Let's build an AI agent that is

optimized for handling PDF data. I'm

going to be taking this dummy invoice

data here, but the steps and logic

you're about to see, you can apply to

any type of PDF data that you want to

input into this workflow. So therefore,

by the end of this video, you're not

only going to learn how to extract data

effectively from PDFs, but you're also

going to learn how we can take that data

and leverage it in other applications.

This specific video, I'm going to show

you how to do with Google Sheets, but as

you already know, you can apply that

logic anywhere else. Let's jump in.

Welcome, Becky. Today's video comes from

a suggestion from our community here

down in the description down below.

Completely free to join. Sarica

basically asks, "How do we take it so we

can get a PDF here and automatically

give it to an agent, get the data,

extract it, and use it?" So, we're going

to show that real quick. So, to do that,

go ahead and start your new agent

workflow. Start stays the same, but the

first little agent that we're going to

build together here is going to extract

the underlying data. Now, I'll be honest

with y'all. At first, I was like, "Oh,

are we going to have to use assistance

API here to get vision context? are we

going to maybe add an extra layer here

in order to see the data that's actually

on the PDF? But in reality, we don't

have to do any of that. What's cool

about these agents is that builtin is

vision context. And if you don't know

what vision context is, essentially when

dealing with a PDF, this is obviously a

textheavy PDF, but some PDFs have maybe

more images or diagrams, whether that's

for real estate or engineering.

Therefore, vision context is important

to understand those diagrams and for the

information we want to get out of it. In

theory, when looking at an invoice PDF

like this, which is very much textbased,

we could programmatically create a

Python script to extract all the

relevant data. But let me just show you

a one-size fit all to get this data. And

the one size fit all is going to be

using this vision context that's built

into agents. So, first thing you want to

do is provide context of what the PDF

is. Mine's an invoice. Look at this

invoice PDF. What the heck is yours? I

don't know. Is it like the most amazing

dog treats and a product description of

dog treats? whatever it is. First line,

context, PDF. Second line, we're going

to identify what information we want to

extract. So, for me, in today's video,

we'll extract the company address, the

total, and the invoice number. But don't

worry, we're going to gut check this to

make sure that the real data is getting

pulled in. It's not hallucinating. So,

this is the address, this is the invoice

number, and then the total is $262.50

USD. That's it. So, what is the relevant

data that you care? This can be much

longer. If anything, we could format

this a little bit better so it's easier

to read, more legible. Add the data

points that you care about that's coming

from that relevant PDF. In addition, you

can also extract different things such

as summary, more open-ended data points.

When I say open-ended data points, I'm

essentially saying maybe more

analysisoriented data points. So, for

example, in the invoice PDF, maybe just

like a give me a two-sided summary of

the client and the customer. It doesn't

really apply too well to an invoice

example, but maybe for that real estate

example, you'd want to get information

like, give me a summary of based on this

location, is this a good purchase, based

on square footage and location

desiraability. Keep that in mind. So,

for me, in theory, I could do summary of

the invoice. Not really relevant. So,

right now, we're just going to extract

three data points that you can expand to

however many data points you care about.

It's safe. With this done, we can scroll

down here. We got the model GBT5. That's

fine. If you find yourself running into

issues or errors or it's not working as

effectively as it should, simply come

over here to reasoning effort and then

increase it to high since the purpose

for me simply is invoice invoice data

and three specific data points.

Honestly, I can leave this as low. Fast

in, fast out. High gives the model just

more IQ. So for your task, does it

require more IQ? Tools. We're going to

go ahead and give no tools here. Don't

worry, it's actually built into the

agent. And then the response format will

be JSON. This is fundamentally important

for the next step here for it to be a

JSON output. All right. So here we go.

This is going to be the structure JSON

output. I can rename this. So maybe we

do like invoice schema voice data to

make it more sense here. Add property.

And the property name is going to be

what we identified in the prompts. So

for one was company. We'll just do

company address. The other one was

invoice

number. And then this one was a number

because it's a number. I'm going to hit

update. And then finally total USD. But

I'm going just do total USD and we can

make this a number as well.

Fundamentally, if you want to deep dive

on what all these mean, I suggest you

just do a screenshot, put it into an AI

chat real quick. String just text number

full, true or false. Enum categorization

object a little bit more complex. Array

is going to be like a list. All you need

to care about for right now because most

of the times you're extracting data is

either going to be a string or a num.

When I say num, I mean number update.

With that done, then the next step here

becomes fundamentally way easier. So, we

got our data being formatted from the

PDF. In the user data section here,

we're going to leverage Zapier. And what

Zapier is going to allow us to do is to

then take it to our Google sheet here.

And with our Google sheet, we should see

the data come in as the second row here

automatically. So, we're going to go

ahead and first create a prompt. Place

this data in the relevant Google sheet

column. Identify where you're placing

the data in the specific software this

data is being placed into. In theory, if

you're not really placing it anywhere,

maybe you're just sending it to a Gmail,

just say that as well. Whatever the use

case for the data is, identify it here

as the first line. Next, we're going to

identify the data again. So, what I'm

going to do to make my life easy is make

sure you use the exact same dictation

that you used in your actual wherever

you're placing the data. So, invoice

number, copy, invoice number, semicolon,

company address, semicolon, total USD,

semicolon. Now, here is where we add

context. Context is going to be the

invoice number found here, invoice

number. So, it wants to place down

there. I don't like that. Don't do that

to me. Place it right here. And then,

what you'll notice here is that we can

add the other ones. Input output parsed

invoice. No, no, not the invoice number.

This will be the company address

address. If you're wondering, Corbin,

how'd you know it was that? Because if

you go back here, go to the invoice

data, we called it company address.

Nice. Coming back over here though,

assembly, open this little up total USD

context here. Total USD. There we go.

So, we got all three data points being

placed. Place this data in the relevant

Google sheet column. There we go. If you

have more data points, add the more data

points. Hit save. Now, this one because

it's a little bit more complex. I'm

going to do to high. In theory, I could

go medium. Test it. See if it works. But

especially in the beginning, y'all

always opt for high. Just get it

working. Once it's working, then you can

kind of play around reasoning effort.

Next, we're going to do a tool of

Zapier. So, first we're going to do add

MCP Zapier. Get your API key. For me,

I've already created one. So, simply go

to connect here. Copy secret. It's a

secret. Zap year. Enter it here. If

you're wondering what that is, why you

even need to do that, essentially, this

is tells OpenAI that you have access to

an actual Zap year account and the

functionalities you're about to see.

We're going to add these tools together.

Don't worry. For now, I'm going to

uncheck these. And the tool tools we're

going to add together is creating a

spreadsheet row. The ability to actually

functionally create a spreadsheet row

within software. And then on top of

that, get data from that spreadsheet at

first. Add. So now coming over to Zapier

MCP, let's make sure we add those tools.

I have already added them. You know what

I'll do? I'll remove them from the

server and we'll add them together. So

we're going to add tool Google sheet. In

theory, we could add all the tools, but

I honestly suggest you not to do that.

First, let's just do lookup spreadsheet

rows. This is going to give the ability

for the artificial intelligence model to

get all the relevant data found in the

spreadsheet. We're going to do configure

here and then within configure, we're

going to make sure we choose the correct

spreadsheet. So, I'm going say set

specific value for this field. This is

so that we make optimized decisions and

it doesn't get confused. For me, I

called it easy data. Why? Because it's

easy. So, therefore, I'm going to go to

here, easy data worksheet. If you have

multiple worksheets, choose one. I only

have one, so I'm going to hit save. Add

another tool. Now, let's give the

functionality for it to actually create

a row. To do this, we can type in create

up here. Find data, take action. What's

nice is that you can see all the things

it can actually do functionally to that

spreadsheet. So, we're going to create a

spreadsheet row. But I want you to

notice a couple things. First thing,

create multiple spreadsheet rows. What

does this mean, Corin? This means in

theory, we could build out an agent here

that essentially if I provided like 10

invoice PDFs, it would be able to loop

through them and create multiple

spreadsheet rows. The logic's a little

bit more complex. So, obviously just do

it with one at a time at first just to

get it working. I'm going to come over

here to configure. In configure, we're

going to select that spreadsheet again,

which is going to be easy data.

Worksheets's going to be the same. It's

fine. It's all good. We're going to hit

save. Now, notice two things. First

thing, notice that we've identified

specifically the actions of create

spreadsheet row, look up spreadsheet

rows. We can give more actions here. And

you know, MCPS can give actions across

all these different apps over 8,000. And

then on top of that, we've identified

the specific area we want to place that

data, which for me was easy data. And

that's what we called our Google sheet

here. Nice. And with this done, because

I essentially readded those, let me see

if the tools are selected here

correctly. So yeah, let me do that

again. Update. Nice. And here we go. We

got our prompt. We got our MCP. We get

the reasoning effort to high. Let's see

if this works. Go to preview. I'm going

to add my invoice data PDF that I showed

earlier. This one right here. So we got

our invoice data PDF. Enter. First thing

it's going to do is extract the relevant

data that we care about, which was the

address, invoice number, and the total

USD. We identified this. And once it

extracts that, it's going to put it into

a JSON output. Don't over complicate it

at all. JSON output is just a way we can

format data. So it effectively can be

run thousands upon thousands of times in

a structured manner. This is how the AI

likes to talk to each other. Okay? Or

software in general as well. Here we go.

Company address invoice 7262.5.

Now right now in the workflow, it

requires me to approve the action using

Zapier. I'll show you how to make it so

that it always just does it

automatically because that can get

frustrating. But we should be

essentially executing right now. And

there we go. Invoice number, the company

address, and then 262.5 coming over

here. 262.5. Nice. Now, some of y'all

might be like, Corbin, where's all the

zeros? Like 0000. Okay, we could format

it that way if we want to. Or

alternatively, the total USD, why is it

formatted that way? We'll just here to

format as currency. Okay, but that

worked perfectly and that executed how

we like. So, in theory, I can close the

preview, come over here, go to Zampier,

simply click approval to never require

approval. So, it just essentially just

does it. You don't need a yes. There we

go. Make sure to leave a like. It is

completely free. Check out that school

community in the description down below.

Let me know what other use cases you

want to see using agent builders. One

big thing I already know half of y'all

are going to be like, "Corbin, this is

cool, but and that but is essentially

you saying, "How do we actually put this

into an internal app for my team, push

this to an actual website?" Basically,

how do we use that chatkit UI? Well,

lucky for you, I've already created an

entire video dedicated to how to take

workflows like this and actually

implement them to real websites or

internal apps for your company. So,

without further ado, I'll see you in the

next Did we just learn how we could take

basically any PDF, extract the data

automatically, and put it anywhere

video? Nice.