198 | Build a Business-Boosting AI Agent That Books Calls While You Sleep-No Code w/ Carolina Posma - Leveraging AI

Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Isar Meitis (00:00):
Hello, and welcome to another live episode of the

(00:03):
Leveraging AI Podcast, a podcastthat shares practical, ethical
ways to improve efficiency, growyour business, and advance your
career.
This is Isar Mateis, your host,and we have an amazing session
for you today.
We've been talking a lot aboutagents, both in previous
episodes as well as on the news,like agents are now, everybody's
interested in agents, everybodywants to build agents.
Uh, everybody wants to know whatagents can do for them.
And the reality is most peopledon't have a clue how to do

(00:26):
that, how to get started, whichplatforms to use, and so on.
And this is exactly what we'regoing to tackle today and we're
gonna demystify the wholeconcept of creating agents.
We're gonna show you exactly howto build an agent that you can
actually use that will be veryuseful for probably most
companies in the world today,big or small.
So what are we going to build?
We're going to build an agentthat will learn everything about
your company just by looking atyour website, and we'll be able

(00:47):
to answer questions on thatwebsite to, well, wherever you
wanna place it for a demo today,we're gonna place that on
Instagram.
So we're gonna build an agentthat will be able to answer DMS
on Instagram.
Based on information from thewebsite, you obviously can build
it and connect it to any otherthing like place it on your own
website and so on.
And the cool thing is you'll seethat the Egen universe, that
while this sounds complicated,it actually really easy and will

(01:07):
do all of that within the next30 to 40 minutes, which is
incredible when you, again, mostpeople think that it's really
hard and really complicated, andyou gotta have an army of coders
to do this.
The reality is it's not, andwe're gonna do this with the
help of Carolina psma, who isour guest today, she is an
expert on doing this.
She has her own agency helpingcompanies do exactly that.
She is a.

(01:27):
Advisor, like an external, uh,promoter of Flo Gent, which is
one of the most commonly usedplatforms today to develop
agents.
And this is the platform we'regoing to use today to show that
to you.
Her background is in marketingand in entrepreneurship.
She's been in multiple marketingroles across multiple, uh,
companies, including several ofher own.
So she understands business, sheunderstands entrepreneurship,
and she understands agents,which makes her the perfect

(01:49):
guest to share with us, uh, herinformation about this.
So I'm really excited about thisone.
Carolina, welcome to leveragingai.

Carolina Posma (02:39):
Thank you so much, ISAR, and thank you for
this, this very kindintroduction.
So what I want to show to youtoday is indeed how you can
build an easy and Instagram aiDM appointment set.
So you already mentioned it inyour description, ISAR, uh, an
AI agent that does sales for youon pil, right?
So that's what we're going toexplore today.
So indeed, uh, pick about me.

(03:00):
Uh, oh, sorry.
Yes.

No, I'm saying that's, that's amazing.
I think it's very exciting.
I just wanna, first of all,thank everybody in the audience,
the people are joining us live,whether on LinkedIn and or on,
uh, zoom.
So thank you so much for beinghere.
If you have, uh, first of all,go ahead and introduce yourself.
So where you are, what yourexperience with agents, so we
have a little idea of who's inthe audience.
And, uh, if you're not in theaudience, then know that we do
this every single Thursday atnoon pm Eastern.

(03:23):
You can come and join us, andthen you can ask questions, you
can chat with the people in thechat.
You can network and so on.
Also, I wanna mention one lastthing.
If you wanna learn more in depthon how to leverage AI in your
business.
There are two ways we can helpyou with that.
One is we have the AI BusinessTransformation course, which we
have been teaching for over twoyears.
The next session opens on thesecond week of August.
It's a four week course, twohours a week.

(03:45):
That will give you an incrediblelevel of understanding in AI and
how to apply it in yourbusiness.
Whether you're just a businessperson or you're in leadership
position in a company.
Uh, in both cases, it will giveyou incredible.
Boost to your understanding ofai.
The other thing is we work withcompanies on custom tailored
workshops to help businessesaccelerate very quickly into the
AI era.
We do this either in person oronline, depending on the setup

(04:08):
of the company.
So if you're interested in anyof those, just look for the
links in the show notes, uh, orjust connect with me on LinkedIn
and ask me whatever you wannaask me.
Uh, and now let's dive intoagents and let Carolina lead the
conversation.
Uh, Carolina, it's your stagefrom now.

Yes.
Thank you so much.
So let's talk a little bit firstabout the difference between
chat GPT, AI automation and AIagents.
If you are a frequent listener,uh, you might already, uh, know
the difference, but I alwayslike to, uh, go back to basics
to just, uh, get everyone on thesame page.
So I will share a couple ofslides first, where we will
discuss this.
So let'S get started.

(04:42):
First a little bit on thedifference between chatt AI
automations and AI agents,because today we're going to
talk about AI agents and we'regoing to create our own AI
agents as well, but.
As I said before, it's alwaysvery useful to get a little bit
back to basics on what are weactually talking about.
So first, where do we stand Nowin the world of ai, many people
have already probably seen this,uh, lead note from the Shopify

(05:04):
CEO telling us that, um, uh,telling all his employees that
before asking for more headcountand resources, teams must
demonstrate why they cannot getwhat they want done using ai.
So this already emphasizes theimportance that huge companies
lay on understanding AI andusing AI and knowing how to get
things done using autonomous AIagents, as it's mentioned over

(05:25):
here.
Then we have companies such asGumbo that are only with four
persons that are raising 17million, uh, in dollars.
So they're building, as theyclaim, the first 10 mil, uh, 10
billion, um, uh, 10 per 10person, 1 billion company.
We have companies looking for AIfirst talent only, and you have
to automate your job as you go.
And even further, there arecompanies like Fire Crawl who

(05:47):
are currently hiring AI agents.
So they're not even, uh, lookingfor people to build AI agents.
No.
They have job agencies for AIagents open.
And the great news here is thatwe are still in the very much
beginning and the people whoadopt this technology right now
will win just like in a previousstance like internet,
smartphone, those are verydisruptive, just like AI are.

(06:07):
And.
When we're talking about AIagents, the first thing that you
need to grasp a little bit isChat g pt.
But I am assuming that everyoneright here has played around
with chat.
GT already has had aconversation already with chat,
GPT, uh, but a lot of peopleare, uh, are calling chat GT and
AI assistant.
And this is where I want to getstarted because in my opinion,
AI uh, chat g PT is just an AIadvisor because chat GT is

(06:30):
brilliant to get advice tobrainstorm ideas, to create
content that will help you whenyou're stuck in anything.
Chat GT is your personal advicethat you can ask anything, but
one of the problems of chat G PTis it can't really do anything.
Yes, if you're very advancedwith Che gt, you may be aware of
custom Gcpt, but in general, CheGT doesn't do, uh, anything.

(06:51):
Then we get on to automations.
And automations was alreadypossible for a very long time.
Uh, let's take this, uh, superstraightforward automation
example where you're connectingtwo different tools.
Let's say you have a, uh, leadform on your website and people
can leave their, uh, informationover their first name, last
name, and you want to have thatadded to, for example, your CRM
such as HubSpots pipe drive forany other CRM.

(07:14):
You can do that using tools likeZapier make and it m uh, tools
that were there for yearsalready.
So putting this information fromone side to the other side was
already possible for years,years, years.
But why is there all the hypessuddenly about AI automations or
AI automations?
So process automations as weknow them, completely changed

(07:36):
with the introduction of bettercost friendly and accessible ai.
2022, the, uh, A-P-A-P-I ofGPT-3 0.5 became available and
suddenly.
AI wasn't anymore just for thebig corporates with huge budget.
No, it was for all of us, foryou, me, everyone who, uh, is,
um, yeah, open to play aroundwith ai, even as an individual.
And what we were seeing was thatAI was getting so much more

(07:57):
smarter and at the same time, somuch more cheaper.
So it became more, more, moreavailable.
And what does AI now allows usto do, to do in these process
automations?
Well, it allows us to interpretinformation, to summarize
information, to turnunstructured information into
structured data.
So what do I mean by that?
Let's say for example, that youhave, uh, uh, someone sending in

(08:17):
a resume or a cfi, it can besent in a PDF and that's a huge
bulk of unstructuredinformation.
It can, the, the name can beanywhere on that PDF, right?
But with AI, it can read thatinformation and it can.
Take out, for example, oh, thisis the first name.
This is the last name.
It can turn unstructuredinformation into structured data
that's so powerful.
It can create new content.
So there's so much that you cando with AI right now.

(08:40):
And when we take a look at ourexample, and please interrupt me
if you have any questions, butif we are looking at this
example from, uh, the lead formto, uh, name, email, website to
your HubSpot pipeline, suddenlyyou could start doing stuff with
that information.
So with the introduction of ai,you can can summarize the
website, for example, you canum, give some qualification
criteria to AI and ju have ajudge if this is a qualified

(09:03):
lead or perhaps you have a salesteam that's pick up the phone
for everyone who is, uh, who isfilling in that lead form and.
AI or an AI can help you createa call script for, for that
salesperson, right?
So there's so much that you cando with ai.
You can interpret information,unstructured information into
structured data.
It's changed the rules.

(09:24):
So right now we discussed chat,GPT, which is your personal
advisor, to ask questions, gethelp.
Then we have automations.
And the old school automationshelp you to stop copy pasting
info and doing the same boringtasks.
It helps you to connectdifferent softwares and tools
with each other.
And then since a couple ofyears, we have AI automations
and that's making yourautomation smarter with ai,

(09:45):
summarizing it, interpreting it.
But something new in the lastcouple of months, year, uh,
happened and that that is AIagents, and this is even a step
further.
So let's move on to the AIagents.
You can think of AI agents asJet GPT, but it doesn't just
chat, it takes action.
So you can train an AI agent,for example, in your company's

(10:07):
knowledge, and you can plug itinto different tools.
It can book meetings, it canupdate your end records.
It uh, records.
It can send messages.
It gets real work done.
Before we dive into AI agentsand how to build, uh, it's
without, if you don't know, haveany technology at all because,
uh, like is, I said, you don'tneed to, uh, to be able to code.
No, you can really, there aresoftwares that allow you to just

(10:29):
look and play with AI agents.
There's something that you firstneed to know it.
Where we stand with AI agentsright now, and what I'm showing
you on the screen right now isthe triangle of talents.
And, uh, in this triangle oftalents, you see a triangle.
It looks a little bit like thesocial hi hierarchy of different
levels that employees can have.
And level one is lowest level.
That's useless.

(10:49):
And that's someone, if you tellthem what to do, they rarely do
it right.
If you go one level up, we'retalking about task conies for
those, uh, employees, personsshould tell them what to do and
how to do it and when to do it,and it'll get done.
Then we have a third level oftalent, which is problem solver.
Tell them what to do, and theyfigure out the how.
The fourth level of, uh, talentis systems thinker.

(11:11):
Tell them.
Um, and I do see that thecredits actually are missing.
This is, uh, shared by shampooon, uh, on, uh, Twitter.
The fourth level is systemsthinker.
Um, tell them the problem.
They set a, a, a system tofigure it out.
People plus process and levelfive is a superstar.
They identify the right problemand get it solved.
So where are AI agents?
I wanna pause

you.
I wanna pause you just for onesecond to add two things, one to
do a quick summary of the greatintro you did of the differences
between AI and agents andautomations, and I think most
people are still struggling withthat, so I think that was great.
The biggest differences betweenagents and just ai, large
language models is, like I said,they can take action.
The second thing is they canfigure out their own steps.
Like if you wanna do amulti-step process with, uh, ai,

(11:52):
large language model, you gottatell it what the steps that it
needs to do.
Even though the reasoning modelssolved it partially, but only
partially, uh, and the agentsknow kind of how to think for
themselves.
You give them the goal, theyfigure out the steps, versus you
have to take them the steps.
And number three is they knowhow to work collaboratively with
other agents.
You can build a hierarchy ofagents that can talk to one
another.
When it's LLM, it's just thatone LLM.
So this is kinda like a quicksummary and my 2 cents to add to

(12:13):
that on what you just said aboutthe, the talent, there was a
very interesting, uh, interviewwith Sam Altman this past week
at the, um, keynote ofSnowflake, the big data company.
And when he was asked about whathe's excited about, that's
probably coming in the nextyear.
So the next 12 months, hebasically thinks that the
biggest thing is that AI will beable to solve really big, really

(12:35):
complex business problems thatwe are either unable to, or it
takes us a huge amount of effortto solve just by giving them
more compute.
So he's basically saying whatyou're talking about right now,
that we are gonna go from leveltwo to level four, maybe five,
uh, by just giving it access tothe right data and just paying
for a lot more tokens.
But then it will know what todo.
And for, if you can solve areally big problem for a really

(12:56):
big company that might be worthtens of millions, hundreds of
millions, billions, depending onhow big the company is.
So saying, okay, I'm gonna throw$2 million worth of compute at
this.
Makes perfect sense in thosescenarios.
And so just to give you abroader idea of where we are,
where we are going, uh, this iskind of where we are now.
I'll let you continue on wherewe, uh, where we are right now.

That's amazing.
That's, I, I don't know.
It's, it's, it's sometimes, soit is honestly also a little bit
scary because even AI agentalready can go to level four,
the systems think go level five,a superstar route.
That's, I, I think that's,that's brilliant when we get
there.
But also, I'm so curious on howthe world will look like by then
on, on, on what will change.
Because so much is going tochange, but at the moment, um, I

(13:37):
think, um, right now, uh, AIagents are still, most of them
are still task monkeys.
I do know that, uh, somecompanies are playing around
with problem solvers already,but, uh, the most AI agents
that, uh, you for example, wouldcreate, uh, in, in softwares
that you see, uh, often around,like for example, N eight N are
still task monkeys.
You need to tell them what todo, how to do it, and when to do
it.
And then your AI agents will getit done every single time, if

(14:00):
you instructed correct.
So with that in mind, um.
My experience, there are two biguse cases for AI agents right
now.
Um, AI agents can really wellconnect with your leads or
customers, uh, via you oftenhear voice, uh, so via calling,
via messages, for example, onInstagram like we're discussing

(14:21):
today, but also in website chatsor in, um, uh, on WhatsApp, for
example.
So, uh, AI agents, uh, are greatright now to really get the
conversation going withcustomers and leads like a sales
employee or like a customersupport employee.
That's the one use case in theother use case is a personal
company assistance, um, doingdifferent stuff in the
backgrounds, uh, more like anAutomation plus, for example,

(14:41):
doing some research for youevery time a lead is coming in
or, um, the, uh, let's say, uh,creating content for you doing
research on that.
So that is another big use case.
If you are not technical at all,um, or at least you consider
yourself to be a non-technicalperson, there are three software
tools that I can recommend youto get started with AI agents

(15:04):
and these software tools are foryour personal company.
Assisted are Zapier.
They, I think they have a reallygood super user and beginner
friendly, um, personal company,assistant AI agent builder, uh,
relay app.
I think, uh, I really like themin the sense that they have a
lot of pre-built, uh, AI agentsalready for you.
So you can just click and playalso for the personal company
assistant.
And if you are looking for an AIagent builder that allows you to

(15:26):
create AI agents that canconnect with your leads on, on
Instagram, WhatsApp, on yourwebsite, then I highly recommend
Flow Gen, uh, ai.
Right.
Also involved as an at Pfizer.
So let's take you look at someexamples that we can create, uh,
using, uh, an AI agent that canconnect with your leads or
customers.
So let's, like I mentioned, thelead generation on your website.

(15:46):
Uh, Instagram BM AI agent thatalso, for example, customer
support for e-commerce stores.
Uh, connect your AI agent withyour Shopify and have it search
for products and have it, uh,updates, uh, order status or
request order status when peopleare asking for it.
Or for example, an AI coachtrains on, on, on your course.
If you are, uh, a consultant orif you're running a course, uh,
imagine that people can chatwith your AI coach version, uh,

(16:08):
on WhatsApp or in Slack, forexample, and you, uh, giving
your students or clients 24 7advice.
Or let's say for example.
Uh, customer support or saleson, on WhatsApp for events.
I think if you're running a hugeconference, maybe you can start,
uh, reaching out to everyone onWhatsApp.
It's the new, uh, the newestupdates and, um, if people are,
have questions about where do Ipark or who are the keynote

(16:31):
speakers, or anything like that,and AI agents can answer that
easily right now.
So just to recap what I, um,showed you here and what I
discussed here.
We went from kechi pt, which isa personal advisor to
automations, to AI automations,and we did discussed two
different types of AI agents,the sales or customer support
employee or the personalassistance.

(16:54):
Any questions so far?

No, I think this is fantastic.
I don't see any questions in theaudience, people saying that
it's interesting and that it'sawesome, so I guess you're on
the right track.
Brilliant.

No, that's good to hear.
I see that NADA has a secondbrain assistant built on nat
com.
That's amazing.
Okay, let's move on then toactually creating our AI agent.
I am going to hide, let me seeif I can hide this.
Voting yes, because then I canactually see what I'm doing.

(17:25):
So if you want to, uh, walkthrough this, uh, together with
me, you are welcome to do itbecause, uh, well, I'm going to
build this live, always nervousfor the live demo, so I hope
everything goes well.
But, uh, I will show you how todo it.
So, um, where we're going tobuild our, what we're going to
build is an Instagram AI agentthat can answer any questions on
your Instagram dm.
So let's say that you have a lotof followers on Instagram, uh,

(17:46):
as a company or you're aninfluencer, then this is a
brilliant use case, uh, to tryand add an AI agent to it.
So, where we're going to buildthis is flow Gen ai.
This like flow and agentcombined, but then without the a
middle flow gen.
And on the flow website, you canclick on get started for free
and then you can add your emailaddress or click on continue

(18:07):
with Google.
And then you will get on a pagethat is the onboarding form.
Like every software, uh, there'san onboarding form and there you
can tell.
What should we call you?
Well tag Lina.
Um, we're going to build an AIagents for Multiply today.
Then, uh, we'll just keep thison one to 10 and we will
continue.

(18:27):
So next, uh, where did you findus?
Uh, referral.
What's the use case?
Uh, Instagram AI agents, but youcan put in any use cases will of
course help the team of flownto, uh.
Uh, builds their software betterfor you.
So, and this is where the magicis going to happen, because

(18:47):
what's going to happen right nowis, uh, if you fill in your
website over here, and we'regoing to do this with website,
if you fill in your website overhere in the background and ai,
uh, uh, or the website will bescraped and an AI agent will be
trained, so to say.
So, uh, it'll be added asknowledge and a prompt will be
generated, uh, on your website.
So you don't need to, uh, haveany tech skills for this.

(19:09):
You can just fill in yourwebsite and automatically, uh,
it'll be threats.
So we're building your agentnow, and in the same time, you
are going to be let, uh, guidedthrough the apps.
I'll just quickly go throughthat.
So, uh, the credits, uh,dashboards, um, if you are doing
this on your, uh, on your own,you can easily read it, but they
have a chat manager.
Uh, you can have differentagents.

(19:29):
You can, uh, connect withdifferent integrations and
settings.
Uh, now click on agents andclick on agents overview.
Then you can open it to see yourfirst agent, but the agent is at
the moment, uh, still beingbuilt.
Uh, this is your agent'splayground.
So once the agent is ready, wecan play around with the agent

(19:50):
and we can chat with it.
In the meantime, let's take alook.
You can edit to your websitehere to edit your agent, and we
will take a look at how to edityour agent.
So what we will see in, uh, acouple of minutes is the agents,
uh, the system prompts being,um, being filled in.
So.
For every agent, the systemprompt are the instructions that
you can explain to the agent onhow to behave.

(20:11):
So you can really think of itas, um, the introduction or the
onboarding guide for any of youremployees.
In your onboarding guide, youalso tell, um, how, uh, someone
should behave or, uh, whatsomeone should do, or, uh,
what's expected in what cases.
And that's the same, uh, uh, forthe system prompt.
So.
Remember that the system promptis really the heart and the
brains as well of your AI agentbecause you are, uh, telling,

(20:31):
its, its personality.
You're telling it, um, the wayto behave, which actions it you
take, um, and all in plainEnglish.
So this is really the same as aschatting with chat GPT, but
then, um, you're, yeah,explaining it to your AI agent
on how to behave.
The next part of every AI agentis that an AI agent should have
before, before you

continue just to talk about the system prompt.
So, and I think you framed itperfectly, right?
It's, it's the onboarding andthe training of a new employee,
right?
But it covers, I.
More than that.
Just like you said, because italso covers personality, it also
covers how you're gonna answerit.
Do you want to be brief?
Do you wanna be detailed?
If you are detailed, howdetailed do you want to be?
What's the format of youranswers?
Like every one of those thingsyou can customize and define for

(21:10):
the agent exactly how to behaveand different than the average
employee.
It's actually gonna follow theseinstructions, uh, exactly how
you tell them.
So if you, my question to youthough, Carolina, is, does flow
gent or do you have a differentresource of all the different
topics that should be in asystem prompt?
Like what, in order to make sureyou're not missing anything

(21:30):
important, do they guide youthrough these different things
or do you look for that onYouTube and start there?

Um, you will see it in a couple of minutes.
Okay.
Uh, but, um, Flo Gent right now,uh, because I filled in the
website, Flo Gent has the, um,system prompt to created
automatically with ai.
So they have a team of AI agentsin the background that go to
your website and create a systemprompt automatically.
So, uh, the rough draft of thesystem prompt is already there,
uh, which makes it more easy toget started with.

(21:58):
Uh, from there on, uh, you canmake any changes.
Um, things that I would includeand, uh, always, and this is
something that uh, yes, that I'mgoing to share, uh, now as well
is, um, always defined the roleof the, of the AI agents.
So who is he or she or it's, or,I don't know how to define it,

(22:18):
but, uh, who are they?
So is it, um, uh, is it.
A personal assistant, is it an,um, a sales employee?
Is it an, uh, appointmentsetter?
Is it a customer supportemployee?
This already sets the stagefirst for your AI agent on how
it should behave in, in, in therest of, of the conversation.
So the role then always defineindeed.

(22:40):
You mentioned it already aswell, the goal.
So what's the goal of the AIagent for a lead generation AI
agent?
You wanted, uh, maybe to collectemail addresses or for a
customer support AI agent, thegoal is to help the customer as
best as possible to solve itsproblems.
Then I always recommend toinclude, uh, the standard
operating procedure, but you canname it a different ways.
You can name it, for example,you can say it's the task or,

(23:02):
um, the standard operatingprocedure or the workflow or
anything like that.
But there you define usually on,um, bot steps.
It should take.
If you wanted to take certainsteps.
So for example, um, for ageneration AI agents, sometimes
you don't want it to, to havecertain steps and you just want
to go with the flow and reallyfollow the conversation.
But sometimes you want to AIagent to first ask a couple of

(23:22):
questions that are, for example,qualifying, and then, uh, go
with flow.
And that's all stuff that youcan also, uh, define as well.
Um, tone of voice, um, examples.
So this is a little hack that I,uh, always use, but.
I love to, uh, add examples to,to, to the system from, because
it makes it very clear for theAI agent how to, uh, behave.
But what I usually do for thatis, uh, screenshot a couple of

(23:43):
old conversations that arealready done by, by human, um,
and then put that in chat g pt,ask to transcribe those
screenshots and then, uh, put itin the, uh, agent's user, uh,
set or uh, user structure.
And then you already have someconversations that you can add
as an example as well to the AIagent.
And what AI of course, does is,is the whole goal of an LM or AI

(24:06):
is to predict your dream outcomebased on the information that
you are providing.
And that's also why you willnotice that AI is always.
Very much a pleaser because it'strying to predict your dream
outcome.
So the more information you cangive it, the better it can
predict your dream outcome.
Because the more context it has,the better you can predict, uh,
the dream outcome.
And the last, um, uh, the lastelement that I would always

(24:30):
include are rules.
And rules are basically, uh,maybe you, uh, will recognize
this as well is are creating AIagents.
Uh, you will notice that it doessome stuff that you don't want
it to do, so that I just put onrules or notes or something like
that.
Never do this, always do that,never do that.
And just when I'm playing aroundwith it, when I'm testing it,
then grills are just a bunch ofrandom stuff usually that I add

(24:50):
when, um.
Um, yeah, when I notice thatit's doing certain behaviors
that I don't want it to.

Yeah.
A a a quick example of somethingthat happens to me all the time,
uh, is these tools like to startwith a prefix to the answer.
So it will say, oh, based on theinformation that you provide
with me and based on, uh, thequestion that you asked, and
based on 1, 2, 3, 4, here's theanswer.
I said, no, I just give theanswer.
And because otherwise you willsay that multiple times in the
conversation, which is reallyannoying.

(25:16):
So basically telling it not todo any prefix and just provide
the answer, I.
Is something that I found veryhelpful.
So that's just an example so youunderstand in your head, uh,
what Carolina is talking about.
But, but there's a lot of thesesmall little tweaks.
Some of them are consistent,meaning once you learn them,
like the example I just gaveyou, uh, by the way, that's the
same thing in, in NA 10automations or make automations,
right?
It will sometimes give you moreinformation what you actually

(25:37):
need in the output, and then itmesses up the next steps of the
automation.
So you need to very clearlydefine what to include, what not
to include, what format toinclude it in.
And these just are part of therules that Carolina is talking
about.

Yes.
Yeah, absolutely.
When you're mentioning this, uh,it reminds me one of the
favorite one is of course the Mdash that's that everyone is now
so afraid of, of using becauseit's, it's clearly signals ai.
So I always try to, to tell theAI agent not to use an an M
dash, you know, like the, the,the, the long dash.
And its answers, it's, it'smostly not listening to me.

(26:09):
I do try to emotionallymanipulate the AI 80 by saying
that if it uses the M dash, thatwill, uh, get fired.
Sometimes it still decides toignore me on that, but, but this
is also a, yeah.
Really famous, uh, example ofone of the notes that you want
to exclude.

There's an interesting question before we
continue.
An interesting question in thechat.
What if there's no personal orcompany website, but, uh, so
that's the question from thechat, so thank you, Netta.
But the, I will broaden that.
Let's say I have a website,which is most companies, but I
wanna add more information tothe agent to have access to more
stuff.
Uh, how can I get, how can Iconnect it to additional data
sources?

Well, that's indeed the next step where we,
uh, are going heading nowbecause Oh, perfect question.
You can that in.
Yeah.
This is actually the perfectquestion because that's
something that you can do in theknowledge base.
And what I actually want toemphasize over here is, um, that
these principles, um, areactually the same for all the
different AI agent builder, ornot all, but most AI agent
builder tools.
So everything that I'mmentioning here, you will indeed

(27:05):
recognize this.
You mentioned as well in n it, nin make.
Um, so the principles are thesame, um, and the elements of an
AI agent are always the same.
Uh, the difference here withinFlow Gent is that, um, it just
makes it, uh, they make it veryeasy to connect it to different
tools like Instagram, uh, andWhatsApp.
And they have a chat manager andthey have certain features.
Uh, for example, for Instagram,uh, something like, for example,

(27:26):
if someone comes to keywords,then you can automatically send
a DM already to them, or, um, sothere is a lot of prebuilt, uh,
features in there that make it,uh, very easy.
Uh, but the elements are, uh,very much the same.
So knowledge base, um, here youwill find, uh, options to expand
your agent's knowledge.
So right now in the backgrounds,um, the AI agent team of flown
already, uh, edits the websiteas knowledge.

(27:47):
So all the content that is onthe website, it created an FAQ
based on the website.
But there's more things that youcan add.
For example, if you are usingNotion, um, you can connect your
notion and then automaticallyit's real-time things, the
notion page that you want the AIagent to have access to.
So this is, uh, one of thefeatures that I very much, uh,
that I find very useful because,um, this way you can make the AI

(28:10):
agent's knowledge very dynamic.
If you make any changes on theNotion pages, the knowledge of
the AI agent will also beautomatically updated.
So, um, this's something that's,that I like.
So connecting notion, but youcan also add, for example, PDF,
uh, you can just copy pastetext, you can add more q and a,
more websites, um, uh, pricingplans.
So these are, and they are, uh,adding more and more, uh, as we

(28:31):
go.
The, the next step is to chooseto pick your channel.
So these are the AI agents thatare really meant to, um, uh, be
able to connect with customersand leads.
So over here, you can, uh, editas a chat bot to your website,
just like the one that we'reseeing over here.
But you can also add it toSlack, uh, to Slack, to

(28:51):
WhatsApp, to Instagram.
And if you are more technical,you can trigger it with a web
hook or, um, access it throughthe flow Gen API or have it
scheduled.
But, um, yeah, the most commonlyused are Instagram, WhatsApp,
slack, and chat bots.
Then of course, one of the, andchat means

it's gonna be like a piece of code I can put on my
website.
That's what it means.

Yes.
Yeah, you, you can copy pastethe codes and then, um, edit to
your website.
So, um, yeah, it's, it's, I willshow you, um, as well, but it's
an, uh, yeah, it's just copypaste a small piece of codes
and, um, one of the mainelements of an AI agent is that
can take action.
So, um, you mentioned already,uh, before as well, is our, an
AI agent should be able to makedecisions.

(29:30):
It can, can, um, yeah, makedecisions based on the goal.
Uh, it should have knowledge.
And one of the big things of anAI agent, which really sets it
apart from just general chatbots or chat assistance, is that
an AI agent can take actions.
So it can connect with yoursoftware, it can connect with
your tools, and it can takeactions in those software and,
uh, those tools.
So in, uh, flow gen, they have,uh, a couple of, um, pre-built

(29:54):
actions already, such as sendinga Slack message, um, car.com.
I think they will release thatnext week.
But, uh, booking an appointment,kindly booking an appointment,
uh, sending email notifications.
But over here, um, they have acustom tool builder, which I
will also show where, which isreally, uh, comparable to
make.com.
You can build multi-step, um,tools.
You, uh, can connect it with anyAPI.

(30:15):
So this is really, um, how youcan make your AI agents so, so
advanced.
So what are things that you needto, to be thinking of?
Uh, for example, I helped acompany that wants the AI agent
to process and create invoicesand process that invoice
information.
So I, um, created a tool that's,uh, connected to their invoicing
software, and then at the AIagent could decide when it was

(30:36):
necessary to actually create aninvoice.
And then at those moments, itcreates an invoice.
Other things that you can thinkof is, for example, adding,
having the AI agent at leads toyour CRM.
So to connect it with HubSpotspower drive.
Um, other things that you canthink of, uh, is, uh, adding
todos, for example, to yourproject management tool
connected with Asana, connectedwith notion mandated com, any
project management tool to f theAI agent add to.

(30:59):
When necessary in a way that youcan have this AI agent actually
decide on, um, on when to do it,is by in the prompt.
So, um, like I showed youbefore, the system prompts, um,
you need to first connect thetools and then in the prompt you
can tell it, use this tool whenyou need to create an invoice or
use this tool when you want tobook an appointment.
So this really, you, yeah, youreally tell the AI agent, and

(31:21):
this reminds me that this is oneof the elements that you should
also, uh, add in a prompt is thetools that an AI agent can use
and when, so the process from

that perspective, again, just to clarify to
people, is you need to know, andit's the same by the way, in any
agent creation platform, whetheryou're creating this.
On flow gent or any otherplatform, you need to define and
create tools first.
And tools are sometimes thetools themselves and sometimes
tools that you make up, right?
So a tool could be connection toSlack, okay?
Slack is a tool, but sometimesit could be a tool that you

(31:49):
define that is something thatthe AI can do.
So a tool is basically, thinkabout it as something the AI can
now do and you can explain to itwhen to use this thing in the
system problem.
So this is, again, going back tothe onboarding, it's the same
exact thing.
If you are a new sales employee,uh, you got a new lead, go and
log it in the CRM, here's howyou do this.
So the, here's how you do this,this is how you build when you

(32:10):
build the tool.
And then just explaining thatthat's what needs to happen,
happens in the system.
Prompt.

Yes.
Yeah, exactly.
So, um, yes, and the last thingthat you want to decide, of
course, is what kind of AI modeldo you want your AI agent to
use?
So right now, uh, if you're onthe free version, uh, and I'm
showing you the free versionright now to show you that you
can just try everything forfree.
You can only choose between fouroh mini and 4.1.
Um, I always recommend in thatcase to use 4.1.
And in general, I like, uh,right now, as it's now June

(32:39):
12th, maybe Jet GPT will, uh,uh, uh, release GPT five soon.
But right now, I, I wouldrecommend, uh, using GPT-4 0.1,
but if you, um, upgrade to apage plan, you can also use
clouds.
Um.
They're adding, uh, O three, uh,GT O three as well.
So if you're not familiar withthe different models, um, 4.1
is, uh, a model that you can'tsee in check GPT, but it's

(33:01):
available for, um, yeah, theAPI, which means that you can,
uh, include it in, inautomations like these.
Um, I personally prefer 4.1over, uh, four oh.
Uh, but when you're creatingyour AI agent, just try the
different models, see howdifferent your agent is
responding.
Um, once you pick your model andcreate optimizer pro after that,

(33:22):
then when it's live, don'trandomly change the, the, the
model again later on because theagent can start behaving
differently, even if, if, ifit's the same prompt when you
change the model.
So that's something to be awareof

a hundred percent.
What it basically means is fordifferent use cases, different
models are gonna work better andyour only way to know, sadly, is
to actually test it outyourself.

Yeah.
Sometimes, um.
For me, I do notice that 4.1 isreally the, it does it all
model, but before we had 4.1,you could basically choose,
choose between 4.0 and or cloud3.7, and I would always go with
cloud 3.7, um, because it's, uh,it's, if it's was more complex
AI agent where needed to make,uh, different tool, uh,

(33:59):
different choices betweendifferent actions.
So in my experience, cloud 3.7was really, uh, good for an AI
agent to actually make thosechoices and know when to use the
tools.
And GT four O sometimes forgotto use the tool, kind of the
was, but GT four, four O wasagain, better for, for
conversational experience, was alittle bit more human, I would
say.

Interesting question from the chat, uh, from
Gwen.
So thank you Gwen, from thequestion, what about data
security and.
Privacy, right?
So you're gonna give it accessto your company website, you're
gonna upload documents, you'regonna give it access as you to
your Instagram account or yourSlack account.
So how do they maintain, wheredo the data go?
Uh, what risks are companytaking by integrating stuff like

(34:38):
that?

Yes, that's a good question.
So, um, I know that they storeall information on, uh, in the
eu.
Um, the software is GDPR proof.
So, um, and with the open AImodels, uh, that's also good to
know that, uh, open ai, if youare using the API of open ai,
which is being used there, um,GPT, uh, or OpenAI isn't, is at

(35:01):
least claiming that they are nottraining on your data.

Yes.
Just like

Google is claiming things and you, you,
you can never be home.
Sure.
But they at least promise thatthey're not training on your
data.
Okay, then, uh, yes.
So then we have, uh, we end theonboarding, uh, sequence here in
the playground and in thebackgrounds, the multiply AI
assistance is already trained.
So if you try this, uh,yourself, you will see, uh, also

(35:28):
an AI agent trained for you.
Right now, this agent isn'tconnected to any tools yet, but
you can chat with it already.
So, for example, what services,uh, do you offer?
It should also be, it's, I thinkthese are the colors already.
Yes.
It's this light blue color fromthe website.
Um, and here we see that it'susing the terrible M dash, but
it's, uh, answering correctly,hopefully.

(35:49):
So we offer a full range of, uh,AI powered solutions.
Um, maybe we can ask, um,something like,

you know what I have, I have an interesting
question to test it with.
Let's try that.
Yes.

I love that.

Uh.
Which podcast episodes wereabout AI agents?
Let's see, because, because Ihave, we have all the episodes
there transcribed and let's seeif it knows how to go through
that data as well.
That will be a very trickyquestion.

Yes.

Uh, because it has hundreds of pages.
So what did it say?
Can

you tell me it?
Okay, so it's search Knowledge.
It's uh, we have great episodes.
So this is the Powerful AIAgents and it's is, oh, I'm, I
should have opted in nta.
It is also sharing the link overhere.
Very cool.
Yes.
But we can maybe ask for more.
Uh, can you give me threeepisodes of the podcast, uh,

(36:40):
that were about AI agents?
Let's see if it can go evenfurther.
Even more.

Yeah.

Search knowledge.
Absolutely.
Here are three episodes.
Let's see if these are indeedabout ai.
H you know this better than Ido.

No, this is really, really cool.

Yes.
So it's also links already.
So this is, uh, yeah, alreadytrained automatically.
And this is just your baseversion.
So this is the draft version ofyour AI agent.
From here on, um, this helps youto easily at least, um, yeah,
get started because that'ssomething you already asked me
as well.
How do I get started or howdoes, does someone get started
with, uh, with creating aprompts?
Well, that's one of the moredifficult things to, to, to

(37:18):
start from completely scratch.
Um, so that's why they, uh, theybuilt this to, to make it easy
to go from zero to one.
And from here on you can, canmake any changes.
Uh, just answer your questionabout the website and that you
can copy the code and then youcan have platform, uh,
instructions over here.
So let's take a look at how youcan change it.
So let's first take a look atthe system prompts, um, and how
they defined it.

(37:38):
So over here, they have to role,um, like I mentioned as well,
lead generation assistants, uh,core responsibilities.
Uh, this is the same as thegoal.
Um, what you see over here isthat there are always hashtags
in front of, of the titles.
And if you're not familiar withthat, that's, uh, something
called markdown format.
And markdown format is a way totell to computers how something

(37:59):
is formatted.
And, um, in short, one hashtagmeans heading one, two, hashtags
heading two.
And this way the AI agents orthe AI will be, uh, better able
to understand the structure ofyour system front.
Because if you don't have this,it'll be just one big, um, yeah,
wall of text basically for theAI agents.
And it doesn't really know howit's structured and, and yeah,

(38:20):
how, how it's, uh, important.
It should, uh, different partsof the, uh, product part.
So, uh, it has coreresponsibilities, um, the tone
of voice.
Um, this tone of voice is basedon your, uh, website.
So the AI agents.
Team that took a look at yourwebsite, and based on that, uh,
it's decided on the tone ofvoice.
So if you have a very playfulwebsite, then the tone of voice

(38:42):
will be different as well.
Uh, conversation flow.
So it's, um, tells us, this is alittle bit like the SOP that I
mentioned.
So what should the, uh, way orwhat way should the conversation
flow?
You can make any changes here.
So if you don't, don't like it,then you, if you don't want it
to ask for an email address forexample, then you can make those
changes, key guidelines.
And here are some FAQ knowledgethat it created already.

(39:03):
But you can also change thatover here in the knowledge base
if you go to

FAQ.
Yeah.
Very cool.
So basically, like you'resaying, going from zero to one
is very, very easy because itdoes, it basically creates a
vanilla agent for the maintopics that they have, like
customer service or engagementor content creation with all the
instructions already built in,which also saves you the hassle.
Of my question before, of whatcategories do I need to have
there?
So these are already there andyou can go and just add and

(39:27):
change and massage and now youunderstand, but the general
structure is already built infor you and not starting with a
blank sheet of paper, which Ithink is fantastic.

Yes.
Yeah, exactly.
Yes.
So it makes it so much more easybecause, oh, I, I must admit, I,
I, I'm getting so lazy becauseof this, this, uh, yeah,
automatically generated frombefore this wasn't there and
that you needed to start fromscratch and really needed to
write everything and test andchange.
But this is already making lifeso much, so much easier.
You are even more, uh, lazy orwant to try more.
You can also click over here andyou can start recording and just

(39:56):
say what you want the AI agentto do, and you can have it
completely unstructured.
You can just say anything thatyou want and then also say, oh,
no, actually I don't want that.
And then it'll also, uh, createa prompt just like this.
Uh, so in inside the prompts,so.
Let's cont, uh, continue byadding it now to Instagram.
Because of the time, I won'tmake any changes to the prompt,

(40:17):
but if we're, uh, if you wouldreally create an Instagram, um,
AI agent, I would, for example,um, ask it to not collect full
name and email because we'realready talking on Instagram.
Um, but um, yeah, make thosechange.
Play around, uh, click alwayson, save and test them.
Play around with how the agentresponds.
But let's edit to Instagram asan example.
You can click on configure soyou're not

watching, there's just a channels section and you
click on that and Instagram isright there, and you click on
configure.
And that's, I assume, more orless it.

Yeah, that's it.
Click connection.
Then.
Make sure this is an, uh,important aspect that the
Instagram profile that you wantto connect it to, um, is the
profile that you are alreadylogged into on your, uh,
accounts, uh, on your browserright now.
So this is demo AI agent.
I also gave the Q profilepicture.

(41:06):
And, uh, let me, now you clickon allow and then the connection
is now being set up.
This may take a few momentsbecause it needs to make a
connection with Instagram.
So let's wait for a little bit,um, from the moment on that you
are connecting it.
Be aware that your AI agent willstart responding from now on
your D app.

(41:26):
So that's something to be awareof.
Um, in the chat manager of Flo,you can always look
conversations and look, meansthat the agent won't respond to
that conversation anymore.
So that's, um, something thatyou, uh, should be aware of.
Um, you can have differentsettings as well, uh, such as
that.
Uh, if you send messages insideof the Instagram app that the
agent is aware that you as ahuman sends those messages, um,

(41:49):
send, uh, you can have it senttwo updates.
So for example, let's say thatyou're an e-commerce store, you
connected with Shopify, and you,um, want the agent to update
saying, Hey, I'm right nowlooking into our systems to see
if, uh, uh, if it's happen, uh,or.
If something is happening, if Ican do something, then you can
have that updated here as well.
I usually have this turned off.
Um, load al messages forconversations.

(42:10):
So the moment, this is also oneof the data question, uh, answer
to one of the data questions.
So if you connect yourInstagram, um, the floating
software won't see, won't seeall your old conversations, and
only if you, uh, allow it.
Then the moment that someonestarts talking to you from the
moment onwards that youconnected it, it can see the
last 20 messages already.
So then at least theconversation will keep on going.

(42:32):
Um, something that I like to usefor influenzas is a random
response delay because ofcourse, an AI agent will start
responding immediately to you,uh, normally, but that's a
little bit inhuman.
And I also help, uh, influencerswho, uh, want to automate their
Instagram dms but don't want.
Their followers to know thatthey are actually automating it.
Um, then you have a randomresponse delay.

(42:54):
So it will take any time betweenI think two or uh, uh, one or 10
minutes to respond.
And you can also set workinghours, for example, if you want
your customer support team toanswer during the day.
But once AI agent to answerduring outside of office hours,
then you can set that over hereas well.
So these are all functions thatmake it very useful, um, to
actually, uh, connect it withInstagram.

(43:15):
Now let's see if our AI agent isgoing to respond on Instagram.
Uh, then I need to quicklysearch for the, what's the AI
agent name?
Demo, AI agent.
If you want, um, anyone, you canjust start chatting with my AI
agent.
Demo AI agent as well.
Uh, let me see.
It's over here.

(43:37):
Follow, and I'll start sendingit a message just to make sure.
I think those of

you're not, or not watching and just listening, we
just open Instagram with adifferent account so we can chat
with that account.
Uh, and then we could just gonnasend it a message and we'll see
how it responds.
Uh, yes.
And we turned off the randomdelay, so it actually answers,
so it actually answersimmediately.

I hope.
So.
Let's see.
It'll take, uh, it'll alwaystake a couple of seconds.
Why?
Um, because in Instagram it'svery common that, um, people are
sending multiple messages orsending.
One big message for them.
Send it in multiple messages,right?
Oh, yes.
So this is a, a built-in delayto make sure that, let's see,
well, to multiply, uh, sorry.

(44:24):
To make sure that, um, all thosemessages are combined sent to
the AI agent, because otherwiseit'll respond to every message
individually.
And that yes, gets a reallyweird conversation.
So it's working.
Welcome to Multiply.
We're excited to help you tospur your business with ai.
What brings you here today?
And now we can also ask, uh,what's happening?
Oh, I think that's okay.

(44:45):
I think Have voice

pipe somewhere.

No, what what happens right now is, uh, the AI
that I'm logged into right now,and I need to stop it, uh, is,
or the Instagram channel thatI'm, uh oh, has its own AI
channel.
It also has an

so there, so the agents are talking to one
another.
This is the end of the worldthat we know it.
So let's, let's pause here.
I think, I think this was greatfrom a demo perspective, I want
to do a quick summary.
Uh, and, and kind of explaineverything we talked about and
then I'll let you tell peoplehow they can find you and work
with you.
But to, to summarize everythingwe talked about, one is we
talked a little bit about whatour agents and how they're

(45:18):
different.
But the most important thing isif we ignore all the explanation
part, the process to create anagent and put it on, in this
particular case, either yourwebsite with a code that will be
copy paste, or on a channel likeInstagram or Slack took maybe
five minutes, right?
Because we give it a link to thewebsite and then we waited and
there was a first draft of theagent ready.

(45:39):
All we told it is what is thegeneral role in this particular
case, like a customer serviceone.
So it built a system prompt foryou.
So if you know nothing and youwant, okay, I wanna start at
least experimenting with thisfive minutes later.
You can have an AI chat thatknows how to work with different
tools.
In this particular case, wedidn't connect any, but we could
have, uh, so let's say I havelike a way to book stuff on my
website or, and I want toconnect it to the back end of

(46:01):
that or a CRM or anything likethat.
Uh, literally five minutes lateryou have a good starting point
and it's extremely powerful ifyou want to get started quickly
and without any previousknowledge.
And the other thing is itcreates a very solid start for
the system problem.
So you can, from there, just youkind of understand what's going
on and you can just add yourstuff.
And you don't have to be anexpert on creating agents.

(46:23):
Uh, Carlita, this was fantastic,like really a very, very well
thought after a introduction andthen the actual demo, if people
wanna follow you, learn fromyou, work with you, uh, connect
with you, what are the best waysto do that?

Yes.
So, um, you can find me onLinkedIn, Karina.
Um, also on YouTube, I create,uh, videos on YouTube also
making easy to followexplanations on how to get
started with, uh, AI agents andmy event to, to share that.
Um, this Monday I'm launchingthe AI Agents Made Easy School
community.
And, uh, yes, I'm very excitedfor that because the goal of

(46:56):
this community is really to, towelcome anyone who is, uh,
doesn't consider themselves as,as technical to, uh, come join
us and learn about AI agents,learn about ai, uh, automations,
and really go from beginner towinner with AI agents and
automation.
So everyone is welcome, even ifyou feel like you have no
technical experience at all.

Fantastic.
Uh, thank you so much.
Thanks everybody who joined uslive.
We had multiple people in a veryactive audience.
Sadly for you, you missed that.
But a lot of good questions andconversation on, on LinkedIn, on
Zoom you were at least able tosee.
Uh, but thank you everybody forjoining us live.
Uh, just for, for yourknowledge, Carolina, we had
people from, uh, uh, let me see,they said in the beginning, but
from many places around theworld from, uh, Singapore to

(47:37):
India to UAE to San Diego, CapeTown.
So you had a co literally peoplefrom all around the world, uh,
in your session.
So this was absolutelyfantastic.
So thank you everybody, uh, forjoining us, uh, for being
active, for asking questions,for chatting with, uh, with me,
and with each other in the chat.
And thank you so, so muchCarolina.
Again, this was absolutelyfantastic.

Thank you so much for having me.

All Episodes

198 | Build a Business-Boosting AI Agent That Books Calls While You Sleep-No Code w/ Carolina Posma

Episode Transcript

Popular Podcasts

Crime Junkie

24/7 News: The Latest

Stuff You Should Know

.css-15opob5{left:0;position:absolute;top:0.8rem;} All Episodes

.css-14f5ked{margin:0;word-break:break-word;display:-webkit-box;-webkit-box-orient:vertical;box-orient:vertical;-webkit-line-clamp:2;overflow:hidden;}198 | Build a Business-Boosting AI Agent That Books Calls While You Sleep-No Code w/ Carolina Posma