57:02automations and systems that you can

57:03build for a given business that work

57:05together as part of a larger system. All

57:07right, guys. So, before we build our

57:08second agent workflow, I want to talk

57:09you through hosting since it's

57:10particularly important that you

57:11understand where your workflows are

57:13actually going to be running once you've

57:14got them set up and deployed for your

57:16clients. So, in order for things to run

57:1724/7, your workflows have to always be

57:19on, which means they need to be hosted

57:21in some kind of service somewhere. So,

57:23when we signed up to N8N and when we

57:24started building, we were automatically

57:26putting N8N on the N8N cloud. So, with

57:29the N8N cloud, you're basically

57:30borrowing some space on their server.

57:32It's like renting an apartment, but it's

57:33for running your workflow. They host

57:35these workflows for you automatically

57:37with no setup required for the hosting,

57:39but you pay based on usage. So, this

57:41works fine when you're learning and just

57:42starting off and building your first few

57:44workflows just like this. But as you

57:45scale up and particularly start to deal

57:47with clients doing a lot of volume,

57:48you're going to hit walls where your

57:49workflows might crash with a large

57:51amount of data. Your bills can spike

57:53unexpectedly. You can't install certain

57:55tools that you need. And ultimately,

57:56you're stuck within N8N's limits and

57:58their pricing structure as well. So, for

58:00running an actual AI agency or AI

58:02automation agency powered by N8N and

58:04selling those workflows to your clients,

58:06pretty much every AI agency I know is

58:08self-hosting their N8N and using that to

58:10run their client workflows. Or in other

58:12words, they own their own home for their

58:14workflows versus renting an apartment in

58:16N8N's cloud. So, when you self-host, you

58:18can run as many workflows as you want,

58:19basically. And instead of paying

58:21surprise bills to be N8N, you pay a

58:23predictable and usually much smaller

58:25monthly fee just for the server access.

58:27You also get the freedom to install any

58:28community tools that you need. And most

58:30importantly, you control everything

58:31about your infrastructure. But if you

58:33had to set all of this self-hosting

58:34stuff up from scratch, that would be a

58:36lot of complicated work for beginners

58:37particularly. But fortunately, there are

58:39a lot of services out there that handle

58:40all of the technical complexity for you

58:42of this self-hosting in order to get the

58:44most scale out of your automations. So,

58:45when it comes to self-hosting,

58:46especially for beginners, Hostinger is

58:48my recommendation since it does

58:49basically all of the heavy lifting for

58:51you automatically with literally like a

58:52one-click N8N installation. And

58:55fortunately for me, Hostinger has

58:56offered to sponsor this video. So,

58:57shout-out to Hostinger for being the

58:58sponsor of this video. So, let me show

59:00you how to get your self-hosted N8N up

59:02and running if you plan on scaling your

59:04agents professionally as an AI

59:05automation agency. So, below this video,

59:07there's going to be a link that will

59:08send you to Hostinger's self-hosted N8N

59:10page. They've got a page specifically

59:11for it. You can select whatever tier you

59:13want, but KVM 2 is a good price for the

59:15value you get. So, uh we can choose this

59:17plan, and then we can select either a

59:18one-month or 12-month or 24-month for

59:21better savings. Your server location of

59:22where you're hosting your systems uh

59:24will automatically be chosen based on

59:26your location. And the closer the server

59:28is physically to you, the better

59:29performance you're going to get. Now,

59:30down here, you'll see N8N was already

59:32selected for me, so I can leave that as

59:33is. And before heading to checkout, you

59:35can get an extra 10% off with code Liam

59:38Ottley, all caps. All right, so you can

59:39put that in there and get an extra 10%

59:41off your hosting uh setup. Then you just

59:42register your account. You can fill in

59:43your billing details and enter your

59:45payment info. Now, we just need to add a

59:47root password. Make sure you save this

59:48somewhere safe. And next, we can get

59:50malware scanning for free, which is

59:52great. And I'm going to select that

59:54before finishing things up. Now,

59:55Hostinger is doing all of the complex

59:57setup for us automatically, which is

59:58awesome. And once it's finished up,

1:00:00we'll be routed to the dashboard. We'll

1:00:02go from there. Now, from here, I just

1:00:03click on manage app. I'll be taken to my

1:00:05own self-hosted n8n instance. So, here

1:00:07is where you enter your info and get

1:00:09your instance up and running. Once

1:00:10inside, we have the option to receive

1:00:11some free features for advanced

1:00:13debugging, search, and organization. And

1:00:15now that I'm in, I can go ahead and

1:00:16create my n8n workflows from here. So,

1:00:18either using a provided template or

1:00:19creating from scratch and building node

1:00:21by node. Now, I have unlimited

1:00:22automation power, running 24/7 with

1:00:25predictable cost. So, that's a key thing

1:00:26for turning this into a business.

1:00:28Occasionally, n8n will release versions

1:00:30with new and useful features, so you'll

1:00:31want to update your n8n from inside your

1:00:33Hostinger account, which there's a

1:00:35tutorial for right here. And if you want

1:00:36to customize your domain, you can follow

1:00:37the guide here for a custom URL that

1:00:39your workflows actually live at. And if

1:00:41you need to reset your password, you can

1:00:42find out how to do that here, of course.

1:00:44These are some helpful guides as well,

1:00:45but what's especially helpful, I find,

1:00:47is the AI agent called Cody, who can

1:00:49walk you through basically anything on

1:00:50the Hostinger platform, which is super

1:00:51great for beginners. Now that you have a

1:00:53solid understanding of the benefits of

1:00:54self-hosting and how to get it up and

1:00:55running, let's get back to building our

1:00:57next AI agent workflow.

1:01:01All right, guys, so we're getting into

1:01:02build number two now, and this is the

1:01:03start of the chain of business-focused

1:01:05builds that we're going to be doing, so

1:01:06it's important you get this right cuz

1:01:07they are all kind of interrelated and

1:01:09part of a bigger system or package that

1:01:11you can sell. And so, this one is a

1:01:13throwback to one of my favorite builds

1:01:14that I've ever done on the channel. It's

1:01:15sort of updated and refreshed to be uh

1:01:17more current and new, but the use case

1:01:19itself is still rock-solid and one of

1:01:20the top ones that you can sell to

1:01:22businesses, and that is a lead

1:01:23generation chatbot that helps to answer

1:01:25questions about the business and about

1:01:27the services on the website. In this

1:01:28case, we're going to be building an AI

1:01:30tool that allows them to pass in their

1:01:31monthly power bill and their address,

1:01:34and we're going to be able to return to

1:01:35them, based off the Google Solar API,

1:01:37which is really the powerhouse of this

1:01:38build. We're going to convert their

1:01:39address into a latitude and longitude

1:01:41coordinate. We're going to pass that

1:01:42into the Google Solar API, and it's

1:01:44going to return It's going to basically

1:01:45look up their house, uh determine what

1:01:47shape their roof is, the different sides

1:01:49on it, estimate how many panels could

1:01:50fit on it, and then also, based off

1:01:52their monthly power bill, be able to

1:01:54estimate a range of different solar

1:01:55packages that they could buy, and the

1:01:57estimated payback periods it'll look up

1:01:58based off their location, how much sun

1:02:00those roofs are going to get, and also,

1:02:02look up based off their state and

1:02:03location and area, any tax incentives uh

1:02:06from the federal government or the state

1:02:07government, and be able to give them a

1:02:08full comprehensive picture of how much

1:02:10they could potentially save based off

1:02:11the power bill and location, sun hours,

1:02:14savings, all of the stuff that is really

1:02:16important for someone to know before

1:02:17they go in and make a solar purchase.

1:02:18So, this is a way of creating a what

1:02:21would essentially be a human kind of

1:02:22solar consultant task to take in the

1:02:24information [music] about them,

1:02:25determine what packages they can look

1:02:26at, and we're going to be able to

1:02:27automate this fully through AI within a

1:02:29chatbot, so that all users can come on,

1:02:31ask a few questions, the chatbot's going

1:02:32to offer them, "Hey, do you want to get

1:02:34a perhaps a a quotation or estimate on

1:02:36how much that you could save and how

1:02:37much solar could cost for your house?"

1:02:38And then after they've gotten that

1:02:39value, we're going to prompt the agent

1:02:41to afterwards ask for, "Hey, if this

1:02:43seems interesting, I can put you in

1:02:44touch with one of our representatives,

1:02:46and they can take you a bit further on

1:02:47this process um to learn more about what

1:02:49this would mean and and look like if

1:02:50you're going to roll this out for your

1:02:51house, different types of panels we

1:02:52could set up, and so on." We're going to

1:02:54be learning a ton of super valuable

1:02:55skills in this that you're going to

1:02:56carry forward into all of your future as

1:02:58an AI automation expert, being able to

1:03:00call custom APIs. In this case, we're

1:03:01calling the Google Solar API and the

1:03:03Google Geocoding API as well. I'm going

1:03:04to teach you a very key skill in making

1:03:06your agent more powerful, which is using

1:03:07the workflow as a tool node, which

1:03:09allows us to build out complex workflows

1:03:11and then connect [music] them as a tool

1:03:12on our agent. I'm also going to be

1:03:13showing you how to set up a simple

1:03:15knowledge base within n8n so that you

1:03:16can load documents in and allow your

1:03:18agent to answer questions over those

1:03:19documents. Great skill to have. And

1:03:21finally, I'm going to be showing you how

1:03:22to take everything we've built here and

1:03:24set it up on a chat widget on a website

1:03:26so that you can take it out of n8n and

1:03:28put it onto a website for customers to

1:03:29use and give you a new deployment option

1:03:31other than the Telegram one we've

1:03:32already done. A website deployment for a

1:03:34chatbot is a very useful one to know how

1:03:36to do. So, to start things off, we're

1:03:37going to follow the same process that we

1:03:38went through last time. We're going to

1:03:39be using the GPT and the relevance tool

1:03:41that I've created here. We're going to

1:03:42be chatting back and forth with this AI

1:03:44automation CTO to figure out how we can

1:03:46uh get the right plan. We're going to

1:03:48feed that into our automation brief, and

1:03:49then we're going to go from there into

1:03:50setting it up with the n8n AI, testing

1:03:52it, [music] and then deploying it to a

1:03:53website. And this is the platform we're

1:03:55going to be using to take out n8n agent

1:03:56and put it on a website, so I'm just

1:03:58going to copy the URL of this and pop it

1:04:00in here as context. Okay, I'm starting

1:04:01off a conversation with this prompt

1:04:02that'll be included in the resources if

1:04:04you want to follow along. Basically

1:04:05saying a a rough scope of the build. I'm

1:04:07trying to build a lead generation agent

1:04:08for a solar company's website. It's

1:04:10going to be deployed via a chat widget

1:04:12using n8n chat UI. It's going to be able

1:04:14to answer questions from a knowledge

1:04:15base, which is going to be basic,

1:04:16[music]

1:04:17about the business and guide the user

1:04:18towards a free solar estimate, which

1:04:20will be done through the Google Solar

1:04:21and Geocoding APIs. We will take in

1:04:23their monthly electricity bill and also

1:04:24their address, then send that into a

1:04:26workflow as a tool AI agent tool. Um

1:04:28you'll see what that means in a second.

1:04:29That we use those APIs to return us a

1:04:31simplified breakdown of the relevant

1:04:32data to them from the Solar API. The AI

1:04:35is going to give them a nice summary of

1:04:36all of the information in the Solar API

1:04:38that that they've returned to us. And

1:04:39then after that estimate has been given,

1:04:41the AI will suggest that they provide

1:04:43their phone number and a name for one of

1:04:45our representatives to get in touch with

1:04:46more information and special offers uh

1:04:48to get them sort of encouraged to give

1:04:50their uh information. And then this lead

1:04:52information will be added to an

1:04:53airtable, which is kind of like a fancy

1:04:54Google Sheet, which you'll see it's a

1:04:55great tool to learn. That's another key

1:04:57one that you're going to be learning in

1:04:58this, is how to set up airtable and

1:04:59connect it to n8n. And that's going to

1:05:01be serving as a CRM or our customer

1:05:02relations manager, which is a basically

1:05:04a database of all the people that we're

1:05:06contacting or working with in our

1:05:07business. And then I've said I would

1:05:09like this to be as lean as possible and

1:05:10simple as possible using a single agent

1:05:12and tools. So, in this case, I know what

1:05:14I want, so I'm telling it quite exactly

1:05:15what I'm looking for, the different

1:05:17services I'm going to use. But in many

1:05:18of your cases, you may not know, and

1:05:19you're going to use this AI automation

1:05:21CTO to be able to help find those kind

1:05:23of geocoding or solar APIs, getting it

1:05:25to search the web and figure things out.

1:05:27So, here we have the research that's

1:05:28been completed, what the APIs are, how

1:05:29they're going to work, the shape, the

1:05:31endpoints, and practical notes. Talking

1:05:32about the architecture of the n8n agent,

1:05:34chat trigger, AI agent, solar estimate

1:05:36tool, and the create lead tool. Yep,

1:05:38that's correct. And an optional uh

1:05:40knowledge base tool via a vector

1:05:41database. Then it's breaking down the

1:05:42second workflow, which is going to be

1:05:44packaged into the solar estimate tool.

1:05:46So, we're going to be using a separate

1:05:47workflow as a tool, but there's a little

1:05:49trick you can do to kind of keep it

1:05:50within the same uh workflow, which is

1:05:52great. So, this is going to be a

1:05:53separate sequence of steps that's a bit

1:05:55bigger, which is going to chop that

1:05:56solar information retrieval from the

1:05:58Solar API into a couple steps. Here we

1:06:00have breaking down kind of the prompt or

1:06:01the glue that's going to stick it all

1:06:02together and how the behavior we expect

1:06:04the behavior to go. So, I'm going to

1:06:06tell it to assume that this is USA only

1:06:07because that's where the Google Solar

1:06:09API is relevant. And don't worry about

1:06:10handling all the edge cases cuz it's

1:06:12starting to build into here uh ways to

1:06:13handle if something goes wrong or if uh

1:06:16there are errors. We don't necessarily

1:06:17need to do that in this version, but it

1:06:18is something that you would do later on.

1:06:20So, I'm asking it to just create an MVP

1:06:22or minimum viable product here so we can

1:06:24test the core functionality. And then I

1:06:25will say,

1:06:30And I've told it to carry over a lot of

1:06:31this research about the APIs into the

1:06:34information that I'm going to put into

1:06:35the form here on the left.

1:06:37Okay, now we have the brief here. If for

1:06:39the model preference, because you guys

1:06:40have already set up your Google account,

1:06:41I'm just going to keep using Google in

1:06:42this case. Um so, we can just say,

1:06:49And we're going to give this a run.

1:06:55Come down and get this raw. We're going

1:06:57to head over to n8n, create a new

1:06:58workflow.

1:07:01Open the AI editor here, paste it in,

1:07:03give it a quick check over. It's got all

1:07:05the details about the Solar API here,

1:07:07that's really important. The output

1:07:08schema, and I think we're good to go.

1:07:23There we go. This is the first attempt

1:07:24it's given us. Uh let's just have a skim

1:07:26through and see what it's asking us to

1:07:27set up. So, the airtable uh lead tool,

1:07:30that makes sense. [music] It's going to

1:07:31need us to set up our airtable

1:07:33integration there. Uh the Google API key

1:07:35and the Google API key. So, just a look

1:07:37through here, give a bit of an

1:07:38explainer. We have the chat trigger

1:07:39here, and if we pop this open here, we

1:07:41can look at the chat trigger. So, this

1:07:42allows us to, in this case, chat away

1:07:45here in this chat session, and this is

1:07:46going to allow us to test the

1:07:47functionality before it's actually put

1:07:49onto the website, so this is our testing

1:07:50window. And this log section will show

1:07:52us what's happening to the AI agent at

1:07:54every step as we go through the chat. We

1:07:55don't need that right away. And once we

1:07:57receive a message, this chat trigger

1:07:58will fire information into the solar

1:08:00lead agent here. We have a handy-dandy

1:08:02prompt. You're a solar energy

1:08:03consultant. Your role is to answer

1:08:05questions. Answer questions in the

1:08:06knowledge base. Well, it seems like it's

1:08:08just put the knowledge base directly in

1:08:09here, um which is is one option. I'm

1:08:11actually going to take that out and give

1:08:12it a proper knowledge base. Important

1:08:14rules.

1:08:15After providing an estimate, that's what

1:08:16we want it to do. So, yeah, that looks

1:08:18pretty good overall. We have the Google

1:08:19Gemini 2.5 flash. Again, it's got the

1:08:21wrong model here, so we'll just change

1:08:23it over to the one we want. We've got

1:08:24our original Google connection here from

1:08:26the previous builds, you should just be

1:08:27able to reselect that.

1:08:30And we have the solar estimate tool

1:08:31here. Now, this is actually not the uh

1:08:36easiest way of doing things because we

1:08:38have a uh just a second workflow here.

1:08:40So, we basically want to connect this

1:08:41whole workflow via this into this. So,

1:08:44normally you'd have to create a separate

1:08:45workflow, you'd go back up to add and

1:08:47create a new one, and then we would we

1:08:48would uh click here, we would add a new

1:08:50one here, and go call n8n workflow tool.

1:08:53So, this is how you can build much more

1:08:54complex tools with multiple stages like

1:08:56this from an agent. So, if I click this,

1:08:58it's basically asking me which workflow

1:09:00am I supposed to call uh with this tool?

1:09:02So, a little trick you can do here is

1:09:03that you can actually set it up by from

1:09:05ID.

1:09:06And because we have this entry point

1:09:07here, we can actually trigger this as

1:09:09the entry point by taking the URL up

1:09:12here and grabbing this section of our

1:09:13URL after the workflow bit, and then go

1:09:15back to this workflow uh n8n [music]

1:09:18workflow tool here, paste in this as the

1:09:19ID, and there we go. It's it's been able

1:09:21to read this and determine what inputs

1:09:23cuz we need to pass some information

1:09:25into that workflow. And the AI has

1:09:27already set up the address as a in the

1:09:28monthly bill case we're looking for

1:09:30actually a number. So, this is all set

1:09:31up correctly as the inputs, and that's

1:09:33why it is automatically popped up here.

1:09:35So, just one more time, this whole

1:09:36workflow that we're working in is this

1:09:39workflow ID up here that we clicked.

1:09:40But, because on this whole canvas, the

1:09:42only option for triggering this workflow

1:09:44via an ID is this one, when we put in

1:09:47the ID of this workflow into this tool,

1:09:50it's going to call the only one that's

1:09:51available. So, it can't enter here

1:09:52because this is not a trigger by another

1:09:55workflow. We're going to take the ID of

1:09:56this whole workflow, pass it in here,

1:09:58and we can sneakily keep this on the

1:10:00same canvas, nice and easy like that.

1:10:02So, we can delete this.

1:10:03Then we can check out how it set the

1:10:05solar API section up. So, we have the

1:10:07information of the monthly bill and the

1:10:09address being passed into this geocode

1:10:11address. So, we've got it using the

1:10:12correct URL that we found in the

1:10:14research section. With indication we

1:10:16have none here for now, and we're just

1:10:17going to be putting in our API key in

1:10:19the query parameters, which will go in

1:10:21the URL that we send off to Google. So,

1:10:22that all looks pretty good. We'll grab

1:10:24our API key in a second. Then we take

1:10:26the solar data, which is going to be

1:10:27grabbing the information from the

1:10:28previous step, taking that latitude and

1:10:30longitude that the geocoding API found,

1:10:32and setting it into the

1:10:34>> [music]

1:10:34>> input to the solar API as the latitude

1:10:37and the longitude here, and our API key

1:10:39as well. Sending that off, then it's

1:10:41going to dump back a ton of information

1:10:43to us. So, we need to pick through that

1:10:44and determine what we actually want. In

1:10:46this case, we want to

1:10:48expect us to have a little bit of issue

1:10:49with this because it's sending back so

1:10:50much data. So, I won't go into this too

1:10:52much for now, but it's basically can we

1:10:53extract only the stuff that's important,

1:10:55send that back to the agent, and the

1:10:57agent can turn those numbers into a

1:10:58nice, clean response to give to the

1:11:00user. We have got a conversation memory

1:11:01as usual here that's handy to have that

1:11:03set up. We're remembering the past 10

1:11:04messages. And one thing we are missing

1:11:06is our knowledge base. So, I'm going to

1:11:08set that up later, but for now we can

1:11:09just pop it in here as a placeholder,

1:11:11simple vector store down here.

1:11:14And we are just going to leave that for

1:11:15now. All right. So, the most important

1:11:17thing for us to do now is to make sure

1:11:18that the solar estimate is working. So,

1:11:20we can do this a few ways. The easiest

1:11:22way to do that is to come in here. We

1:11:23need to change these inputs to being

1:11:26created by the AI. So, that's an

1:11:27important button to know exists that

1:11:30when we have certain fields for say for

1:11:32these tools that we're using, we can

1:11:34either hardcode the inputs, so we can

1:11:36have it set as fixed, and we can always

1:11:39pass in the same value. We can have it

1:11:40as an expression from variables earlier

1:11:42in the workflow. Or, in this case, as a

1:11:44tool for an AI agent, we can let the

1:11:45model define this parameter. And what's

1:11:47helpful is to add in a description of

1:11:49the tool. We can say this is

1:11:51>> [music]

1:11:51>> the address of the user. And this goes

1:11:54back to the descriptions that we write

1:11:56for tools. The descriptions of the

1:11:58variables or the parameters that you

1:11:59pass into the tool is also helpful to

1:12:01give the model more context on what is

1:12:03supposed to go where and the purpose of

1:12:04each of these. I'm adding must be as a

1:12:06number here,

1:12:07giving an example of the format. It's

1:12:09just some more context around how it

1:12:10should be structuring each of those when

1:12:11it passes it in.

1:12:14So, if we go stop and we execute the

1:12:17step, now it pops up with things that we

1:12:19can put in. Here we can see this is as a

1:12:21number. Remember in this button down

1:12:22here, it says it's a number, and the

1:12:24address is going to be a string. So, I

1:12:25like to go random USA address. [music]

1:12:28Let's just go maps and find some random

1:12:30guy's house. Let's go Utah. Where is a

1:12:33city?

1:12:35This one looks good. Actually, yeah,

1:12:37let's go a few panels. So, we can see

1:12:39these different sides to the roof.

1:12:40Google's actually going to be able to

1:12:41figure that out, which is pretty

1:12:42helpful. So, we're going to copy this

1:12:43address. Sorry to this person.

1:12:47Monthly bill of $200. Actually, before

1:12:49we can test this, we should get our API

1:12:50key set up from Google. So, let's just

1:12:52back out of here, head to

1:12:54Google Cloud Platform that we've been

1:12:57using previously. All right. So, now

1:12:58you've got that project that you created

1:12:59before that you've already got the

1:13:00billing set up for. We want to go down

1:13:02to APIs and services. You want to enable

1:13:04APIs and services up here. And we're

1:13:05going to look for the geocoding API.

1:13:09You can click this.

1:13:11You can click enable here. Then we're

1:13:12going to go back and go solar.

1:13:16And we're going to

1:13:18And we're going to enable this. Then we

1:13:20can go into manage. We can go to

1:13:21credentials.

1:13:23And then you can create a new API key.

1:13:26We can copy this, hit back, open up this

1:13:29geocoding one.

1:13:31Paste the key in there.

1:13:32Then go to the solar one. Same again.

1:13:36And now, while we're here, we may as

1:13:37well point out and link back to some of

1:13:38the concepts we learned earlier. We have

1:13:40this as a get request, and get requests

1:13:42don't have a request body. We don't send

1:13:44away a big payload of information. What

1:13:46this is doing, it's just essentially

1:13:47adding on little things to the end of

1:13:49the URL. So, a get request will take

1:13:51this, and then it's going to be

1:13:53attacking on the latitude and longitude

1:13:55to the end of that. They're called get

1:13:56parameters or query parameters, and

1:13:58they're going to be added onto the end

1:13:59of the URL, and that includes our API in

1:14:01this case, which is not like the most

1:14:03safe way to do it, but for these

1:14:05purposes it will be fine. Now, we can

1:14:07try to save this. Okay. So, I've just

1:14:08cleaned this up a bit so that we can run

1:14:09through some testing loops as we work to

1:14:12getting this custom workflow working.

1:14:14Now, if we right click on here and go

1:14:15execute step,

1:14:17sorry, if we open it up,

1:14:18we change these to AI.

1:14:22We can save that, and then right click

1:14:24on execute step. We can put in the

1:14:26address and put in like $200 monthly

1:14:28bill. We execute this, and then the

1:14:29error of any errors that return in here

1:14:32are actually going to be hidden in here.

1:14:33So, you'd expect to see dick, dick,

1:14:34dick, but actually you need to click

1:14:36into here, and you can see that the

1:14:37resource you're requesting could not be

1:14:39found. If you click on view sub

1:14:40execution, we get to go on a bit more of

1:14:42a kind of debugging mode. You can see we

1:14:44have the editor tab here, the executions

1:14:46and evaluations. Our editor and [music]

1:14:48executions are the main ones you're

1:14:49going to need to worry about. So, in