Everyone’s talking about AI Agents, But few are showing how to actually use it in a way that saves time, uncovers real insights, and drives business decisions.
In this live episode of Leveraging AI, Pooja Jain — founder of PowerUp AI and one of LinkedIn’s rising AI educators — is going to take you step-by-step through the exact process she uses to build custom AI agents. No fluff. No code. Just the “how to” you’ve been missing.
You’ll see a live demo of a real AI agent built in Relevance AI that handles competitive analysis — scanning websites, doing sentiment analysis, pulling customer feedback, and even giving positioning suggestions based on gaps in the market. Yes, it actually does things (not just spits out summaries).
Meet Pooja: A former Procter & Gamble leader, Pooja now trains executives and C-suite leaders across Europe to integrate no-code AI and automation into their businesses. She’s already taught 170+ leaders — how to lead AI initiatives without writing a single line of code. She knows what works (and what doesn’t), and she’s here to show you the difference between tools, automation, and real AI agents.
About Leveraging AI
- The Ultimate AI Course for Business People: https://multiplai.ai/ai-course/
- YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
- Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/
- Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events
If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!
Mark as Played
Transcript
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
(00:04):
Leveraging AI Podcast, thepodcast that shares practical,
ethical ways to leverage AI toimprove efficiency.
Go your business and advanceyour career.
This Isar Metis, your host, andwe have.
Maybe the top exciting topic ofAI in 2025, which is agents,
which is going to talk abouttoday, and agents are already
here and I know a lot of peopleare hearing about this.
(00:27):
It's something that is comingand it's the next stage, and
it's the evolvement of LLMs, butagents are already here and
multiple people and companiesare already generating
significant business value byleveraging.
Agents across multiple aspectsof the business.
But for most people and mostcompanies, there's still this
elusive concept that peopledon't understand and definitely
(00:50):
do not know how to implement.
Now, this gap between the peopleand the companies who have
agents to those who don't isspreading every single day, and
it's widening.
And you want you, yourself, as aperson with capabilities as well
as your companies to be on theright side of that gap.
(01:11):
Meaning you want to be on thatfast bullet train, that
accelerating and providing moreand more capabilities to your
business versus being standingon.
The train station and watchingthat train getting further and
further away from you,especially if you know that your
competitors might be riding thattrain, meaning they'll be able
to do more and more than you canfor less money and then be a lot
(01:35):
more competitive than you cannow, developing agents sounds
like complex, and it sounds verytechnical and it sounds like
only big companies likeMicrosoft with a lot of people
who write code can actuallycreate them.
But the reality is there aremultiple tools out there today
that allow you to write reallypowerful agents with either no
code or low code.
(01:55):
Today we're gonna focus on nocode at all.
So one of the most powerfultools out there that enable to
do that is called relevance ai.
And our guest today, Puja Jain,is a relevance ai.
Expert and she has beendeveloping agents for multiple
companies on relevance for awhile now across different
aspects and different businessesand so on.
(02:17):
Now, in addition to the factthat she knows relevance really,
really well, she spent years indeveloping and implementing AI
solutions at Proctor and Gamble.
So in addition to her recentexperience and just building
them on relevance, she hasenterprise level experience in
understanding what is requiredto actually develop agents that
actually work in a businessenvironment.
(02:38):
And she knows how to do it in astep-by-step in a tool that
anybody can use, which makes herliterally the perfect guest to
talk about, this topic with us.
So this is exactly what we'regoing to do today.
We're going to show you anentire process, beginning to
end, how to develop an AI agent,what to think about and how to
use relevance in order to dothat.
But having that knowledge andseeing this will allow you to
(03:00):
develop other agents and onother tools because the concepts
are exactly the same.
Now, as I mentioned, sinceagents are maybe the most
transformational technology weever created, at least that I
know of so far, then.
This is a really important topicand hence I'm really, really
excited to welcome Puja to theshow.
(03:21):
Puja, welcome to Leveraging ai.
Pooja Jain (04:06):
Thank you so much, Isar, how are you?
Isar Meitis (04:09):
I'm doing awesome.
I'm really, really excitedabout, this session.
I myself am tinkering withdifferent tools.
I'm definitely not close to yourlevel, in doing this, so I'm
really excited.
I'm sure a lot of people arereally excited as well.
Our top performing episodes inthe past few months are all been
around agent development, so I'msure there's a lot of people who
are really curious, about thistopic.
It's everywhere in the news ifyou're following an ai, but
(04:31):
still, I think most people nowget a feel on how to use large
language models and imagegeneration.
I think 99% of the people don'thave a clue how agents work and
how to create them.
And so I think this is gonna bereally fascinating.
before we get started, a few,messages.
First of all, thank you toeverybody who are joining us
live.
Whether you're joining us onLinkedIn or joining us on Zoom,
we really appreciate you beinghere.
(04:52):
I know all of you have otherstuff to do on, Thursday at
noon, Eastern time.
But, feel free to introduceyourself.
Tell us what you know, what youdon't know about agents, and
what you wanna know from thissession as well.
That will help me, guide theconversation, maybe even more.
Tell us where you're from so weknow where people are joining
from.
That's always fun for me to see.
There's always people from allover the world, at least on
this, on the, LinkedIn side ofthings.
(05:13):
There's always a lot of peoplefrom, interesting places.
So introduce yourself, make newfriends.
In addition, one last thing thatI will say, is that if you are
still watching us live, the nextcohort of the AI Business
Transformation course startsthis coming Monday.
So when this becomes a podcast,if you're just listening to this
after the fact, you missed thatparticular cohort, but the
(05:33):
cohort starts on May 12th,meaning you still have a few
days.
This is this coming Monday, thiscourse.
Is really transformational forpeople, for companies, for
careers, for complete teams, andwe've been teaching it.
I've been teaching it for thepast two years every.
Single month, sometimes twice amonth.
So hundreds or maybe thousandsof business people have been
through it and have reallychanged their businesses with AI
(05:56):
based on the information thatthey've learned.
So if you have not startedproper implementation of AI in
your business, this is anincredible opportunity to get a
huge amount of information, bothon the tactical side of what
tools to use, as well as on thestrategic side, on what's the
right process to implement AIcompany wide.
If this is interesting to you,I'll drop the link in the chat.
(06:18):
If you're listening to thisafter the fact, then there's
gonna be a link in the shownotes to tell you when the next
cohort is, which will probablybe in August.
because we in between, weusually teach private courses
and we're usually fully booked.
that's it.
With all of that being said, Iwill give puja the microphone.
I'm really excited to see whatyou have prepared for us and,
get ready everybody to learnhow.
(06:39):
Simple.
It actually is to create agentsthat are extremely powerful.
Pooja Jain (06:43):
Thank you so much, IAR for having me here today.
And yeah, super excited to sharewhat I am working on and what I
have learned about agents sofar.
Honestly, speaking agent AI isevolving.
It's evolving on a weekly basisnow.
We are seeing something newevery week coming up, so it is
even new for me, although I havebeen working in the automation
space for many years now.
(07:03):
But yeah, agent AI is somethingnew.
And today, what I'm going toshow you is first like
clarifying a bit.
So what are AI agents actually,and how do they differ from say,
an automation workflow becausethat's where I think most of the
people are still strugglingwith.
And secondly, showing a livedemo in relevance AI about this
market research AI agent that Ihave built, and what is the
(07:26):
concept around it?
Should we get started?
Isar Meitis (07:29):
Yeah, let's dive right in.
I'm, I really wanna learn aswell.
Pooja Jain (07:33):
Perfect.
Let me share my screen quickly.
Isar Meitis (07:36):
Sure.
for those of you who are,listening to this as a podcast,
and I know thousands of you are,we will explain everything that
we're seeing on the screen soyou're not missing anything.
But if you do wanna watch whatwe're doing, you can either
switch to watch this on YouTube.
So go to our YouTube channel,then multiply AI YouTube
channel.
There's a link in your shownotes, or if you are driving
right now or mowing the yard orat the gym wherever you are,
(07:58):
when you cannot do this, thenyou can keep on listening to us
and then you can decide later onif you also want to watch the
YouTube video.
Pooja Jain (08:05):
Great.
Okay.
So firstly, what exactly is anAI agent?
of course, like by now almosteveryone is aware of what is
Chad GPT?
So Where do agents fit in here?
So imagine when Chad G PT comesin, even now, to get something
beneficial out of Chad G pt.
You have to give in a properprompt.
You have to add in your businessknowledge, and this becomes a
(08:27):
repetitive process, right?
And Chad g PT is still a chatbot.
So it only talks, it does notexecute.
And this is why AI agents arereally changing, the landscape
for the businesses because theycan execute.
So if you want to understand AIagents in very simple terms,
think of them as let's say ifchat GPT gets knowledge, so your
(08:49):
business specific knowledge, itgets a memory.
I mean, a chat g PT has amemory, but really, you know,
something that is a memoryspecific to your business, to
your style of working as well astwo hands that it can start
executing.
Actions in your systems.
Now, it can be something assimple as writing an email or,
(09:09):
managing your calendar, whichcan be chaotic for a lot of
people.
it's a big productivity hack IfJGBT can start managing your
calendar or something as complexas, doing a proper research and
then going into your CRM andupdating the fields there.
So this is what, AI agents areall about, like they can
(09:29):
understand, of course, becausethe underlying power is still
LLMs, but in addition to that,they can also execute actions in
the systems in your business,ecosystem.
Isar Meitis (09:40):
Yeah.
I'll add, two, two more things.
like when I think on agentsversus large language models.
So you mentioned, one veryimportant thing, which is the
ability to take action.
The two other things that agentsdo that LLMs don't do, one is
make their own decisions.
So they're more autonomous.
They don't have to be, by theway, but they can be more
autonomous than just a largelanguage model, meaning you can
(10:01):
give them a broader task andthey will figure out the steps
on their own without you havingto define an exact path on how
to get there.
And the other one is.
If you build it correctly, theycan be multi-layered, meaning
you can have an organization ofagents working together, like we
have teams of humans.
You have a manager, you havesomebody writing content, you
have an editor, you have anevaluator, you have a designer,
or a similar parallel in anyother part of the company, the
(10:24):
same thing.
You can build agents and thenthey can work collaboratively
versus just working on theirown, which is what LLMs do.
So these are the maindifferences.
Pooja Jain (10:31):
Very good point.
ISAR because, and this is alsoone of the major differentiators
when you start thinking aboutagents from workflow automation.
so workflow automations arereally like, very static.
So let's say if this happens,then they should happen.
There is no autonomous behaviorthere.
you have to predefine everythingand they execute, which is
fantastic for a lot of businessprocesses because not every
(10:53):
business process needs, thatlevel of autonomous behavior.
However, AI agents, they canselect their own tool.
They can select as I mentioned,they can select their tools.
they can select the next stepand so on.
So where do you really need AIagents?
So AI agents make sense whenyour process is super complex,
involves multiple steps andmultiple level of decision
(11:13):
making.
If this is the case, then youshould start thinking about AI
agents.
Otherwise, if your process ismore about, say, repetitive
steps, that more or less staythe same.
You are good to go with aworkflow automation.
Isar Meitis (11:26):
Yeah.
I always tell people now startedcalling, Zapier and make like AI
agents are like, no.
If it's doing just a step bystep process, you don't need AI
in it or you can bring AI in itinto specific steps if you're
trying to analyze what surgeonin an email as an example.
But the rest is just old schoolautomation that existed for a
decade now.
Pooja Jain (11:43):
Yeah, absolutely.
Oh, by the way, Zapier has alsointroduced their AI agent
builder, which is a no-code aswell.
I think that most of theseworkflow automation builders are
now also moving towards agentai.
Yeah.
Great.
And then second point, that is,I just mentioned about this AI
agency, so that is exactly whatI would be showing today.
(12:04):
So my agent is about a marketresearch AI agent, or rather a
team of AI agents.
And when you start thinkingabout this.
you have to really think of itlike a team of, humans that are
running this.
So let's think of it, there is amanager agent, and manager agent
is the one that is coordinatingwith all the subagents in this
(12:25):
team.
I, as a human, interacts onlywith the manager agent.
So I only give my, request or mycommands or whatever I want to
do, only to the manager, agent,manager.
Agent then, picks like the nextsubagent.
for example, in this specificuse case, so I build this for
competitor research, there areseveral subagents, like industry
(12:45):
analyst, competitor tracker,customer sentiment analyzer,
social media auditor, and then areporting agent that is really
gathering information from allthese subagents, synthesizing it
and sharing it with the manager.
It If you think about it, it'sexactly like how you would work
in a team of humans.
And that is actually the basicsof even when you start setting
(13:08):
these up in any platform.
So manager agent is really whoyou should be telling in detail.
think of it like a jobdescription or a very detailed
SOP that is literally yourprompt that manager agent should
understand and then the manageragent in return, from based on
this prompt, should be able toexecute these, or rather pick
(13:30):
the right subagent for yourtask.
We will see this in a demo.
I think that will make it, moreclear.
Isar Meitis (13:36):
Yeah.
Pooja Jain (13:36):
anything, to add here?
Isha.
Isar Meitis (13:38):
No, I think it's great.
I really think what you said isthe important thing, think about
what humans you would need andeven make it more granular,
because in humans, in manycases, there's one human that
does several of this.
So think about it more on thetask level rather than on the
human participant level.
And because every agent will bedoing one task that will enable
it to be very good at thatparticular task.
(13:59):
And then the manager agent, andthen you can also add layers,
like a improver, like somebodythat reviews the work and adds
Yeah.
Comments and so on.
there's different layers you canadd, but still each agent will
be focused on one task andthere's will be one or more
agents that will help tocoordinate the task to make it
more effective this time around.
As well as moving forward.
Pooja Jain (14:20):
Yeah, that's a great explanation.
I agree to that because when youlook at it, I think in an
organization you would not see ateam that where one person is
just focusing on industryanalysis.
Isar Meitis (14:29):
Yes.
So
Pooja Jain (14:30):
it can be equated to, I don't know, a very well
funded startup where you have aone person doing one task.
Isar Meitis (14:37):
Yeah.
Pooja Jain (14:38):
So now let's go into relevance ai and I will start
describing you how relevance AIis, but I'm hoping that there's
an interesting
Isar Meitis (14:47):
question, which I know the answer and we're gonna
actually demo that, but it's agood question to ask.
The question is, is the humansets up the subagents and not
the manager agent spins them upin real time?
And the answer is yes.
and I think it's a yes, but, soI will let puja answer the rest
of it.
Pooja Jain (15:01):
Yes.
I'll be showing that during thedemo, and if it is not answer,
happy to take that at the end.
Isar Meitis (15:06):
Okay.
Perfect.
Pooja Jain (15:08):
Okay.
So what you are seeing on mysystem is really the relevance
AI interface.
It looks a bit messy becauseI've not really organized my
workspace here.
but relevance AI is a no code,AI agent builder.
And if you go into relevance ai,they have now terms and terms of
templates.
So when you start going, likeyou can create a free account if
(15:29):
you want to get started, andthey have a lot of agent
templates, which are fantasticway to start.
So if you're just starting out,I think the best way would be
take one of these templates andstart editing it for your
requirement.
That is the easiest, right?
Then there are a lot of tools,so these are really the
integrations that are alreadypresent in relevance.
(15:49):
So for example, if you want toextract something from LinkedIn,
they already have a tool for it.
All you have to do is select it,use it in your agent or in your
project, and adapt the prompts.
Now let's go into this agentthat I am working on, which is
the co research agent.
(16:09):
Okay, so firstly, let's see,maybe I will first run this demo
so we see the output and then Iwill start explaining it one by
one.
So this competitor, analystagent, as I explained, is a team
of subagents that is doingindustry analysis competitor.
So it can go onto yourcompetitor's website, track the
latest updates.
(16:30):
It then goes into, sayTrustpilot or any other review
website that you have givenaccess to, gets the latest
reviews.
I have added an additional layerthere.
So not just get the reviews,because let's say if the
competitor is doing well insomething, I want to understand
why or if there are any gaps.
I also want to understand why.
So it has an additional layer ofsentiment analysis, tracking.
(16:54):
I have also trained it to goonto the LinkedIn or any other
social media, for example, togather what is happening at the
social media of the competitor.
And finally, it synthesizeseverything and sends me over on
Slack.
So let's give it a demo.
So let's say I just give, use ademo use case.
Okay.
I'm a sales head for, AIpowered, CRM for SMPs.
(17:18):
So I'm simply describing what mycompany's doing in very simple
terms.
Nothing like, we do not need apage long prompt for this.
And I would like to.
Get the competitor insights, sayon Pipedrive, because Pipedrive
is like the CRM, which is very,like a competitor for me, right?
So I would just simply run that,okay.
(17:41):
As you can see, I've simply,talked with this manager agent,
right?
So what you saw on my window wasthe manager agent.
I simply, communicated what Iwant to do and it has started
working in the background.
Now, first thing is, it isextracting the industry news for
me.
so this is another subagent thatis running in the background.
Now let's open another window soI can show you what is
(18:05):
happening, in the background.
Okay, so I just, ran this likebefore the session because
sometimes, it is extracting alot of information.
Yeah.
It'll take time from the web.
It takes time.
So let's see what it did.
It extracted the industry news,which is specific to pipe drive
and which is very specific tothe AI initiatives of Pipedrive.
(18:27):
Because I specifically saidthat, I am doing this AI powered
CRM.
Isar Meitis (18:31):
Yeah.
Pooja Jain (18:32):
So it just extracted everything along with, the
source.
Then, so those of you who arejust listening,
Isar Meitis (18:38):
there are multiple articles with each and every one
with several bullet points andtopics and summary, and then
there's a final analysis of allof them together.
Pooja Jain (18:47):
Yes.
So this is one agent that didits job, second agent, which is
the Trustpilot Reviewsummarizer.
What it did, it went on toTrustpilot, it got the latest
rating of Pipedrive.
And based on that, it gave melike the key positive points.
what is happening there?
key pain points, so for example,it find out that based on the
(19:08):
reviews that there is arecurring issue, in the
limitation of the form ofcomponents, particularly
regarding constant tracking forsubscriptions and so on, pricing
transparency and the learningcurve.
So these are already very good,insights for me.
Now, let's go further.
It then did a competitor,product benchmarking.
So what it did, it went to pipedrive's, website extracted, all
(19:31):
the latest CRM features thatpipe drive release because that
is what is relevant to me.
It summarized that, and finally,it summarized everything like,
synthesized all this informationfrom different subagents,
summarized it and gave me like areport here.
for example, the strengths, whatare the pain points, current
capabilities, and so on.
(19:52):
Now, one of the questions that Ioften get here is, how is it
different from, a workflowautomation, right?
So one thing here is, let's say,if had this been in a workflow
automation, but if I want torerun this process, it would
always run everything from startto end.
Whereas now, let's say if I amnot happy with the, Trustpilot,
(20:15):
tracking, for example, I cansimply say, I would like to
rerun or redo the, sentimenttracking.
And what would happen here isbecause the manager agent here
is responsible for choosing ordelegating to the subagents, it
would identify that only thatparticular subagent needs to be
(20:36):
reactivated now.
So it is only using theTrustpilot review, summarizer
subagent and not rerunning thewhole process.
And that basically is the bigdifference when it comes to
workflow automation because hadthis been like a, Zapier flow or
make flow, it would have rerunthe whole thing and not just one
particular component of it.
Isar Meitis (20:57):
Yeah.
And this connects beautifully toa question that was in the chat
that said, That so far, likewhen you just started running
this, it looks similar to deepresearch.
How is that different?
And I think you answered somecomponents, but I wanna dive
deeper into that because I thinkit's important for people to
understand.
First of all, deep research isan agent.
So the biggest differencebetween deep research and just
(21:18):
using chacha PT research searchis exactly that concept, that it
understands your question andthen it says, oh, the person
wants to understand this topic.
What do I need to do?
Which is now an autonomous thingto actually give him the
information that he needs.
The biggest difference betweendeep research and this, and
there are several differentdifferences.
Difference number one is you canin advance define all these
(21:39):
subagents and you will knowexactly what it will do.
Meaning in deep research, youcan't control what's.
Sources it will go to which one,it's not gonna go to what topics
you're interested in and so on.
So when you build an agent for aspecific topic, like in this
particular case doingcompetitive research, you can
build it the way you want it.
So think about it like acustomized version of deep
(21:59):
research.
That's basically what it'sdoing.
So that's difference number one.
Difference number two is, asPuja said, you can go back and
forth with specific componentsof this because they're
standalone agents, which is notpossible, or it's possible to an
extent, with deep research, butit will still not be as tailored
in specific as the specificagent that you develop.
(22:22):
And number three, which whatyou're gonna see in a minute,
this will actually go and updateyour CRM and do other things
that obviously deep researchwill not do.
So think about this as a.
Multi-layered, multi-level, morecustomized version of deep
research that can also then goand do stuff like write it to
you in Slack or update your CRM.
(22:42):
Anything you want to add, Puja,because I think it's an
important topic for people tounderstand.
Pooja Jain (22:45):
Yeah, I think that summarize it beautifully.
another point, human in theloop.
Deep research does not havehuman in the loop, right?
You have no control.
You cannot really say, okay,there is no concept of asking
for permission.
I will show you in a second thatas a custom AI agent, you can
really, train it to ask forpermissions in certain areas
(23:09):
like where you want it to, notrun on its own.
Awesome.
okay.
Yes.
Isar Meitis (23:14):
Another question that is also I think very
important and then we cancontinue is, there was a
question about hallucinations.
Is there a way to know or toreduce or to verify the
information that is coming fromthese research agents?
Pooja Jain (23:28):
it is possible to, of course, like you can train
your AI agent to ask for thesources.
So for example, when I wasbuilding this agent, I was
really not sure if it is gettingthe right information from the
trustpilot.
So I was checking it again andagain while building it.
I was always verifying, forexample, the stars that it shows
me here are correct, the reviewsare correct or not.
(23:49):
And the date, because I wantedto only go one month, get the
reviews one month older only.
So that is something I wasalways manually checking.
I think this is a very importantstep.
Like you have to be verycautious when you are designing
your AI agent again, human inthe loop.
I think this is one of the mostor the ultimate guardrails that
can be part of your AI agents.
(24:11):
Second, the system prompts,they're the best ways to control
your AI agents.
I will show you in a secondwhere you can add your system
prompts.
So whenever you are designing anAI agent system, prompts are
really the controllers.
That is the information and, thekind of, positive prompts as
well as negative prompts thatyou add in your ai, system
(24:32):
prompt.
They are the ones that would becontrolling everything for you.
Isar Meitis (24:36):
Awesome.
So I'll add my 2 cents.
Yes.
When I do deep research on stuffI really care about and I really
need to verify the information,I usually run it on three
different AI deep researchtools.
So think about running threedifferent agents to do the
research for you.
And then I have a fourth processthat actually creates a
comparison table and checks.
If all three have the sameinformation, when all three have
(24:58):
the same information, then it'svery unlikely they all made up
the same stuff.
And so that's most likelyaccurate information.
And when there are outliers,when only one of them finds a
piece of information, the samefourth agent go and checks that
information is correct.
So again, now I get a secondverification if the information
is really there from thatparticular source and all the
(25:19):
outliers are not found, are justbeing thrown to the trash.
And I get a summary of that.
So it just depends on how.
Buttoned up, you need thisinformation to be right.
let's say there's 500 Trustpilotreviews.
If 10 of them are made up, 15%are made up, still not a big
deal, you're still getting the500.
But if you're looking onsomething that you're gonna make
a very important businessdecision on, you want it to be a
(25:40):
hundred percent accurate and not90% accurate, and then you can
add these additional layers andsteps in order to dramatically
increase the chances that theinformation is correct.
Pooja Jain (25:49):
Great point.
Great.
So let me show you the conceptof human in the loop, which is
absent on, definitely on theworkflow automation because
there is no way that you can addlike a, Human, approval step in
there, or even in the LLMchatbots, you give them a
question and they answer thereis no place to even ask for
approval.
(26:09):
But here it is the case.
So let's say in this demo thatI'm running, I am happy with the
output that I have got so far.
Now I want a summary of it, andI want this on my Slack channel.
I have said this AI agent in away that it does not spam my
Slack channel with everything.
It always asks for approval.
(26:30):
And only when I'm happy with theoutput.
let's see.
Okay.
I should not be writing please,but I'm used to.
I do this,
Isar Meitis (26:39):
I do the same thing.
By the way, the jury's still outon that.
So those of you who don't knowthis whole discussion, there was
a big debate last week, based ona post from several different
people, including Sam Altman, onthe cost of saying, please, and
thank you to these chatbots.
And there is inconsistency inthe knowledge whether that
actually makes the resultsbetter.
I think the best analysis thatI've seen, came out from, brain
(27:04):
Dead.
it'll come back to me in aminute, but he basically said it
depends on the use case.
Sometimes it helps and sometimesit doesn't.
uhhuh, but I still do thisbecause that's how I'm used to
typing.
And so I'm just like you.
Pooja Jain (27:15):
is this a report from Ethan Molik?
Isar Meitis (27:16):
Ethan.
Ethan Molik.
Yes.
Thank you.
Ethan Molik.
Pooja Jain (27:18):
yeah.
I read it as well.
It's interesting that there area few things we cannot get rid
of, which is nice as well.
yeah, that's fine.
They're going to be ourteammates, right?
Isar Meitis (27:27):
Yeah.
Pooja Jain (27:28):
Okay, so I simply asked that, now I'm happy with
the output.
Now, please send me a summary onthe Slack channel and as you can
see it called the Slack channel.
so it called the relevantaccount, from the Slack channel.
Created a summary, a very nicesummary of the key trends,
vulnerabilities, pricing,transparency, and so on and so
forth.
(27:48):
And now, but it did not run onits own because this is, like I
have said this pro stepspecifically for human approval.
So only once when I click onapprove, it would go to my Slack
channel and send all thisinformation.
Otherwise it would not.
These two steps.
this single step actuallyprobably should clarify for most
people what is human in the loopfirst thing and second, how an
(28:10):
AI agent differs from, say l andm chat bots, deep research or,
even a workflow automation.
So I would not say approve,because my Slack channel is full
right now.
I don't wanna open that, but Ihope the concept is clear.
Isar Meitis (28:24):
yeah.
Pooja Jain (28:26):
Wonderful.
I think what will beinteresting,
Isar Meitis (28:28):
is to dive into what are the instructions that
make these agents do it, likethe actual creation of the
agents.
I think that's gonna be, yeah.
I think the output are nowhopefully clear to people.
So let's dive a step deeper andshow how they're actually built.
Pooja Jain (28:41):
Yeah.
So let's go into the build.
Great.
So first let me Demo a bit aboutthe interface that you are
seeing here, and then I will godeeper into it.
So what you are seeing right nowis really the AI agent a as it
should look, and theinstructions here, as you can
see, they are like super longinstructions.
These are the system prompts andwhenever you are designing an AI
(29:03):
agent system.
Prompts are the key.
Like they are the ones that arecontrolling everything.
As I mentioned, the systemprompts are the controller.
So make sure first, in yoursystem prompts, for example, to
mention what would be thefunction of this particular AI
agent.
when I explain it to myaudience, I always say, think of
(29:23):
it like a very well defined jobdescription.
So you have to define what youragent should be doing.
What tools does it have accessto?
So I already defined, you haveaccess to industry news.
So basically telling my AIagent, so as an onboarding
process, who all are there inyour team and when you should be
calling them.
(29:44):
That comes later.
But what are your core functionor your core to-dos?
when you go on a website, whatyou should be looking for, then
what you should do and what youshould not do.
So this, what you should not dois super important.
So for example, do not assumeanything.
ask if something is not clear ormissing, because a lot of times
(30:06):
what I see is, let's say when Idid not add this, negative
prompt, it was always assuming alot of information.
Which is something I do not wantwhen I am running it for a
business process.
So now adding this negativeprompt made such a huge
difference.
So this manager agent is nowcoming back to me asking me what
is missing?
And you can add a lot ofguardrails in there.
(30:27):
So for example, if you aredesigning a customer chatbot, a
chatbot that is talking to yourcustomers, you can define
specifically like a relevanceclassifier.
So to classify if whatever thecustomers are asking this
chatbot are actually relevant,and, they're not asking for a
piece of recipe to, to your, Idon't know, CRM chat bot,
because that happens.
(30:47):
you should be defining all thishere.
Second the tools.
So really what, so just beforewe
Isar Meitis (30:53):
jump to the tools I want Yes.
I wanna go back because I thinkit's very important what you
said, Just think about definingin simple English, but in a very
well structured way, all thethings that you want the agents
to do and all the things you donot want the agent to do.
And so the more detailed youbecome, the better and more
accurate the output is going tobe, I assume.
(31:15):
But that's an assumption thatyou have some kind of a template
or some kind of a custom GPTthat helps you write these and
actually get to all the details.
Pooja Jain (31:23):
So I am actually using Entropic console for,
okay.
Refining my prompts.
I should I demo that?
How it looks like?
Isar Meitis (31:30):
if that one, why not?
let's show everybody, let's showeverybody the what's happening
behind the magician's curtain.
Pooja Jain (31:37):
Yeah, so if you're not aware, anthropic console is
really, the, what is happeningin the background of cloud.
Okay?
Yeah.
You can simply go in there,create an account, it asks for
your credit card, but it is verycheap to run.
But the benefit here is that itis super good with writing and
refining the prompts,specifically those advanced
level prompting, like chain ofthought prompting or the tree of
(31:59):
thought prompting, which is whatyou need when you are, designing
AI agents for, complex businessprocesses, right?
So what I do is I write like abasic prompt, everything that I
want this agent to do becausethat's, of course, like ropy
cannot understand that I takethat prompt.
then there is a generate aprompt.
You simply go in there, copypaste your prompt, and what it
(32:21):
would do is.
It would generate a very, likedetailed prompt for you.
and you can even say, if youwant a chain of thought
prompting style or three ofthought, whatever you want, you
can add it here.
Isar Meitis (32:33):
Awesome.
So this
Pooja Jain (32:34):
is the hack.
Isar Meitis (32:35):
Yeah.
This is a great hack.
another thing that helps a lotwhen you're working on more
complex stuff, is I really likeusing.
Canvas in Chachi pt.
So you can do this as step one,bring it to Chachi pt, ask you
to open it as canvas.
And then as you're testing theagent and something doesn't
work, or it doesn't act exactlylike you want, you can go back
(32:56):
to the canvas, highlight thatsection and say, Hey, I'm using
this as an agent, and it's doingthis and that behavior.
How can I change this segment ofthe instructions in order to
prevent this behavior or toenhance this behavior or to add
this functionality?
And it will add it right inthere in the canvas, which then
makes it very easy to copy andpaste it back into, whatever
tool you're using.
So being able to do iterative AIassisted process, I think the
(33:20):
best tool that we have right nowis Chachi PT Canvas.
I agree with you that gettingvery detailed prompts, console
from Claude is fantastic.
Pooja Jain (33:29):
it is.
So Claude.
Any, entropic anyways, is, Doinga wonderful job in, agent AI
stuff.
They're like the way they'redeveloping them.
So they are like this improvingan existing prompt.
the way they have fine tunedtheir models is really to cater
to the agent ai.
So this is like my go-to, portalor go to website when I want to
(33:50):
design a system, a very goodsystem from.
Isar Meitis (33:53):
Awesome.
Cool.
So we were just talking aboutthe system from, for the agent,
we're about to move to tools andexplain what tools are.
Pooja Jain (34:01):
Yeah.
let's say you have this newperson joining your team.
You have explained the job, youhave explained the role.
The next step there is you givethis person the access to all
the key systems, right?
This is exactly what tools isall about.
You have to think what your AIagent needs.
So for that, the diagram that Idescribed before really helps me
out.
like when I am starting, I amalways thinking, okay, if it has
(34:23):
to get, say the sentimentanalysis, then it should go to
Trustpilot or maybe Capterra orany other of these.
Review website.
If I want to run a social mediaanalysis, then I have to give it
access to LinkedIn.
So this is the thought process.
then it needs also access to mySlack channel and Adding tool is
very simple.
All you have to do is add it,click on add a tool.
(34:45):
As you can see, they, likerelevance and all the agent
builders by now have this.
But relevance for example, hasaccess to, I don't know, tons of
tool.
All you have to do is select,add, link that to your account.
Some of these tools they arefree to use, but some of these
tools they would need your API.
So that depends really on yourprocess.
(35:06):
But yeah, you can do a lot here.
as a starter, I think it neverhappened to me so far that a
tool that I needed was not here.
So I think this is a very goodlibrary of existing tools.
and then once you select thetool, the very important step
here is you have to define whereyou want this tool.
when do you want this tool torun?
(35:28):
So simply, describe this in a,how tool is described to the
agent.
So how does your agentunderstand this tool?
and then the use case, ofcourse, like there is a
possibility here, I think can doit here.
No, maybe I'll show it later.
Okay.
I'm not able to find, but yeah,there is an option here, where I
(35:48):
can really, say, okay, can thistool run manually or does this
tool need my approval to run?
So that is something you canedit within the interface.
Isar Meitis (35:58):
Yeah.
I'll say something importanthere.
So it's really a two stepprocess, right?
Step number one is justconnecting a third party tool,
right?
This could be your CRM, yourERP, your email platform, or
something that's not from yourcompany, like doing research on
a specific platform and so on.
And then the second thing isreally explaining to the agent.
How to use the tool because thefact you have access to Slack
(36:21):
still doesn't mean that youunderstand what you need to do
once you get into Slack.
And so these are like the twodifferent layers of creating
tools for specific agents.
Now, the other thing that youneed to remember is depending on
exactly what you're trying todo, the tools are reusable.
Like you can create a tool thatconnects to Slack or email or
ERP or CRM or whatever.
(36:42):
That can be used by multipleagents, but you don't have to,
depending on how you define theinstructions, you can use it in
multiple tools or just in one orhowever you wanna build this.
You can connect to the samepiece of software with different
definitions of how to use themand then create, quote unquote
different tools for thedifferent agents to use.
Pooja Jain (37:01):
Yes, that is basically how these tool works.
so for example, if you see likethis post to Slack, I've just
given it a very simpledescription here.
So when the user approves thepost, create the comprehensive
summary and output of the agentin the Slack channel, that's it.
That's what it needs to know.
It does not have to be superlong.
So what needs to be verydetailed is system prompts.
(37:22):
After that, it has to be verycommunicative with what you want
it to do.
Isar Meitis (37:26):
Fantastic.
just to explain what I wassaying before with this example.
In this example, what this tooldoes is it creates a summary of
whatever the input was into ashort slack message, in a
specific message, in a specific,channel, right?
So you can use this acrossmultiple agents, not just j that
we use.
If what you wanted is to postthis thing on Slack, sometimes
you want an agent that willrespond to things on Slack, and
(37:48):
then you'll build a differentquote unquote tool.
It's still connected to Slack,but the definition of how to use
it will be different and it willbe a different tool within your
relevance environment.
Pooja Jain (37:58):
Yeah.
So for example, When I, look forSlack, as you can see there is
post to Slack.
I mean it sees my, but there isan ad emoji reaction.
There is even a separate toolfor that.
Yeah.
Or create a new channel, sendmessage to a Slack channel,
right?
You have different sub tools,let's say for the same Slack.
So I think Yeah, that's a greatpoint is platform you're
Isar Meitis (38:18):
connecting to.
Yeah.
For the
Pooja Jain (38:19):
same platform.
Yes, for same platform.
Based on what you want to do,delete messages, get the file so
you can do whatever in here.
Isar Meitis (38:28):
Yeah.
Pooja Jain (38:29):
Okay.
I hope this clarifies theconcept of tools.
So as you can see, yeah, thereare different tools by the way,
like it is so scalable.
So since I'm running this fordemo, I did not add a lot of
tools.
Otherwise it becomes very slow.
But let's say if you want to getthe funding news about your
competitor, you can add a toolhere that goes to Crunchbase and
(38:49):
start getting the funding news.
Same if you want to have anothersocial media.
I only added LinkedIn, but ifyou want to have another social
media where you want to trackyour competitors, you can add
that.
So it is exactly as is Idescribed, like Lego blocks keep
on adding or keep on deletingbased on your use case.
Moving on, Now knowledge thinkof it as a brain of your agent.
(39:13):
For this particular use case, Ihave not added any knowledge,
but let's say if you arebuilding a customer support AI
agent, which you want to bespecifically trained on your
business, then knowledge plays abig part because what you have
to do is make sure that it hasaccess to your business,
specific information to yourwebsite, I don't know, to your
(39:34):
vision, mission company onepagers and so on.
And as you can see, you can addknowledge in multiple formats.
So PDF website, existingknowledge.
The knowledge here is really therag.
So retrieval, augmentedgeneration, that means once you
have a query or once the userhas a query, the LLM model
(39:56):
really goes in there, finds themost relevant, answer, search
for it, or generate, retrievesit and then generates a respons
for, so that is knowledge foryour AI agent.
any questions here?
Isar Meitis (40:13):
no.
I think that's prettystraightforward.
I think people are very muchunderstanding of this from just
using LLMs, right?
It's the same concept or usingcustom GPTs or any of these
tools.
It's just adding informationthat it will make it more
specific to you versus just thegeneric universe.
Pooja Jain (40:26):
Yes.
And the last part here istriggers.
So triggers are, so for example,here I'm triggering the, or I am
starting this AI agent bychatting with it.
That is one way.
But let's say, when you think ofautomation, maybe you want your
AI agent to run every week, ormaybe you do not want to come to
the relevance AI interface, butrather trigger it via some third
(40:48):
party.
So then you can use one of thesetools and use that directly.
Maybe a Telegram channel,WhatsApp, these are premium, so
they need really high credits.
but yeah, that is a way to alsoinvoke your AI agent.
So this is, another part of it.
Isar Meitis (41:07):
Yeah, so those of you who are familiar with the
Zapier and makes of the world,it's the same concept, right?
An email comes in, it's the sameconcept, configure one agent, an
email comes in from a specifictopic, starts the agent.
A meeting is set by a specificperson, starts an agent, a
message on Slack from a specificchannel like each and every one
of these things, the tools weuse every single day.
(41:28):
let's take our example of aresearch thing.
You just created a new accounton the CRM.
As soon as it's created, it willtrigger, the agents, the
original logo.
Go do the steps, do theresearch, do the thing.
We'll create a summary for youin the CRM without you having to
do anything in between, otherthan just creating, the new
account.
So you can think about every oneof the systems you use daily as
(41:48):
both something it collaboratewith, but also as a trigger that
will actually initiate theprocess.
Pooja Jain (41:54):
Yeah, that's a very good description now.
The final part how it looksreally in the AI agent
interface, right?
as you can see here, if you'reaware of the ER and make, you
know that er, and make, you haveto really, tie down the tools or
connect the tools with oneanother, and it runs in a very
specific sequence, right?
But here I have just given itthe access to tools.
(42:16):
they run in no particular order.
for example, if my question isin a way that I want only the
Trustpilot summary, it'll onlyrun this tool.
So there is no particular order.
And another thing is, you havethe option to actually decide.
So whether this tool all runsalways, or it requires an
approval as I did for Slack, forexample.
(42:38):
Or you let the agent decide soyou can really make a
deterministic, always run this.
Or you give the autonomy to yourAI agent to decide, or you, add
human in the loop here.
Everything is doable here.
So that is, I think, a very,major difference when it comes
to, what I observe, when I workwith me or with, for example,
(42:58):
these AI agent builder,relevance.
Isar Meitis (43:02):
Fantastic.
Puja, this was an amazingoverview for beginners.
I think we covered reallyeverything people need to know
to get started.
Uh, Renee on LinkedIn literallysaid, I just signed up for a
test account.
I'm gonna start playing withthis.
So you got at least one personexcited enough, and less scared
of building agents to actuallytake action.
That is wonderful.
(43:23):
so it's fantastic and I'm suremore people will do it.
There was a question earlierthat I wanted to wait for the
end to answer, but I think It isinteresting.
There are other tools out therelike relevance, right?
So I'm trying to see what theymentioned, like Mind Studio
Lindy, and there's A bunch ofothers.
I don't know if you just doveinto relevance and that's like
your universe.
No.
Or you play with some of theothers And do you know, do you
(43:44):
have a reason why or preferenceswhy use relevance versus the
others or not?
Pooja Jain (43:49):
No.
No.
So I am an AI trainer.
I play with all of them.
Okay.
So why IT tool stack reallyinvolves relevance, Lindy and
eight n and now also Zapieragents, for example.
Isar Meitis (44:01):
Yeah,
Pooja Jain (44:02):
why I chose relevance for this particular
use case was because there arefew benefits of relevance.
They are cheap, so it's a$20subscription every month and you
get 10,000 credits, which isgood enough to run such kind of
use cases.
Whereas when it compares toLindy or Zapier, that is very
complex, that is very, expensiveto run.
(44:22):
So they consume a lot ofcredits, and I think their
monthly subscription is alsohigher than this.
That is one Second is comparedto NA 10, NN is very popular
right now, but I personally canbuild a lot of NA 10 workflows.
But NA 10 is still a bit moretechnical.
So when I, when it comes to
Isar Meitis (44:39):
way more tech, way more
Pooja Jain (44:40):
technical, yeah, it is way more technical.
So when it comes to training thenon-technical, audience, they
want something that they can seeand intuitively understand.
So this is why I find relevanceto be slightly better, for
non-technical folks.
But once you get used to of,these AI agent concept, I mean
you can use whatever, and itthen is very, efficient in terms
(45:00):
of price, in terms of creditconsumption.
You can literally run n it andfor four,$4 per month.
So that's the benefit of it.
Isar Meitis (45:09):
one small thing about NA 10, since you mentioned
that NA 10 is an open sourcetool, so you can pay them to run
it on their platform, which isstill cheap, but for very
little, you can host it on yourown and then You can run it
unlimited.
Amount, maybe not at the highestspeed, but speed is not a big
deal here.
Like even with what we just did,most of the agents, the speed is
not critical.
(45:30):
So you can still use arelatively cheap hosting plan,
four, five,$6 a month and run asmany agents as you want.
And yes, instead of getting ananswer in five seconds, it'll
take 10, 20, 40.
Who the hell cares?
Like it does the work for me.
I don't have to do it.
and it costs me four bucks amonth.
So there's benefits in runningNA 10, but I agree with Puja a
hundred percent.
It's more technical and not assimple to use as relevance as an
(45:52):
example.
yeah.
Puja, if people want to knowmore about you, work with you,
learn from you, I know you'relaunching a course.
what are the best ways toconnect with you and do more
with you based on the amazinginformation that you have?
Pooja Jain (46:05):
Sure.
So if I would love to connectwith you guys on LinkedIn.
I'm super active there.
I always share some practicaltips on AI agents and business
use cases.
Please connect with me.
I am launching a course nextweek.
It is called Six Week AI RevenueAccelerator, really focused on
small and medium businesses andwhat we are going to build in
these six weeks, starting fromthe basics to really building AI
(46:28):
agents for content creation,lead generations SalesOps and,
customer success, and personalproductivity.
Oh, wow.
So end-to-end course.
and I'm running it, inpartnership with Valeria.
I believe some of you who hasbeen on this podcast
Isar Meitis (46:42):
as well, so previous guest, fantastic.
Puja.
Again, thank you so much.
This was absolutely amazing.
Thanks everybody who joined us.
We had multiple people on theZoom and on LinkedIn, and I
think this is the most activechat I've seen in a long time on
both platforms.
So I didn't ask you all thequestions.
I chatted with a lot of them andjust answer them, but great
participation.
Pooja Jain (47:02):
Can I go and check the questions?
Maybe I can answer some of them,or, do they after?
No, I, all
Isar Meitis (47:06):
the stuff that I knew how to answer quickly, I
didn't answer, but you can alsoanswer them, right?
they stay on LinkedIn, so youcan just go there and answer the
questions.
I'll try my best.
yeah.
so thanks everybody for beingwith us on the live.
If you're not here, you shouldjoin us.
Like we do this every week,every Thursday.
And at noon Eastern.
I will remind you, that, ourcourse also starts on Monday.
So you have two courses to pickfrom and to be completely
(47:26):
honest, it doesn't matter whichone you pick, but take a course,
accelerate your AI knowledge soyou can do more with AI in your
business and for your owncareer.
So pick a course and go do it.
It makes a very big differencein your ability to actually do
the things that you need to doin a much more effective way.
and that's it.
Thanks everybody for joining us.
Thank you again, Puja.
Have an awesome thank you.
Rest of your day, everyone.
Pooja Jain (47:47):
Thank you so much for having me.
Bye-bye.
Hello, and welcome to another live episode of the
(00:04):
Leveraging AI Podcast, thepodcast that shares practical,
ethical ways to leverage AI toimprove efficiency.
Go your business and advanceyour career.
This Isar Metis, your host, andwe have.
Maybe the top exciting topic ofAI in 2025, which is agents,
which is going to talk abouttoday, and agents are already
here and I know a lot of peopleare hearing about this.
(00:27):
It's something that is comingand it's the next stage, and
it's the evolvement of LLMs, butagents are already here and
multiple people and companiesare already generating
significant business value byleveraging.
Agents across multiple aspectsof the business.
But for most people and mostcompanies, there's still this
elusive concept that peopledon't understand and definitely
(00:50):
do not know how to implement.
Now, this gap between the peopleand the companies who have
agents to those who don't isspreading every single day, and
it's widening.
And you want you, yourself, as aperson with capabilities as well
as your companies to be on theright side of that gap.
(01:11):
Meaning you want to be on thatfast bullet train, that
accelerating and providing moreand more capabilities to your
business versus being standingon.
The train station and watchingthat train getting further and
further away from you,especially if you know that your
competitors might be riding thattrain, meaning they'll be able
to do more and more than you canfor less money and then be a lot
(01:35):
more competitive than you cannow, developing agents sounds
like complex, and it sounds verytechnical and it sounds like
only big companies likeMicrosoft with a lot of people
who write code can actuallycreate them.
But the reality is there aremultiple tools out there today
that allow you to write reallypowerful agents with either no
code or low code.
(01:55):
Today we're gonna focus on nocode at all.
So one of the most powerfultools out there that enable to
do that is called relevance ai.
And our guest today, Puja Jain,is a relevance ai.
Expert and she has beendeveloping agents for multiple
companies on relevance for awhile now across different
aspects and different businessesand so on.
(02:17):
Now, in addition to the factthat she knows relevance really,
really well, she spent years indeveloping and implementing AI
solutions at Proctor and Gamble.
So in addition to her recentexperience and just building
them on relevance, she hasenterprise level experience in
understanding what is requiredto actually develop agents that
actually work in a businessenvironment.
(02:38):
And she knows how to do it in astep-by-step in a tool that
anybody can use, which makes herliterally the perfect guest to
talk about, this topic with us.
So this is exactly what we'regoing to do today.
We're going to show you anentire process, beginning to
end, how to develop an AI agent,what to think about and how to
use relevance in order to dothat.
But having that knowledge andseeing this will allow you to
(03:00):
develop other agents and onother tools because the concepts
are exactly the same.
Now, as I mentioned, sinceagents are maybe the most
transformational technology weever created, at least that I
know of so far, then.
This is a really important topicand hence I'm really, really
excited to welcome Puja to theshow.
(03:21):
Puja, welcome to Leveraging ai.
Pooja Jain (04:06):
Thank you so much, Isar, how are you?
Isar Meitis (04:09):
I'm doing awesome.
I'm really, really excitedabout, this session.
I myself am tinkering withdifferent tools.
I'm definitely not close to yourlevel, in doing this, so I'm
really excited.
I'm sure a lot of people arereally excited as well.
Our top performing episodes inthe past few months are all been
around agent development, so I'msure there's a lot of people who
are really curious, about thistopic.
It's everywhere in the news ifyou're following an ai, but
(04:31):
still, I think most people nowget a feel on how to use large
language models and imagegeneration.
I think 99% of the people don'thave a clue how agents work and
how to create them.
And so I think this is gonna bereally fascinating.
before we get started, a few,messages.
First of all, thank you toeverybody who are joining us
live.
Whether you're joining us onLinkedIn or joining us on Zoom,
we really appreciate you beinghere.
(04:52):
I know all of you have otherstuff to do on, Thursday at
noon, Eastern time.
But, feel free to introduceyourself.
Tell us what you know, what youdon't know about agents, and
what you wanna know from thissession as well.
That will help me, guide theconversation, maybe even more.
Tell us where you're from so weknow where people are joining
from.
That's always fun for me to see.
There's always people from allover the world, at least on
this, on the, LinkedIn side ofthings.
(05:13):
There's always a lot of peoplefrom, interesting places.
So introduce yourself, make newfriends.
In addition, one last thing thatI will say, is that if you are
still watching us live, the nextcohort of the AI Business
Transformation course startsthis coming Monday.
So when this becomes a podcast,if you're just listening to this
after the fact, you missed thatparticular cohort, but the
(05:33):
cohort starts on May 12th,meaning you still have a few
days.
This is this coming Monday, thiscourse.
Is really transformational forpeople, for companies, for
careers, for complete teams, andwe've been teaching it.
I've been teaching it for thepast two years every.
Single month, sometimes twice amonth.
So hundreds or maybe thousandsof business people have been
through it and have reallychanged their businesses with AI
(05:56):
based on the information thatthey've learned.
So if you have not startedproper implementation of AI in
your business, this is anincredible opportunity to get a
huge amount of information, bothon the tactical side of what
tools to use, as well as on thestrategic side, on what's the
right process to implement AIcompany wide.
If this is interesting to you,I'll drop the link in the chat.
(06:18):
If you're listening to thisafter the fact, then there's
gonna be a link in the shownotes to tell you when the next
cohort is, which will probablybe in August.
because we in between, weusually teach private courses
and we're usually fully booked.
that's it.
With all of that being said, Iwill give puja the microphone.
I'm really excited to see whatyou have prepared for us and,
get ready everybody to learnhow.
(06:39):
Simple.
It actually is to create agentsthat are extremely powerful.
Pooja Jain (06:43):
Thank you so much, IAR for having me here today.
And yeah, super excited to sharewhat I am working on and what I
have learned about agents sofar.
Honestly, speaking agent AI isevolving.
It's evolving on a weekly basisnow.
We are seeing something newevery week coming up, so it is
even new for me, although I havebeen working in the automation
space for many years now.
(07:03):
But yeah, agent AI is somethingnew.
And today, what I'm going toshow you is first like
clarifying a bit.
So what are AI agents actually,and how do they differ from say,
an automation workflow becausethat's where I think most of the
people are still strugglingwith.
And secondly, showing a livedemo in relevance AI about this
market research AI agent that Ihave built, and what is the
(07:26):
concept around it?
Should we get started?
Isar Meitis (07:29):
Yeah, let's dive right in.
I'm, I really wanna learn aswell.
Pooja Jain (07:33):
Perfect.
Let me share my screen quickly.
Isar Meitis (07:36):
Sure.
for those of you who are,listening to this as a podcast,
and I know thousands of you are,we will explain everything that
we're seeing on the screen soyou're not missing anything.
But if you do wanna watch whatwe're doing, you can either
switch to watch this on YouTube.
So go to our YouTube channel,then multiply AI YouTube
channel.
There's a link in your shownotes, or if you are driving
right now or mowing the yard orat the gym wherever you are,
(07:58):
when you cannot do this, thenyou can keep on listening to us
and then you can decide later onif you also want to watch the
YouTube video.
Pooja Jain (08:05):
Great.
Okay.
So firstly, what exactly is anAI agent?
of course, like by now almosteveryone is aware of what is
Chad GPT?
So Where do agents fit in here?
So imagine when Chad G PT comesin, even now, to get something
beneficial out of Chad G pt.
You have to give in a properprompt.
You have to add in your businessknowledge, and this becomes a
(08:27):
repetitive process, right?
And Chad g PT is still a chatbot.
So it only talks, it does notexecute.
And this is why AI agents arereally changing, the landscape
for the businesses because theycan execute.
So if you want to understand AIagents in very simple terms,
think of them as let's say ifchat GPT gets knowledge, so your
(08:49):
business specific knowledge, itgets a memory.
I mean, a chat g PT has amemory, but really, you know,
something that is a memoryspecific to your business, to
your style of working as well astwo hands that it can start
executing.
Actions in your systems.
Now, it can be something assimple as writing an email or,
(09:09):
managing your calendar, whichcan be chaotic for a lot of
people.
it's a big productivity hack IfJGBT can start managing your
calendar or something as complexas, doing a proper research and
then going into your CRM andupdating the fields there.
So this is what, AI agents areall about, like they can
(09:29):
understand, of course, becausethe underlying power is still
LLMs, but in addition to that,they can also execute actions in
the systems in your business,ecosystem.
Isar Meitis (09:40):
Yeah.
I'll add, two, two more things.
like when I think on agentsversus large language models.
So you mentioned, one veryimportant thing, which is the
ability to take action.
The two other things that agentsdo that LLMs don't do, one is
make their own decisions.
So they're more autonomous.
They don't have to be, by theway, but they can be more
autonomous than just a largelanguage model, meaning you can
(10:01):
give them a broader task andthey will figure out the steps
on their own without you havingto define an exact path on how
to get there.
And the other one is.
If you build it correctly, theycan be multi-layered, meaning
you can have an organization ofagents working together, like we
have teams of humans.
You have a manager, you havesomebody writing content, you
have an editor, you have anevaluator, you have a designer,
or a similar parallel in anyother part of the company, the
(10:24):
same thing.
You can build agents and thenthey can work collaboratively
versus just working on theirown, which is what LLMs do.
So these are the maindifferences.
Pooja Jain (10:31):
Very good point.
ISAR because, and this is alsoone of the major differentiators
when you start thinking aboutagents from workflow automation.
so workflow automations arereally like, very static.
So let's say if this happens,then they should happen.
There is no autonomous behaviorthere.
you have to predefine everythingand they execute, which is
fantastic for a lot of businessprocesses because not every
(10:53):
business process needs, thatlevel of autonomous behavior.
However, AI agents, they canselect their own tool.
They can select as I mentioned,they can select their tools.
they can select the next stepand so on.
So where do you really need AIagents?
So AI agents make sense whenyour process is super complex,
involves multiple steps andmultiple level of decision
(11:13):
making.
If this is the case, then youshould start thinking about AI
agents.
Otherwise, if your process ismore about, say, repetitive
steps, that more or less staythe same.
You are good to go with aworkflow automation.
Isar Meitis (11:26):
Yeah.
I always tell people now startedcalling, Zapier and make like AI
agents are like, no.
If it's doing just a step bystep process, you don't need AI
in it or you can bring AI in itinto specific steps if you're
trying to analyze what surgeonin an email as an example.
But the rest is just old schoolautomation that existed for a
decade now.
Pooja Jain (11:43):
Yeah, absolutely.
Oh, by the way, Zapier has alsointroduced their AI agent
builder, which is a no-code aswell.
I think that most of theseworkflow automation builders are
now also moving towards agentai.
Yeah.
Great.
And then second point, that is,I just mentioned about this AI
agency, so that is exactly whatI would be showing today.
(12:04):
So my agent is about a marketresearch AI agent, or rather a
team of AI agents.
And when you start thinkingabout this.
you have to really think of itlike a team of, humans that are
running this.
So let's think of it, there is amanager agent, and manager agent
is the one that is coordinatingwith all the subagents in this
(12:25):
team.
I, as a human, interacts onlywith the manager agent.
So I only give my, request or mycommands or whatever I want to
do, only to the manager, agent,manager.
Agent then, picks like the nextsubagent.
for example, in this specificuse case, so I build this for
competitor research, there areseveral subagents, like industry
(12:45):
analyst, competitor tracker,customer sentiment analyzer,
social media auditor, and then areporting agent that is really
gathering information from allthese subagents, synthesizing it
and sharing it with the manager.
It If you think about it, it'sexactly like how you would work
in a team of humans.
And that is actually the basicsof even when you start setting
(13:08):
these up in any platform.
So manager agent is really whoyou should be telling in detail.
think of it like a jobdescription or a very detailed
SOP that is literally yourprompt that manager agent should
understand and then the manageragent in return, from based on
this prompt, should be able toexecute these, or rather pick
(13:30):
the right subagent for yourtask.
We will see this in a demo.
I think that will make it, moreclear.
Isar Meitis (13:36):
Yeah.
Pooja Jain (13:36):
anything, to add here?
Isha.
Isar Meitis (13:38):
No, I think it's great.
I really think what you said isthe important thing, think about
what humans you would need andeven make it more granular,
because in humans, in manycases, there's one human that
does several of this.
So think about it more on thetask level rather than on the
human participant level.
And because every agent will bedoing one task that will enable
it to be very good at thatparticular task.
(13:59):
And then the manager agent, andthen you can also add layers,
like a improver, like somebodythat reviews the work and adds
Yeah.
Comments and so on.
there's different layers you canadd, but still each agent will
be focused on one task andthere's will be one or more
agents that will help tocoordinate the task to make it
more effective this time around.
As well as moving forward.
Pooja Jain (14:20):
Yeah, that's a great explanation.
I agree to that because when youlook at it, I think in an
organization you would not see ateam that where one person is
just focusing on industryanalysis.
Isar Meitis (14:29):
Yes.
So
Pooja Jain (14:30):
it can be equated to, I don't know, a very well
funded startup where you have aone person doing one task.
Isar Meitis (14:37):
Yeah.
Pooja Jain (14:38):
So now let's go into relevance ai and I will start
describing you how relevance AIis, but I'm hoping that there's
an interesting
Isar Meitis (14:47):
question, which I know the answer and we're gonna
actually demo that, but it's agood question to ask.
The question is, is the humansets up the subagents and not
the manager agent spins them upin real time?
And the answer is yes.
and I think it's a yes, but, soI will let puja answer the rest
of it.
Pooja Jain (15:01):
Yes.
I'll be showing that during thedemo, and if it is not answer,
happy to take that at the end.
Isar Meitis (15:06):
Okay.
Perfect.
Pooja Jain (15:08):
Okay.
So what you are seeing on mysystem is really the relevance
AI interface.
It looks a bit messy becauseI've not really organized my
workspace here.
but relevance AI is a no code,AI agent builder.
And if you go into relevance ai,they have now terms and terms of
templates.
So when you start going, likeyou can create a free account if
(15:29):
you want to get started, andthey have a lot of agent
templates, which are fantasticway to start.
So if you're just starting out,I think the best way would be
take one of these templates andstart editing it for your
requirement.
That is the easiest, right?
Then there are a lot of tools,so these are really the
integrations that are alreadypresent in relevance.
(15:49):
So for example, if you want toextract something from LinkedIn,
they already have a tool for it.
All you have to do is select it,use it in your agent or in your
project, and adapt the prompts.
Now let's go into this agentthat I am working on, which is
the co research agent.
(16:09):
Okay, so firstly, let's see,maybe I will first run this demo
so we see the output and then Iwill start explaining it one by
one.
So this competitor, analystagent, as I explained, is a team
of subagents that is doingindustry analysis competitor.
So it can go onto yourcompetitor's website, track the
latest updates.
(16:30):
It then goes into, sayTrustpilot or any other review
website that you have givenaccess to, gets the latest
reviews.
I have added an additional layerthere.
So not just get the reviews,because let's say if the
competitor is doing well insomething, I want to understand
why or if there are any gaps.
I also want to understand why.
So it has an additional layer ofsentiment analysis, tracking.
(16:54):
I have also trained it to goonto the LinkedIn or any other
social media, for example, togather what is happening at the
social media of the competitor.
And finally, it synthesizeseverything and sends me over on
Slack.
So let's give it a demo.
So let's say I just give, use ademo use case.
Okay.
I'm a sales head for, AIpowered, CRM for SMPs.
(17:18):
So I'm simply describing what mycompany's doing in very simple
terms.
Nothing like, we do not need apage long prompt for this.
And I would like to.
Get the competitor insights, sayon Pipedrive, because Pipedrive
is like the CRM, which is very,like a competitor for me, right?
So I would just simply run that,okay.
(17:41):
As you can see, I've simply,talked with this manager agent,
right?
So what you saw on my window wasthe manager agent.
I simply, communicated what Iwant to do and it has started
working in the background.
Now, first thing is, it isextracting the industry news for
me.
so this is another subagent thatis running in the background.
Now let's open another window soI can show you what is
(18:05):
happening, in the background.
Okay, so I just, ran this likebefore the session because
sometimes, it is extracting alot of information.
Yeah.
It'll take time from the web.
It takes time.
So let's see what it did.
It extracted the industry news,which is specific to pipe drive
and which is very specific tothe AI initiatives of Pipedrive.
(18:27):
Because I specifically saidthat, I am doing this AI powered
CRM.
Isar Meitis (18:31):
Yeah.
Pooja Jain (18:32):
So it just extracted everything along with, the
source.
Then, so those of you who arejust listening,
Isar Meitis (18:38):
there are multiple articles with each and every one
with several bullet points andtopics and summary, and then
there's a final analysis of allof them together.
Pooja Jain (18:47):
Yes.
So this is one agent that didits job, second agent, which is
the Trustpilot Reviewsummarizer.
What it did, it went on toTrustpilot, it got the latest
rating of Pipedrive.
And based on that, it gave melike the key positive points.
what is happening there?
key pain points, so for example,it find out that based on the
(19:08):
reviews that there is arecurring issue, in the
limitation of the form ofcomponents, particularly
regarding constant tracking forsubscriptions and so on, pricing
transparency and the learningcurve.
So these are already very good,insights for me.
Now, let's go further.
It then did a competitor,product benchmarking.
So what it did, it went to pipedrive's, website extracted, all
(19:31):
the latest CRM features thatpipe drive release because that
is what is relevant to me.
It summarized that, and finally,it summarized everything like,
synthesized all this informationfrom different subagents,
summarized it and gave me like areport here.
for example, the strengths, whatare the pain points, current
capabilities, and so on.
(19:52):
Now, one of the questions that Ioften get here is, how is it
different from, a workflowautomation, right?
So one thing here is, let's say,if had this been in a workflow
automation, but if I want torerun this process, it would
always run everything from startto end.
Whereas now, let's say if I amnot happy with the, Trustpilot,
(20:15):
tracking, for example, I cansimply say, I would like to
rerun or redo the, sentimenttracking.
And what would happen here isbecause the manager agent here
is responsible for choosing ordelegating to the subagents, it
would identify that only thatparticular subagent needs to be
(20:36):
reactivated now.
So it is only using theTrustpilot review, summarizer
subagent and not rerunning thewhole process.
And that basically is the bigdifference when it comes to
workflow automation because hadthis been like a, Zapier flow or
make flow, it would have rerunthe whole thing and not just one
particular component of it.
Isar Meitis (20:57):
Yeah.
And this connects beautifully toa question that was in the chat
that said, That so far, likewhen you just started running
this, it looks similar to deepresearch.
How is that different?
And I think you answered somecomponents, but I wanna dive
deeper into that because I thinkit's important for people to
understand.
First of all, deep research isan agent.
So the biggest differencebetween deep research and just
(21:18):
using chacha PT research searchis exactly that concept, that it
understands your question andthen it says, oh, the person
wants to understand this topic.
What do I need to do?
Which is now an autonomous thingto actually give him the
information that he needs.
The biggest difference betweendeep research and this, and
there are several differentdifferences.
Difference number one is you canin advance define all these
(21:39):
subagents and you will knowexactly what it will do.
Meaning in deep research, youcan't control what's.
Sources it will go to which one,it's not gonna go to what topics
you're interested in and so on.
So when you build an agent for aspecific topic, like in this
particular case doingcompetitive research, you can
build it the way you want it.
So think about it like acustomized version of deep
(21:59):
research.
That's basically what it'sdoing.
So that's difference number one.
Difference number two is, asPuja said, you can go back and
forth with specific componentsof this because they're
standalone agents, which is notpossible, or it's possible to an
extent, with deep research, butit will still not be as tailored
in specific as the specificagent that you develop.
(22:22):
And number three, which whatyou're gonna see in a minute,
this will actually go and updateyour CRM and do other things
that obviously deep researchwill not do.
So think about this as a.
Multi-layered, multi-level, morecustomized version of deep
research that can also then goand do stuff like write it to
you in Slack or update your CRM.
(22:42):
Anything you want to add, Puja,because I think it's an
important topic for people tounderstand.
Pooja Jain (22:45):
Yeah, I think that summarize it beautifully.
another point, human in theloop.
Deep research does not havehuman in the loop, right?
You have no control.
You cannot really say, okay,there is no concept of asking
for permission.
I will show you in a second thatas a custom AI agent, you can
really, train it to ask forpermissions in certain areas
(23:09):
like where you want it to, notrun on its own.
Awesome.
okay.
Yes.
Isar Meitis (23:14):
Another question that is also I think very
important and then we cancontinue is, there was a
question about hallucinations.
Is there a way to know or toreduce or to verify the
information that is coming fromthese research agents?
Pooja Jain (23:28):
it is possible to, of course, like you can train
your AI agent to ask for thesources.
So for example, when I wasbuilding this agent, I was
really not sure if it is gettingthe right information from the
trustpilot.
So I was checking it again andagain while building it.
I was always verifying, forexample, the stars that it shows
me here are correct, the reviewsare correct or not.
(23:49):
And the date, because I wantedto only go one month, get the
reviews one month older only.
So that is something I wasalways manually checking.
I think this is a very importantstep.
Like you have to be verycautious when you are designing
your AI agent again, human inthe loop.
I think this is one of the mostor the ultimate guardrails that
can be part of your AI agents.
(24:11):
Second, the system prompts,they're the best ways to control
your AI agents.
I will show you in a secondwhere you can add your system
prompts.
So whenever you are designing anAI agent system, prompts are
really the controllers.
That is the information and, thekind of, positive prompts as
well as negative prompts thatyou add in your ai, system
(24:32):
prompt.
They are the ones that would becontrolling everything for you.
Isar Meitis (24:36):
Awesome.
So I'll add my 2 cents.
Yes.
When I do deep research on stuffI really care about and I really
need to verify the information,I usually run it on three
different AI deep researchtools.
So think about running threedifferent agents to do the
research for you.
And then I have a fourth processthat actually creates a
comparison table and checks.
If all three have the sameinformation, when all three have
(24:58):
the same information, then it'svery unlikely they all made up
the same stuff.
And so that's most likelyaccurate information.
And when there are outliers,when only one of them finds a
piece of information, the samefourth agent go and checks that
information is correct.
So again, now I get a secondverification if the information
is really there from thatparticular source and all the
(25:19):
outliers are not found, are justbeing thrown to the trash.
And I get a summary of that.
So it just depends on how.
Buttoned up, you need thisinformation to be right.
let's say there's 500 Trustpilotreviews.
If 10 of them are made up, 15%are made up, still not a big
deal, you're still getting the500.
But if you're looking onsomething that you're gonna make
a very important businessdecision on, you want it to be a
(25:40):
hundred percent accurate and not90% accurate, and then you can
add these additional layers andsteps in order to dramatically
increase the chances that theinformation is correct.
Pooja Jain (25:49):
Great point.
Great.
So let me show you the conceptof human in the loop, which is
absent on, definitely on theworkflow automation because
there is no way that you can addlike a, Human, approval step in
there, or even in the LLMchatbots, you give them a
question and they answer thereis no place to even ask for
approval.
(26:09):
But here it is the case.
So let's say in this demo thatI'm running, I am happy with the
output that I have got so far.
Now I want a summary of it, andI want this on my Slack channel.
I have said this AI agent in away that it does not spam my
Slack channel with everything.
It always asks for approval.
(26:30):
And only when I'm happy with theoutput.
let's see.
Okay.
I should not be writing please,but I'm used to.
I do this,
Isar Meitis (26:39):
I do the same thing.
By the way, the jury's still outon that.
So those of you who don't knowthis whole discussion, there was
a big debate last week, based ona post from several different
people, including Sam Altman, onthe cost of saying, please, and
thank you to these chatbots.
And there is inconsistency inthe knowledge whether that
actually makes the resultsbetter.
I think the best analysis thatI've seen, came out from, brain
(27:04):
Dead.
it'll come back to me in aminute, but he basically said it
depends on the use case.
Sometimes it helps and sometimesit doesn't.
uhhuh, but I still do thisbecause that's how I'm used to
typing.
And so I'm just like you.
Pooja Jain (27:15):
is this a report from Ethan Molik?
Isar Meitis (27:16):
Ethan.
Ethan Molik.
Yes.
Thank you.
Ethan Molik.
Pooja Jain (27:18):
yeah.
I read it as well.
It's interesting that there area few things we cannot get rid
of, which is nice as well.
yeah, that's fine.
They're going to be ourteammates, right?
Isar Meitis (27:27):
Yeah.
Pooja Jain (27:28):
Okay, so I simply asked that, now I'm happy with
the output.
Now, please send me a summary onthe Slack channel and as you can
see it called the Slack channel.
so it called the relevantaccount, from the Slack channel.
Created a summary, a very nicesummary of the key trends,
vulnerabilities, pricing,transparency, and so on and so
forth.
(27:48):
And now, but it did not run onits own because this is, like I
have said this pro stepspecifically for human approval.
So only once when I click onapprove, it would go to my Slack
channel and send all thisinformation.
Otherwise it would not.
These two steps.
this single step actuallyprobably should clarify for most
people what is human in the loopfirst thing and second, how an
(28:10):
AI agent differs from, say l andm chat bots, deep research or,
even a workflow automation.
So I would not say approve,because my Slack channel is full
right now.
I don't wanna open that, but Ihope the concept is clear.
Isar Meitis (28:24):
yeah.
Pooja Jain (28:26):
Wonderful.
I think what will beinteresting,
Isar Meitis (28:28):
is to dive into what are the instructions that
make these agents do it, likethe actual creation of the
agents.
I think that's gonna be, yeah.
I think the output are nowhopefully clear to people.
So let's dive a step deeper andshow how they're actually built.
Pooja Jain (28:41):
Yeah.
So let's go into the build.
Great.
So first let me Demo a bit aboutthe interface that you are
seeing here, and then I will godeeper into it.
So what you are seeing right nowis really the AI agent a as it
should look, and theinstructions here, as you can
see, they are like super longinstructions.
These are the system prompts andwhenever you are designing an AI
(29:03):
agent system.
Prompts are the key.
Like they are the ones that arecontrolling everything.
As I mentioned, the systemprompts are the controller.
So make sure first, in yoursystem prompts, for example, to
mention what would be thefunction of this particular AI
agent.
when I explain it to myaudience, I always say, think of
(29:23):
it like a very well defined jobdescription.
So you have to define what youragent should be doing.
What tools does it have accessto?
So I already defined, you haveaccess to industry news.
So basically telling my AIagent, so as an onboarding
process, who all are there inyour team and when you should be
calling them.
(29:44):
That comes later.
But what are your core functionor your core to-dos?
when you go on a website, whatyou should be looking for, then
what you should do and what youshould not do.
So this, what you should not dois super important.
So for example, do not assumeanything.
ask if something is not clear ormissing, because a lot of times
(30:06):
what I see is, let's say when Idid not add this, negative
prompt, it was always assuming alot of information.
Which is something I do not wantwhen I am running it for a
business process.
So now adding this negativeprompt made such a huge
difference.
So this manager agent is nowcoming back to me asking me what
is missing?
And you can add a lot ofguardrails in there.
(30:27):
So for example, if you aredesigning a customer chatbot, a
chatbot that is talking to yourcustomers, you can define
specifically like a relevanceclassifier.
So to classify if whatever thecustomers are asking this
chatbot are actually relevant,and, they're not asking for a
piece of recipe to, to your, Idon't know, CRM chat bot,
because that happens.
(30:47):
you should be defining all thishere.
Second the tools.
So really what, so just beforewe
Isar Meitis (30:53):
jump to the tools I want Yes.
I wanna go back because I thinkit's very important what you
said, Just think about definingin simple English, but in a very
well structured way, all thethings that you want the agents
to do and all the things you donot want the agent to do.
And so the more detailed youbecome, the better and more
accurate the output is going tobe, I assume.
(31:15):
But that's an assumption thatyou have some kind of a template
or some kind of a custom GPTthat helps you write these and
actually get to all the details.
Pooja Jain (31:23):
So I am actually using Entropic console for,
okay.
Refining my prompts.
I should I demo that?
How it looks like?
Isar Meitis (31:30):
if that one, why not?
let's show everybody, let's showeverybody the what's happening
behind the magician's curtain.
Pooja Jain (31:37):
Yeah, so if you're not aware, anthropic console is
really, the, what is happeningin the background of cloud.
Okay?
Yeah.
You can simply go in there,create an account, it asks for
your credit card, but it is verycheap to run.
But the benefit here is that itis super good with writing and
refining the prompts,specifically those advanced
level prompting, like chain ofthought prompting or the tree of
(31:59):
thought prompting, which is whatyou need when you are, designing
AI agents for, complex businessprocesses, right?
So what I do is I write like abasic prompt, everything that I
want this agent to do becausethat's, of course, like ropy
cannot understand that I takethat prompt.
then there is a generate aprompt.
You simply go in there, copypaste your prompt, and what it
(32:21):
would do is.
It would generate a very, likedetailed prompt for you.
and you can even say, if youwant a chain of thought
prompting style or three ofthought, whatever you want, you
can add it here.
Isar Meitis (32:33):
Awesome.
So this
Pooja Jain (32:34):
is the hack.
Isar Meitis (32:35):
Yeah.
This is a great hack.
another thing that helps a lotwhen you're working on more
complex stuff, is I really likeusing.
Canvas in Chachi pt.
So you can do this as step one,bring it to Chachi pt, ask you
to open it as canvas.
And then as you're testing theagent and something doesn't
work, or it doesn't act exactlylike you want, you can go back
(32:56):
to the canvas, highlight thatsection and say, Hey, I'm using
this as an agent, and it's doingthis and that behavior.
How can I change this segment ofthe instructions in order to
prevent this behavior or toenhance this behavior or to add
this functionality?
And it will add it right inthere in the canvas, which then
makes it very easy to copy andpaste it back into, whatever
tool you're using.
So being able to do iterative AIassisted process, I think the
(33:20):
best tool that we have right nowis Chachi PT Canvas.
I agree with you that gettingvery detailed prompts, console
from Claude is fantastic.
Pooja Jain (33:29):
it is.
So Claude.
Any, entropic anyways, is, Doinga wonderful job in, agent AI
stuff.
They're like the way they'redeveloping them.
So they are like this improvingan existing prompt.
the way they have fine tunedtheir models is really to cater
to the agent ai.
So this is like my go-to, portalor go to website when I want to
(33:50):
design a system, a very goodsystem from.
Isar Meitis (33:53):
Awesome.
Cool.
So we were just talking aboutthe system from, for the agent,
we're about to move to tools andexplain what tools are.
Pooja Jain (34:01):
Yeah.
let's say you have this newperson joining your team.
You have explained the job, youhave explained the role.
The next step there is you givethis person the access to all
the key systems, right?
This is exactly what tools isall about.
You have to think what your AIagent needs.
So for that, the diagram that Idescribed before really helps me
out.
like when I am starting, I amalways thinking, okay, if it has
(34:23):
to get, say the sentimentanalysis, then it should go to
Trustpilot or maybe Capterra orany other of these.
Review website.
If I want to run a social mediaanalysis, then I have to give it
access to LinkedIn.
So this is the thought process.
then it needs also access to mySlack channel and Adding tool is
very simple.
All you have to do is add it,click on add a tool.
(34:45):
As you can see, they, likerelevance and all the agent
builders by now have this.
But relevance for example, hasaccess to, I don't know, tons of
tool.
All you have to do is select,add, link that to your account.
Some of these tools they arefree to use, but some of these
tools they would need your API.
So that depends really on yourprocess.
(35:06):
But yeah, you can do a lot here.
as a starter, I think it neverhappened to me so far that a
tool that I needed was not here.
So I think this is a very goodlibrary of existing tools.
and then once you select thetool, the very important step
here is you have to define whereyou want this tool.
when do you want this tool torun?
(35:28):
So simply, describe this in a,how tool is described to the
agent.
So how does your agentunderstand this tool?
and then the use case, ofcourse, like there is a
possibility here, I think can doit here.
No, maybe I'll show it later.
Okay.
I'm not able to find, but yeah,there is an option here, where I
(35:48):
can really, say, okay, can thistool run manually or does this
tool need my approval to run?
So that is something you canedit within the interface.
Isar Meitis (35:58):
Yeah.
I'll say something importanthere.
So it's really a two stepprocess, right?
Step number one is justconnecting a third party tool,
right?
This could be your CRM, yourERP, your email platform, or
something that's not from yourcompany, like doing research on
a specific platform and so on.
And then the second thing isreally explaining to the agent.
How to use the tool because thefact you have access to Slack
(36:21):
still doesn't mean that youunderstand what you need to do
once you get into Slack.
And so these are like the twodifferent layers of creating
tools for specific agents.
Now, the other thing that youneed to remember is depending on
exactly what you're trying todo, the tools are reusable.
Like you can create a tool thatconnects to Slack or email or
ERP or CRM or whatever.
(36:42):
That can be used by multipleagents, but you don't have to,
depending on how you define theinstructions, you can use it in
multiple tools or just in one orhowever you wanna build this.
You can connect to the samepiece of software with different
definitions of how to use themand then create, quote unquote
different tools for thedifferent agents to use.
Pooja Jain (37:01):
Yes, that is basically how these tool works.
so for example, if you see likethis post to Slack, I've just
given it a very simpledescription here.
So when the user approves thepost, create the comprehensive
summary and output of the agentin the Slack channel, that's it.
That's what it needs to know.
It does not have to be superlong.
So what needs to be verydetailed is system prompts.
(37:22):
After that, it has to be verycommunicative with what you want
it to do.
Isar Meitis (37:26):
Fantastic.
just to explain what I wassaying before with this example.
In this example, what this tooldoes is it creates a summary of
whatever the input was into ashort slack message, in a
specific message, in a specific,channel, right?
So you can use this acrossmultiple agents, not just j that
we use.
If what you wanted is to postthis thing on Slack, sometimes
you want an agent that willrespond to things on Slack, and
(37:48):
then you'll build a differentquote unquote tool.
It's still connected to Slack,but the definition of how to use
it will be different and it willbe a different tool within your
relevance environment.
Pooja Jain (37:58):
Yeah.
So for example, When I, look forSlack, as you can see there is
post to Slack.
I mean it sees my, but there isan ad emoji reaction.
There is even a separate toolfor that.
Yeah.
Or create a new channel, sendmessage to a Slack channel,
right?
You have different sub tools,let's say for the same Slack.
So I think Yeah, that's a greatpoint is platform you're
Isar Meitis (38:18):
connecting to.
Yeah.
For the
Pooja Jain (38:19):
same platform.
Yes, for same platform.
Based on what you want to do,delete messages, get the file so
you can do whatever in here.
Isar Meitis (38:28):
Yeah.
Pooja Jain (38:29):
Okay.
I hope this clarifies theconcept of tools.
So as you can see, yeah, thereare different tools by the way,
like it is so scalable.
So since I'm running this fordemo, I did not add a lot of
tools.
Otherwise it becomes very slow.
But let's say if you want to getthe funding news about your
competitor, you can add a toolhere that goes to Crunchbase and
(38:49):
start getting the funding news.
Same if you want to have anothersocial media.
I only added LinkedIn, but ifyou want to have another social
media where you want to trackyour competitors, you can add
that.
So it is exactly as is Idescribed, like Lego blocks keep
on adding or keep on deletingbased on your use case.
Moving on, Now knowledge thinkof it as a brain of your agent.
(39:13):
For this particular use case, Ihave not added any knowledge,
but let's say if you arebuilding a customer support AI
agent, which you want to bespecifically trained on your
business, then knowledge plays abig part because what you have
to do is make sure that it hasaccess to your business,
specific information to yourwebsite, I don't know, to your
(39:34):
vision, mission company onepagers and so on.
And as you can see, you can addknowledge in multiple formats.
So PDF website, existingknowledge.
The knowledge here is really therag.
So retrieval, augmentedgeneration, that means once you
have a query or once the userhas a query, the LLM model
(39:56):
really goes in there, finds themost relevant, answer, search
for it, or generate, retrievesit and then generates a respons
for, so that is knowledge foryour AI agent.
any questions here?
Isar Meitis (40:13):
no.
I think that's prettystraightforward.
I think people are very muchunderstanding of this from just
using LLMs, right?
It's the same concept or usingcustom GPTs or any of these
tools.
It's just adding informationthat it will make it more
specific to you versus just thegeneric universe.
Pooja Jain (40:26):
Yes.
And the last part here istriggers.
So triggers are, so for example,here I'm triggering the, or I am
starting this AI agent bychatting with it.
That is one way.
But let's say, when you think ofautomation, maybe you want your
AI agent to run every week, ormaybe you do not want to come to
the relevance AI interface, butrather trigger it via some third
(40:48):
party.
So then you can use one of thesetools and use that directly.
Maybe a Telegram channel,WhatsApp, these are premium, so
they need really high credits.
but yeah, that is a way to alsoinvoke your AI agent.
So this is, another part of it.
Isar Meitis (41:07):
Yeah, so those of you who are familiar with the
Zapier and makes of the world,it's the same concept, right?
An email comes in, it's the sameconcept, configure one agent, an
email comes in from a specifictopic, starts the agent.
A meeting is set by a specificperson, starts an agent, a
message on Slack from a specificchannel like each and every one
of these things, the tools weuse every single day.
(41:28):
let's take our example of aresearch thing.
You just created a new accounton the CRM.
As soon as it's created, it willtrigger, the agents, the
original logo.
Go do the steps, do theresearch, do the thing.
We'll create a summary for youin the CRM without you having to
do anything in between, otherthan just creating, the new
account.
So you can think about every oneof the systems you use daily as
(41:48):
both something it collaboratewith, but also as a trigger that
will actually initiate theprocess.
Pooja Jain (41:54):
Yeah, that's a very good description now.
The final part how it looksreally in the AI agent
interface, right?
as you can see here, if you'reaware of the ER and make, you
know that er, and make, you haveto really, tie down the tools or
connect the tools with oneanother, and it runs in a very
specific sequence, right?
But here I have just given itthe access to tools.
(42:16):
they run in no particular order.
for example, if my question isin a way that I want only the
Trustpilot summary, it'll onlyrun this tool.
So there is no particular order.
And another thing is, you havethe option to actually decide.
So whether this tool all runsalways, or it requires an
approval as I did for Slack, forexample.
(42:38):
Or you let the agent decide soyou can really make a
deterministic, always run this.
Or you give the autonomy to yourAI agent to decide, or you, add
human in the loop here.
Everything is doable here.
So that is, I think, a very,major difference when it comes
to, what I observe, when I workwith me or with, for example,
(42:58):
these AI agent builder,relevance.
Isar Meitis (43:02):
Fantastic.
Puja, this was an amazingoverview for beginners.
I think we covered reallyeverything people need to know
to get started.
Uh, Renee on LinkedIn literallysaid, I just signed up for a
test account.
I'm gonna start playing withthis.
So you got at least one personexcited enough, and less scared
of building agents to actuallytake action.
That is wonderful.
(43:23):
so it's fantastic and I'm suremore people will do it.
There was a question earlierthat I wanted to wait for the
end to answer, but I think It isinteresting.
There are other tools out therelike relevance, right?
So I'm trying to see what theymentioned, like Mind Studio
Lindy, and there's A bunch ofothers.
I don't know if you just doveinto relevance and that's like
your universe.
No.
Or you play with some of theothers And do you know, do you
(43:44):
have a reason why or preferenceswhy use relevance versus the
others or not?
Pooja Jain (43:49):
No.
No.
So I am an AI trainer.
I play with all of them.
Okay.
So why IT tool stack reallyinvolves relevance, Lindy and
eight n and now also Zapieragents, for example.
Isar Meitis (44:01):
Yeah,
Pooja Jain (44:02):
why I chose relevance for this particular
use case was because there arefew benefits of relevance.
They are cheap, so it's a$20subscription every month and you
get 10,000 credits, which isgood enough to run such kind of
use cases.
Whereas when it compares toLindy or Zapier, that is very
complex, that is very, expensiveto run.
(44:22):
So they consume a lot ofcredits, and I think their
monthly subscription is alsohigher than this.
That is one Second is comparedto NA 10, NN is very popular
right now, but I personally canbuild a lot of NA 10 workflows.
But NA 10 is still a bit moretechnical.
So when I, when it comes to
Isar Meitis (44:39):
way more tech, way more
Pooja Jain (44:40):
technical, yeah, it is way more technical.
So when it comes to training thenon-technical, audience, they
want something that they can seeand intuitively understand.
So this is why I find relevanceto be slightly better, for
non-technical folks.
But once you get used to of,these AI agent concept, I mean
you can use whatever, and itthen is very, efficient in terms
(45:00):
of price, in terms of creditconsumption.
You can literally run n it andfor four,$4 per month.
So that's the benefit of it.
Isar Meitis (45:09):
one small thing about NA 10, since you mentioned
that NA 10 is an open sourcetool, so you can pay them to run
it on their platform, which isstill cheap, but for very
little, you can host it on yourown and then You can run it
unlimited.
Amount, maybe not at the highestspeed, but speed is not a big
deal here.
Like even with what we just did,most of the agents, the speed is
not critical.
(45:30):
So you can still use arelatively cheap hosting plan,
four, five,$6 a month and run asmany agents as you want.
And yes, instead of getting ananswer in five seconds, it'll
take 10, 20, 40.
Who the hell cares?
Like it does the work for me.
I don't have to do it.
and it costs me four bucks amonth.
So there's benefits in runningNA 10, but I agree with Puja a
hundred percent.
It's more technical and not assimple to use as relevance as an
(45:52):
example.
yeah.
Puja, if people want to knowmore about you, work with you,
learn from you, I know you'relaunching a course.
what are the best ways toconnect with you and do more
with you based on the amazinginformation that you have?
Pooja Jain (46:05):
Sure.
So if I would love to connectwith you guys on LinkedIn.
I'm super active there.
I always share some practicaltips on AI agents and business
use cases.
Please connect with me.
I am launching a course nextweek.
It is called Six Week AI RevenueAccelerator, really focused on
small and medium businesses andwhat we are going to build in
these six weeks, starting fromthe basics to really building AI
(46:28):
agents for content creation,lead generations SalesOps and,
customer success, and personalproductivity.
Oh, wow.
So end-to-end course.
and I'm running it, inpartnership with Valeria.
I believe some of you who hasbeen on this podcast
Isar Meitis (46:42):
as well, so previous guest, fantastic.
Puja.
Again, thank you so much.
This was absolutely amazing.
Thanks everybody who joined us.
We had multiple people on theZoom and on LinkedIn, and I
think this is the most activechat I've seen in a long time on
both platforms.
So I didn't ask you all thequestions.
I chatted with a lot of them andjust answer them, but great
participation.
Pooja Jain (47:02):
Can I go and check the questions?
Maybe I can answer some of them,or, do they after?
No, I, all
Isar Meitis (47:06):
the stuff that I knew how to answer quickly, I
didn't answer, but you can alsoanswer them, right?
they stay on LinkedIn, so youcan just go there and answer the
questions.
I'll try my best.
yeah.
so thanks everybody for beingwith us on the live.
If you're not here, you shouldjoin us.
Like we do this every week,every Thursday.
And at noon Eastern.
I will remind you, that, ourcourse also starts on Monday.
So you have two courses to pickfrom and to be completely
(47:26):
honest, it doesn't matter whichone you pick, but take a course,
accelerate your AI knowledge soyou can do more with AI in your
business and for your owncareer.
So pick a course and go do it.
It makes a very big differencein your ability to actually do
the things that you need to doin a much more effective way.
and that's it.
Thanks everybody for joining us.
Thank you again, Puja.
Have an awesome thank you.
Rest of your day, everyone.
Pooja Jain (47:47):
Thank you so much for having me.
Bye-bye.
Popular Podcasts
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
Ridiculous History
History is beautiful, brutal and, often, ridiculous. Join Ben Bowlin and Noel Brown as they dive into some of the weirdest stories from across the span of human civilization in Ridiculous History, a podcast by iHeartRadio.