The core theme is about leveraging AI, specifically Large Language Models (LLMs), not as a complete replacement for human effort, but as a powerful tool to augment existing business processes, enabling smaller companies to achieve capabilities previously exclusive to large enterprises. The focus is on a practical, iterative approach to building AI solutions, starting with simple "reactive AI" and progressing to "proactive AI" and "actionable AI" to solve specific business problems, particularly in customer success and operations.
Mind Map
Click to expand
Click to explore the full interactive mind map • Zoom, pan, and navigate
Literally for the first time in history,
if you have 20 bucks a month, you're
able to do things that only gigantic
companies would be able to do just by
using chatbt. All of this, just for
context, would have never been possible
pre- AAI because this is where true
intelligence [music] is being layered on
top of the data that we're giving it.
>> I think so many people fail at trying to
implement AI because they
I'm Alex Lieberman joined by my
co-founder Arman Hzarani. We're the
co-founders of 10X. We help companies go
from AI apps into AI native, build
custom solutions, and uh do enablement
and trainings to train people on the
technology that we built for them.
You're probably wondering, why is he
wearing this hat? I'll tell you why. I'm
in New York City. It's like 30°. I don't
want to freeze my you know what off. But
this isn't about me. It's about you. You
have given us your time watching this
video and we wanted to make it as easy
as possible for you. So, we turned this
whole episode into a playbook on our
website. Just click on the link in the
description and you can get the whole
playbook with prompts, with steps, all
the details you need to apply this to
your work. Check it out. The first thing
that I'll say is, and this is something
we say all the time at 10x, AI is an
incredible technology, but it is just
that it is a technology. And when we
think about technology, we think of
technology as a tool that helps to solve
human beings problems. And that's also
why we don't think AI is always the
right solution. We think AI is a hammer.
When you have a nail, it's great. When
you have a screw, go get a screwdriver.
And so the the entire way I want to
frame this conversation is around what
is the problem we're facing at 10x? What
is our hypothesis? And then how are we
building solutioning to solve that
problem. So the problem is very simple.
We are in a client services business at
10X. We have several customers and next
year we will have even more than
several. And Arman and I as we grow our
organization, we want to have a
consistent pulse on the happiness of our
customers. We want to have a pulse on
how our customer success or and our
strategists are working with customers.
But the question is is how do you do
that as you scale to hypothetically
hundreds of clients? And our hypothesis
is that AI makes it more possible than
ever before to maintain that pulse and
deliver truly six-star service to
customers as you scale. And so this
workflow is not only um going to provide
you kind of a framework for how we uh
approach building solutions in general,
but also how you can deliver great
customer success and keep a pulse on it
as you scale your organization.
Okay. So, the I want to refer back to um
we did a an episode with Wade Foster,
the CEO of Zapier, several weeks ago,
and he used this visual and if you
joined that episode, you have seen this
visual, but I think it's really
important to reiterate. So basically
what this is is it is showing the
spectrum of processes in a business. And
the idea is on the far left you have
what he calls determinism. And
determinism is the traditional way that
technology has worked to enable
processes. Very simply if X happens do
Y. If A happens do B. It is
deterministic. It's finite. It is black
and white. What happens on the fully
opposite side of the spectrum is
inference. And inference is just another
way of describing the power of AI and
what AI is capable of. And the the
beauty of AI is I could go tell
ChachiPT, hey, I want you to build a
travel itinerary for me. And I and my
rough preferences are I like adventures.
I like staying in four-star hotels. and
I want it to be a place that I've never
been before and here are the places I've
been just go build something.
Deterministic technology could have
never done that. AI is capable of doing
that because it is more intelligent and
it is the technology is probabilistic.
The reason that this diagram I think is
really helpful is while everyone has
talked about that 2025 was the year of
AI agents, I would actually argue that
most of the solutions that companies
should be thinking about building as it
relates to AI right now are either
something you would call an AI workflow
or an agentic workflow. And what that
basically means is if you have a
process, let's say you have a 10-step
process, the majority of that 10-step
process is still going to use technology
and automations in the way that they've
always been used, which is in a
deterministic fashion. If this happens,
do this. And the idea is that as AI gets
better, more and more of this
intelligence, more and more of this new
technology will be able to be sprinkled
into steps of the process where it makes
sense to use a technology that for all
intents and purposes behaves like a
junior employee who is really motivated,
really smart, but can be forgetful and
go off the rails sometimes. So that is
why I think it's really important to
understand this graphic because I do not
think it's realistic for people to think
that you're just going to build AI
products, you're going to tell it to do
something, you're going to go uh grab
coffee, grab lunch, come back to it, and
it's done. Humans are still very much in
the loop of this entire thing. Arman,
anything you would add?
Yeah. I mean, [clears throat]
I think that if you ever catch yourself
thinking about AI
in this net new magical way, like if you
ever start thinking like talking about
it in all the marketing terms that Sam
Alman and and Daario use just like like
I think it's very important to just come
back to earth and realize that like I
think so many people fail at trying to
implement AI because they immediately go
all the way to the agent side here and
they try to oneshot everything. And I
think that it's really really important
to note that if I historically told you
that you can get 10 5% improvements
across your business by automating one
step of the process using AI, that
historically would be gamechanging.
But people always try to like automate
the entire thing. And when when AI can't
do it in one shot, they think that it's
a failure. And so that's the first
thing. The other thing to note is I
think one helpful mental model that I
always find myself going back to is if I
want to delegate some of my work to AI,
how would I delegate this to a human,
right? And how junior is this human and
how much experience do they does this
human have? What what context does this
human have? and using that as a as a
framing for the thought process around
how to delegate to an AI whether it's a
workflow AI uh agentic workflow or full
agent. Um I think that's a really
helpful way to think about it.
>> Yep. Absolutely. And in a in a few
minutes we're going to talk about as I
go through this workflow of measuring
customer health and kind of how do we
gradually make it more powerful. Um, as
we go through those steps, I think
thinking about the junior employee
analogy is gonna be really helpful. Um,
so let's keep going. Also, Arman, I can
only see my slides, so if there's
anything in the chat that people are
asking. Yeah,
>> let me quickly answer Josh Dance's
question. I think it's a really good
one. Like, what is the difference
between agentic uh workflow and a full
agent, right? Um,
a full agent is I go into clawed code,
okay? and I tell Claude code, I want you
to build Facebook.
Okay, that's it. I just go in and I try
to oneshot the entire task.
Historically, building Facebook is
actually many many steps, right? But I
can just trust AI to do the entire thing
for like come up with its own plan,
follow its own steps and figure it out,
right? That's that that would be a full
agent. I would argue that we're not
there yet, right? So, what you can do is
you can come up with what are the steps
that I know this agent is going to need
to follow. I know that the first thing
it's going to need to do is do research
about what the hell it what does it mean
to build Facebook, right? So, go do
research and build a PRD, like a product
requirements doc. That's step one. Then
step two will always be take that PRD
and design the technical architecture
for building Facebook. Step three is to
go through step by step of that
technical architecture doc in order to
complete it. Right? And so that would be
like a loop over those things. And so
the difference between these two things
is for the agentic workflow I am telling
the system for step one I always want
you to do this in that step it is
agentic. So in that step the AI is able
to go and do its own thing whatever but
I will take the output of that step and
give it to step two and step two will
always have the same objective and so
on. With an agent again it is creating
its own plan. It is doing its own thing.
Again I think a helpful way to think
about this is like an employee. So with
a very senior employee you can tell them
hey let's say I'm a VC. I want you to go
deploy $2 million of capital, right? In
order to deploy $2 million of capital as
a venture capitalist, you have to talk
to hundreds of companies. You have to
come up with thesis around the market.
You have to uh make a decision on where
you want to invest, right? So, there's
many, many steps. But with a really
senior partner at a VC firm, you can do
that. But with more junior VCs, you're
going to tell them, I need you to talk
to a hundred customer or 100 companies
every every month. I need you to write
up memos for each of them. I need you to
write me a thesis and I need you to come
to me with an exact document doing the
diligence that we need for these
companies and then together we'll make a
decision. And the difference between
those things is I have a lot of
structure with the junior person. I have
very little structure with the senior
person. That's how I think about the
difference between agentic workflows and agents.
agents.
>> Love that. Um cool. Let's keep it going. >> Yep.
>> Yep.
>> Okay. [laughter] Oh, is that a picture
of me? This this everyone is my
co-founder Arman. uh he uh is c he's
currently on safari in Kenya. Um no so
Arman always says this really good line
which I think is such a good framing for
not just how we're going to go through
this workflow but how I think about
building any sort of AI products or just
honestly any products or solutions in
general which is Arman. Do you know what
I'm about to say?
>> Yes, for sure. You you can say it.
>> How do you eat an elephant?
>> I don't know. One bite at a time.
>> Yes. And I think when people are
thinking about building new AI workflows
or processes or products, they think
about like the pie in the sky dream of
what they want to build. So at if we
even just talk about what is the perfect
pie in the sky dream for what a customer
health and happiness um system would
look like at 10X. Here's what I would
imagine. We have a an application that
ingests any information that relates to
us interacting with our customers. So,
it pulls transcripts from notion uh from
our uh notion because that's where we
transcribe our calls. It pulls all
tickets from linear, which is where we
project manage the software we're
building for clients. It pulls messages
from Slack. It pulls emails from Gmail.
And then in a beautiful dashboard, it
visualizes our interactions like average
response time. Uh uh it buckets like it
has um a red light uh where there are
certain clients named that uh are in the
red like they're high risk yellow light
green light and we have this beautiful
dashboard that's visualized all of our
internal data and then even more
valuable than this is this dashboard
then can be talked to and I can say to
the dashboard hey based on what you're
saying about client red what is the next
action action we should take and
actually can you take that action Can
you actually do that for us? That's the
pie in the sky. And honestly, what what
I'm going to show you by the end isn't
that far from there. And that is
ultimately our goal for where we want to
get to with 10x. But that is the
elephant. And the only way to get to the
elephant is take one bite at a time. So
Arman always talks about this when we
talk about product at 10x is how you
scope down to the most important, most
atomic unit of what you're trying to
build. How do you start there and then
build up from there? And that is why
we're going to eat the elephant one bite
at a time. So this is bite number one.
That's that's someone biting. So what is
the first question that we are trying
that I'm trying to answer for our
customers at 10X? Very simply, we create
software for our customers, either AI
software or traditional software. And so
my number one question as I'm thinking
about um customer success and are our
account managers uh successfully
managing our clients is are we shipping
software like we promised? And so this
gets into how I think about building out
this AI solution that I'm talking about
and where I want to start. The first
level of building a solution is what I
call reactive AI. And reactive AI very
simply allows you to talk to your data.
So before doing anything else, I want to
scope down to the simplest way for me to
talk to our data and specifically our
customer interaction or engagement data
so I can easily understand what is the
state of the software that we're
shipping for our clients. So I'm going
to quickly demo that and then we're
going to build up from there. And what
I'm going to play out for you, and this
is what Arman was referring to before,
um, the analogy is there are basically
five steps that I think about taking in
any sort of process build, product
build, whether it's AI or just
traditional technology. The first is how
do we connect to the right sources?
Meaning, how do we connect to the right
information so we pulling in the right
data to learn the things we need to
learn about our customer interactions.
Second, how do we create a great prompt?
And what a great prompt allows us to do
is to get insights from that data.
Third, test the workflow we're building.
Again, the smallest version of the
workflow. Fourth, iterate. Because one
of the expectations I always try to set
with people is you are not going to
oneshot whatever the workflow or product
you're building uh is, you're not going
to oneshot it. It is going to take
iteration. the more complex or the
bigger the thing you're building, the
more iterations it will take to get it
to work right, which is why you want to
scope down to the simplest use case
first. Once you iterate and get it to a
great place, that is when you can add
functionality. And think about this
again, go back to Armon, think about
this as a junior employee. The first
thing you want to do with a junior
employee is give them the right
information. Give them the context that
they need to operate within the uh rails
that you've put them inside of. Then
create a great prompt. What's a great
prompt? It is what are the very explicit
instructions you have given to an a
junior employee to do then test the AI
test the human they go off they do the
work they come back the work isn't
exactly done properly that's where you
iterate iteration said differently for a
human is giving feedback you go through
this loop once the feedback is clearly
worked they're doing the work well you
add functionality or in terms of a human
you add responsibility because you built
No, I'm just laughing at the said
differently. Um, [laughter]
>> that that's that's an Alexism. Okay. So,
what we're going to do is I'm going to
show you how we start with this the
smallest unit here which is talking to
our data and specifically for us that is
talking to linear to understand are we
shipping software at the speed that we
want to to make clients happy. So, let
me get out of here. The one thing the
one thing that I do want to add as
you're following this up is we have a
few questions in the chat of like is
this a product that 10X is building? Is
this just a bunch of Zapier
integrations? Like this is
like we are we are live building this
for our own company. Like everything
that Alex is showing you we are building
for ourselves. We already have a lot of
it built for ourselves, but we're
walking you through how we think about
building this because right now we've
done it for client success and client
support and all that, but we this is how
we do it for every part of our company.
And the goal is that by the end of this,
not only will you be able to literally
copy and paste this tool for yourselves,
um, and I know that there are a bunch of
companies that have this exact product
basically that we're building on top of
Xavier for ourselves, um, they have
this, but they charge a bunch of money.
you guys will be able to basically just
copy and paste this yourselves. That's
the first thing that I think is really
great. But also um you'll be able to
identify other opportunities in your
company and likewise build solutions to them.
them.
>> Yep. Absolutely. Okay. So the first
thing I want to do again is talk to my
data. And to us the for me the most
important data that we can have access
to is effectively our project management
board which because we're we're building
software is linear. I want to be able to
ask linear questions and and I really
would love to get like a report that
just tells me how are we moving along
with every client and what are risks
that uh you would dig into to learn more
about. And just think about this in
context. Let's assume in the future we
have a hundred clients and we have a
hundred different linear boards. It is
not going to be realistic for Arman or I
to get in the weeds or look at every
individual client's linear board. But we
want to make sure we are pushing forward
the work we're doing with clients just
as effectively as when we had one
customer. And so where I always start
again is I connect to uh I connect to
the right source. And so in linear let
me just find it here. Uh sorry in um in
claude uh and catch has the same thing.
These companies have connectors. Uh I I
Arman I laugh doesn't like these things
because connectors are basically just
MCP. MCP from Arman's point of view is
just a glorified version of APIs. But
all this to say that there are these
connections that Chashibbt and Claude
have created to talk to your
applications. Linear is one of them.
Linear also is a connector in chash up.
So what I'm going through is
interchangeable. So once I've made the
connection to linear that means now I
can access data that we have around
project management with our customers.
Then the next thing that I always do is
I create a great prompt and I people
have all these formulas for what is what
is a great prompt. My general formula is
this. If I'm talking to a junior
employee, how do I
increase the odds of them comprehending
what I'm saying so that there isn't
error because things got lost in
translation and how do I treat getting a
prompt in the same way? So, what I
basically said here is I want you to
create a prompt that helps me understand
how much software we're shipping for
clients and how many story points we've
completed this month for each client. I
want the output to be anonymized since
I'm showing this to a group of people.
create a prompt that I can feed to
Sonnet 4.5 that using the linear
connector will allow me to understand
the state of each customer's linear
board, how fast or much we're shipping,
and how many story points we've
completed this month. And just uh for
context, because people may not know
what story points are. Story points are
just a way we measure how much output we
are creating for a client, how much
software we're actually building for a
client. And then I said the linear
boards I'll want to monitor are
attached. This is a really small but
like I think is one of the underrated
things that LLMs have made super easy is
there's no easy way for me in linear to
like copy the name of all of our
different boards and paste it in. And so
now I just take screenshots and LLMs are
incredible at taking screenshots and
turning it into text. So I just took
screenshots of all of our boards,
attached it into Claude and it turned
that into a text list. So that is the
prompt that I create and then I take the
output of that prompt and I just in a
different chat feed it back to the LLM.
So let me just go here. So basically
what the output was and I'm not
scrolling up because it actually has the
names of our clients but basically it
gave this really thoughtful
prompt for telling Claude how it wants
to uh use Claude to generate a report by
accessing our linear data. So what are
the output requirements? Anonymization,
report format. Um based on the number of
story points we create for a client,
mark them as uh green like uh high
velocity, yellow uh medium velocity or
red low velocity. After that, add the
average story points per client, top
three performing clients. Additional
insights, flag any clients with zero
activity this month. note if any clients
have large backlogs. Uh identify any
patterns and then what that actually
results in is this is the output of
the uh prompt that I gave. So this is
basically a our linear customer board.
We're talking with the data by customer.
It shares how many issues have we
completed. What is each task that we've
done for each client by client? what is
the activity level, what is the current
project that we're working on, what are
the key themes of each project, and then
it also will share um what are potential
issues that we should be flagging. And
so like um what I can look at here is
there are certain companies that have
issues that uh they have not reviewed
our work that we've done for them in a
long time. And so thinking about how do
you actually turn this into action? What
I'm trying to understand here is where
are their bottlenecks? Where are we
getting slowed down? And now that I have
this at scale, now I can zoom into who
are the two to four clients where things
are getting stuck in a certain part of
the process. Now I can go into Slack,
ask the specific technical strategist,
hey, what's happening here? And I've
been able to focus my time on the things
that actually matter because creating
this integration allowed me to scale to
hundreds of clients but focus down on
who are the few that I actually need to
care about right now.
So that is that is the first example of
talking to your data and again creating
a connection with whatever is your
source of truth like for customer
interactions. And the one thing I'll say
going back to the slide that I had about um
um
start small, test, iterate and then uh
increase complexity. The way I would
increase complexity here is we started
with linear because in my mind linear is
the source of truth for us is of how
much work are we actually doing. But
then the next way to add complexity here
is not just to do things like make the
AI more proactive, have it take action,
but also make it multi-threaded. And
when I say multi-threaded, get other
data inputs that work into understanding
our customer health. So not just linear,
but what is our average response time in
Slack with our customers? How many Slack
messages have we had with them? Take a
look at our notion meeting transcripts
with them. Are there any signals you got
from there? And that's what we're going
to go into in a minute. Um, I'm going to
pause there, see if there are any
questions before we kind of dial up what
this workflow looks like as we start
introducing other variables. One thing
that I want to highlight here as well as
we're waiting for questions to come in
in the chat is [clears throat] that
that
we get a lot of questions about like,
oh, like I'm using chatbt. What's next?
Right? Like I'm using Claude. What's
next? Arman, you're you're um you're
pretty frozen right now. >> Oh.
Let me hotspot.
>> Well, while we It's all good. We'll keep
going and Arma will work on his Wi-Fi.
But um so any questions on this
integration between Claude and Linear
before we kind of ramp up to making the
the workflow not only multi-threaded but
also more proactive and actually be able
Uh Josh said where do you go to view the
report? So in this example I am going to
claude to view the report and typically
again like claw claude and GPT I use
them interchangeably right now. They're
kind of my daily drivers. So I just
always have it open. In GPT there is the
ability and this takes things from let
me just share my um let me share my deck
again because I think this is an
Okay so going back to the different
levels of creating AI products or
processes there's level one is reactive
AI which is what we just built. You talk
to your AI and you ask it for insights,
but it you are pushing the AI to do
something. Level two is proactive AI and
that is the idea that AI works in the
background based on a schedule that or
some trigger that you've dictated to it.
And so I don't I don't know if they have
this with Claude. It may have been added
in with skills, but with GPT there's a
way to add in recurring jobs. So if we
ran the same exact flow where we're
connected with linear and we have this
report run in linear, we can run it as a
daily job where every day a chat is
created in GPT that delivers this report
based on the last day or week's activity
within linear. So to answer your
question, that's how it would currently
be done. If you use other tools like
Zapier, which we're going to go to in a
minute, then you can have the output
happen via email, via Slack. You have
more flexibility than building kind of
th this workflow I just shared within
GPT or uh Claude. To Mark's question,
linear is think of it as like Air Table
or Monday.com. It's a project management
tool, but specifically focused on
engineers and product managers.
>> Arman, you were going to add something before.
before.
>> Yeah. Um, so everyone can hear me,
right? I'm I'm clear yet again.
>> Yeah, you're good.
>> Okay. So, one thing that I just want to
highlight is we get questions a lot from
from clients, companies, everybody that
>> they basically say, okay, I'm using Chad
GBT, I'm using Claude. What's next?
Like, I'm doing this and I'm and I'm
really good at it. Like, what is the
next level? And I think that there
absolutely are next levels, but there's
always more opportunity to get out of
Chad Shept and Claude. And I think one
traditionally what Alex just showed
would really only be done by like a
full-time person. Like that would take
Think about before AI, which like
literally is hard for me to wrap my head
around, but like before AI,
a person would have to go through this
like linear system, which is basically
like it's a it's a giant project
management system. They would have to go
through the project management system
and they would need to go through client
by client, column by cl column, task by
task and they would need to copy and
paste. Okay, this was done, this was not
done. Okay, for each client, how are
they feeling? Let me look at the slack.
This would be a dayong job. It would be
incredibly expensive. And so the only
companies that would be able to actually
afford having a message like this sent
to the co-founders of the business every
day, you'd have to be gigantic. But for
the first literally for the first time
in history, a company like 10X and a
company like of any size, if you have 20
bucks a month, you can afford to have
this message sent to you. And I think
that is what is incredible here is that
you're able to do things that
traditionally only gigantic companies
would be able to do just by using
chatbt. And so I always think that
there's more opportunity. And then we'll
see in level two that it's even better.
So Alex, I'll throw it back to you.
>> Yeah. So level two, we're going to ramp
this up a lot. And so we're going to do
two things. one is we're going to make
this proactive so that this customer
health report um is generated with a
level of frequency that we want. From my
perspective, Arman and I would want this
weekly to really keep a good uh pulse on
the business on a weekly basis and we're
going to make it multi-threaded. So when
I think about it, going back to what I
was saying before, there there are three
or four data sources that together give
us a great picture of what is the
directional health of a customer given
what they're saying, how they're saying
it, and how we are pushing forward the
projects that we're working on for them.
And so my next goal here with proactive
AI is to have something running in the
background with a level of frequency and
pulling in all of the data sources that
tell us more about our customers. So let
me let me keep this going.
Um so we we had the first bite of the
elephant. Are we shipping software where
we uh like we promised by through Claude
setting up the linear connector getting
a great prompt asking it for a great
report in that report having it point
out what are specific clients where
there is a backlog of PRs in review so
that I can reach out to or Armen can
reach out to our technical strategist
say hey why is so and so client not
pushing forward the work that we're
giving them and uncover some bottleneck
there. Next step is give me this insight
without being asked and
that is where we get into proactive AI.
And so for this I'm using Zapier and
basically when I use Zapier you can
basically assume I am building something
that is code on the back end but all the
code has been abstracted away for
someone like me that is not an engineer.
So if you have access to engineers, you
could build what I am building or you
could use Zapier and do it yourself as a
non-technical person. And I'm sure Arman
could give you several ideas for as an
engineer how you would build this. So
let me walk you through my process for
making this multi-threaded proactive AI
with Zapier.
So let me share this first.
Um stop sharing presentation.
presentation. Nope.
Nope.
>> As you're pulling this up, I wanna I
want to mention a question that we've
seen a few thing a few times in the
chat. Like folks are asking things like,
"Will this work with Jira? Will this
work with Asana?" Um, these ideas that
we're presenting here are foundational
to AI development in business. So like
level one, you can ask AI questions
about your data. Level two, Alex is
going to walk through like like AI will
proactively tell you things about your
data. These these levels, these systems,
these frameworks work regardless of
where your data is, what your data is,
right? Like right now we're just talking
about this one specific use case, but it
can work like even this specific use
case can work with
linear, Asana, Jira, and so on. but they
can also work with your ERP and CRM and
they can also work with your Slack and
they can also work with your email and
your calendar and so really like like I
think it's it's important to dig into an
example to see the the power and the
fire but definitely pull out zoom out
and think about like where else where
else can this apply and the answer is
probably yes.
>> Yep. Absolutely. So this again where I
start with everything is with giving the
right information which I already set up
in in Zapier which you'll see in a
second and creating a great prompt. And
generally like you can have Zapier and
we'll show it in a second create the
prompt for you. I've always just found
that I feel the most comfort with like a
GPT 5.1 or a set 4.5 creating the prompt
because oftentimes like in Zapier or any
of these tools they are just using those
models to create the prompt as it is. So
basically here I I fed exactly what I
wanted created that I described to you
all think hard and create an amazing
prompt I can feed to Zapier. I wanted to
ingest a bunch of signals from client
interactions and use that to create a
weekly comprehensive client health
report that provides overall client
health as well as deep insights client
by client. I want you to take liberties
to make this as specific, deep, and
actionable as possible. But there are
things I think should be included by
client, customer health score, key
signals, key quotes, areas for concern,
potential issues, number of Slack
interactions, average Slack response
time, number of story points completed,
blah blah blah blah. I I fire this
prompt off and then I was like, "Oh,
damn. I also wanted notion. So then I
updated the prompt while it was
mid-thinking and said, "Oh, I also
wanted to pull insights not just from
Slack and Linear, but also call
transcripts from notion." And so then
what it ended up giving me was this just
like really in-depth prompt for scoring
clients based on the inputs. And then
the output format is the overall
summary. So average client health score,
count of clients by status, short
narrative summary, key emerging risks,
key positive trends. Again, all of this
just for context would have never been
possible pre- AAI because this is where
true intelligence is being layered on
top of the data that we're giving it.
And then we have a client byclient
breakdown. So for each client, health
score, status light, renewal, churn,
risk, expansion potential, oneline
summary, key signals this week, key
quotes, uh slack activity, linear
progress, call insights for every single
client. So now let me go to the next
what how that actually shows up in uh
Zapier. So let me share this step. So
we're in Zapier now. There's the agent
builder which is uh functionality in
Zapier. And what I basically did is I
fed the prompt to uh Zapier in agent
builder. And as you can see the trigger
says on demand right now. I just did
that so we could run this um agent if we
wanted to right now. But obviously you
can also do it as you can see you can
schedule by Zap year. You could have it
due every week. Choose value. I would
want it Mondays at
8:00 a.m. so that Arman and I can get it
when we get to the office first day of
the week. And now that is the recurring
trigger. So I'm not going to run it now
because it's going to take time, but I'm
going to show you what the output looks
like. So let me go over here. Um, so
what I basically did is we also set it
up so that it emails to us so that it
would send us emails and also our core
team a Slack message with what weekly uh
customer um like customer sentiment
looks like. And so this is basically
what we're sent. We have the executive
summary which has top three risks. So
one client who's working on field rep
testing, another one who's working on an
eBay integration, another one who's
dealing with performance delays, the top
three opportunities. So what are actual
opportunities for how we can wa wow
clients over the next week. What we then
have is client by client breakdown. So
client A and this is ordered by risk. So
health score 6 out of 10, churn risk
medium, trending down. So that is not a
good sign. So like this is something
that Arman and I would probably pull the
technical strategist in to go through
this breakdown and be like tell us where
the concern that's created in this
report is wrong. Where is it right? What
are the actions we're taking to have
this trend back up and get this from a
red to a yellow. And so then it also
shares immediate actions required daily
check-ins until the field rep testing is
done. Escalate location in uh issues to
senior engineering. Prepare contingency
plan if testing is delayed. And this
goes through with every single client
ingesting Slack, ingesting linear, and
ingesting our Notion call transcripts
all to work together this full picture
of how we're doing with clients. Uh,
Arman, anything you would add? Yeah, I
mean I I just again I like I'm going to
sound like a broken record. I still
think it's insane that like historically
the president and Alex and I talk about
this all the time like there's this
thing called the presidential daily
brief, right? Every single day the
president wakes up and they get a huge
like stack of papers and it is
structured exactly how the president
wants with all the top news, all the
things that they need to know. So they
wake up, they read this presidential
daily brief. This is a thing since I
forget which president started it. I
once I heard about that I was like that
would be incredible to get that right
and there are a ton of newsletters right
like we all know them like Morning Brew
and all the other ones um and you can
wake up and you can read those but what
if you could have your own right and I
think that it's incredible that
basically what Alex just built is that
for customers right for the customer
success and the and deuce has a question
here about like what is uh that's linear
and call transcripts like can you
include email and CRM The answer is yes.
Like uh Daniel put a list of all the
apps from Zapier. You can absolutely add
those other data sources as well. And
then Doug asked how is it assigning
churn risk and a health score. Did you
tell it what the formula would be? And I
I believe that um that Alex, you did
have a like you did have information on
how to assign that client risk. Yep. And
the entire structure of the response
like everything is in that prompt. So
you can make this completely custom to
what you want it to be for customer
success, for your sales pipeline, for
every part of your company. You can have
this happen.
>> Yeah. So what I would say for the churn
risk question because this is a really
uh good one. In a perfect world, the way
we will ultimately use this and
associate churn risk is look do a look
back on all previous clients who churned
and what were the signals that we saw
from that client. whether it was um
things they said on calls, uh what the
average number of touch points were over
Slack, uh what um what uh linear
momentum like in terms of uh number of
issues resolved or number of story
points complete and that will become
like basically those traits will become
the bar of what creates churn risk for
us. What I did here is I basically uh
gave it conservative estimates of what I
would consider high, middle, and low
risk. But I also said use your best
judgment. So I said like if we are if we
in the last week have done more than uh
50 story points worth of work that is
low turn risk. If we've done uh between
25 and 50 that's medium turn risk. And
if we've done um less than 25, that's
high churn risk. And I basically did
that for churn. I did that for Slack.
And I did that for call transcripts. And
like I said, it was directional. But you
would also be surprised how good these
models are at just like directionally
figuring out what signals lead to an
increased risk of churn. But what I did
make sure of here is again, this is for
me, what's most important here is that
it's directional. I would rather there
be false positives than false negatives.
And so I would rather call out too many
clients that are high risk of turn and
we go and check in on them than not call
out enough. And so that was how I
structured that prompt. Um someone else
had a question about this. Um
did you tell Yeah. So the the whole
chorus thing I would also say is again
if I wanted to also specify it further I
would go back to typically chat GPT or
claude and I would say these are the
this is the type of business we have.
This is what we deliver them. Uh these
are the systems by which we engage with
our clients. Can you create a list of
signals that put someone into a high
middle or low churn risk bucket? And I
would use that in the instructions for
Zapier. And as we get more actual
information over time, I would just
finagle the instructions to be more
accurate to the real data that we have.
>> Yeah. And hopefully that one thing that
I think is interesting is like I would
recommend testing more guidance and less
guidance. And so I would actually
recommend testing and Alex, I don't know
if we've done this, but but we should do
it as well.
almost giving no guidance on what on
what defines churn risk but saying I
want you to be a little extra
conservative because I think that AI
tools part of what's incredible about
them I always call them warm-blooded
right like >> yep
>> yep
>> AI is warm-blooded and there's this
within that warm-bloodedness is the fact
that you can actually sometimes
sometimes AI will tell you things you're
like wow I didn't catch that like I
would not have considered this signal to
be like I didn't consider that the fact
that we're not responding thing in 30
minutes to our clients to be a churn
risk, but but the AI is and maybe that's
actually a good thing to to flag, right?
And so I think that there's there's
value in that. And so like Alex said, I
think testing more guidance, less
guidance, different types of guidance,
different types of conservatism within
that guidance um I think is is is a part
of the process.
>> Yep. Absolutely. So I want to bring this
to like the final uh part of the step.
So you saw the last output which was
basically that is the customer health
report uh that organized clients by
churn risk based on what it deems to be
different risk factors of not enough um
not enough kind of like product being
pushed based on our linear board. Uh not
enough interaction or fast enough
interaction based on Slack. Uh any
negative signals or positive signals in
our notion meeting transcripts. The
final piece of this and let me just pull
up the slide again is so so we've had
two bites of the elephant. The third
bite uh which is like what Austin or
what Arman and I would ask ourselves is
like okay
you we we have a sense of are we
shipping software like we promised by
having this integration between linear
and claude and and asking the data
questions that helped us understand like
oh there's a backlog of PRs for review
that a client hasn't reviewed great that
leads us to a conversation then we were
like we want a higher fidelity and
proactive AI so this job runs on a
weekly basis on Mondays. Um, and it
ingests information from four different
sources and works it together in a
report which you saw what the final
report looks like. And to someone's
question, you can always improve the
inputs of what is high, middle, or low
churn risk based on you getting
information in the business and feeding
those insights back into the
instructions for the agent. The final is
take action. So don't just give me
insights, but take action on those
insights. And so I'll just quickly show what
what
>> where did you get these photos?
>> Um I looked up open mouth on Google. I
asked myself what would make people
smile in their office chair and what
relates back to taking one bite of a t
at a time of the elephant.
>> Love [laughter] it.
>> Yep. Yeah. My brain is crazy. Um so then
the final piece here is active AI. And
so what we wanted to do is we had this
report. This report tells us what are
the action items that we want the AI to
take. Well, how can it help us with
those action items? And so, just to
quickly share my screen again, let me uh
go back to Zapier,
Zapier,
please hold um
um
as Okay, pulling that up. Um, if anyone
has questions, please continue to put
them into the chat. Uh, it's it's really
helpful for us to guide the conversation
in the way that in the way that you all
want and um to to make sure this is
super useful for everybody.
>> Yeah. So, what I'm going to do is I'm
actually just going to show you the
build of this uh in Zapier so that you
can uh you can do this yourself. So, let
me just share my screen of Zapier and
we're going to
do this. So basically the way I wanted
to set this up is every time the weekly
sentiment report runs. So every time
just again using this this report runs
that was created from that uh zap year
agent that we built that in in uh
ingests all the information uh creates a
breakdown etc. I want us to get drafts
of emails that our customer success
people can send to clients based on the
action items that it's detailed. So I'm
I'm going to show basically what the
what the end instructions it created are
for the agent, but I actually want us to
just go through the flow quickly
together. So basically it created
instructions of when the weekly customer
sentiment insight agent completes its
analysis, retrieve the generated action
items and recommendations from its
output. Call the weekly customer
sentiment insights agent to get the la
the latest customer sentiment data and
action items. Parse through the action
items and recommendations provided by
the analysis. For each action item,
create a personalized email draft that
account managers can send to clients. It
creates the draft in Gmail. Structure
each email draft to include professional
subject line, personalized greeting,
context about the sentiment insight that
trigger triggered the recommendation,
specific action or solution being
proposed, a clear call to action for
next steps, professional closing, and
then uh save all email drafts in a
format that account managers can easily
access and customize before sending. The
final goal is transform customer
sentment insights into actionable ready
to send email drafts that account
managers can use to proactively address
client needs and improve customer
relationships. And so very simply how I
created this is I basically said here um
I have a weekly what what is it called?
Weekly customer sentiment insights agent
on Zapier. I want you to um every time
that agent runs, I want a new agent to
draft emails
that um directly
um relate to the action items that you
called out in the report. And I want
these drafts to be actionable, thoughtful,
thoughtful,
and specific slashcustomized so that a
technical strategist at my company can
send it off to the customer with minimal
edits. And then you start building. And
what's nice about these um
these tools like Zapier um or Lindy or
NAND or Gumloop now is they have these
agent builders. So it's basically like
you're in a chatgpt or claudeesque
experience but it is specified to their
platform. So it's actually calling
tools. And so what it basically did is
it built the instructions here and then
it's going to ultimately put the tools
that it needs access to in the
instructions. And those tools are going
to be it needs access to Gmail. So it's
going to ask me for access to Gmail. It
also is going to ask for access to the
weekly sentiment insights report. So the
other agent that I created in Zapier
because that's where it's going to
ingest context from to draft these
emails. And it will set up as you can
see it's setting up a web hook which is
basically it is creating a trigger where
when the weekly sentiment insights agent
fires off it is going to trigger this
agent. So as you can see the tools this
agent can use are the sentiment uh
sentiment driven email drafts when the
weekly customer sentiment sentiment
insights agent completes its analysis
and generate a report automate
automatically trigger this workflow. So
basically this start starts when the
other agent finishes and then it
delivers emails as drafts in our inbox
related to those action items. Any
We were just getting advice to use YAP
to text.
>> Yep, we we agree. We actually took a
poll of the team on uh under underhyped
and overhyped things in AI right now.
And one of our team members said that
Yap to text is very underhyped. Um, and
even uh Ryan Carson uh who we had on a a
previous episode of Human in the Loop,
uh he basically yapped to text
everything as an engineer. Um which I
agree with all this, but I would say in
an office environment that we're in, it
would be pure chaos if everyone was
yapping to text.
Um Daniel Ree, do your clients have KPIs
for AI use? What metrics are they
looking at to judge or evaluate success?
Arman, any thoughts here? Yeah, I mean I
think that like there
business is business, right? Like when a
successful business makes more money
than it spends very simply. Um people
want to be happy. People want to work
less and make more money. Like there are
facts of nature that will always be the
case. And I don't think that AI changes
that. And so I think that companies
still have their own success metrics
that they've historically had. And
whether you use EOS or um OKRs or
whatever, like a company has their
goals. We if we're successful as 10X in
in helping companies adopt AI, but also
other people and their own companies
when they're trying to adopt AI, they
will know when they're successful
because that initiative will drive the
company closer to the goals.
>> And they'll know they're not successful
>> if they're just like saying the word AI
a lot and and it doesn't impact the goals.
goals.
Um, two quick things. One is I want to
answer Cat's question and then Arman,
something for you to think about as I
answer it is if I was giving you this
prompt, if I gave you this prompt to
like build a customer health and happiness
happiness
monitor for the business,
how would you go about doing it roughly
speaking given you're an engineer and
you know how to do this technically?
What would be like how would you think
about building this and would it be
different from how I thought about
setting this up? So, just something to
think about. Kat, to your question, why
wouldn't you want to create one agent to
create the report and email it? Is it
just to try and keep workflows as simple
as possible? The answer here is in
Zapier when you chat with the agent. So
like when I chat with the agent here,
just to give you a quick example, um
let's just share my screen one more time
quickly. Uh here. So when you chat with
the agent, you can the a when you chat
with the agent, you can make changes and
when it makes changes, it makes changes
to the entire instruction uh of of that
agent. So like it changes the system
prompt for that agent. And so my whole
>> just to sorry just to jump in like this
agent that you're chatting with is an
agent building agent, right?
>> Correct. Exactly. So this is like Zapier
basically like historically with Zapier.
Zapier, you can go in, you can like
click click click to build these
workflows, but then they introduce their
own Zapier agent that allows you to
build and edit the workflows. And so
that's the one that that Alex is is
messing around with right now.
>> Exactly. And my I don't know if I'm
doing this sublim subliminally because
Arman has always thought about doing
this like with chats in whether it's in
different terminals with cloud code or
different chats with GPT but it's like
whenever I'm working on uh a net new
thing I don't want to pollute the
previous thing that I've worked on. And
so my fear is is if I worked on this
email generator within the the weekly
customer sentiment insights one and I
asked the agent building agent to make a
change. My fear is is it makes a change
to the initial agent that not only
screws up the email generator but it
also screws up the the report generator.
And so now I have nothing that works
versus containing kind of the poison to
this second flow. >> Yeah.
>> Yeah.
Um, so to answer your question, like
let's say I was to build this from
scratch. If I'm being super honest, I
would just use something like Zapier,
right? Um, and earlier we had a question
about like why Zapier and there's
Zapier, there's NAN, there's make,
there's all these different things. Um, currently
currently
my read is that Zapier is the most
robust. Uh, it has it's just been around
for way longer. So the connections are
great. Um, they their team is building
with AI first in mind. like I just we've
just used all the products and we and we
like it the best. We think it's the most
robust um for this use case. This is
what I would do. What we're thinking
about internally and again these are
real use cases like Alex and I on Monday
will get that message in Slack. Like
this is actually what we use internally.
if I but we're also thinking about like
what does it look like to build an
internal operating system for 10x that
not only surfaces these insights but
even more right and so I think if we
were to make it more robust what you
would do is you would think about it the
same exact way that you do that you did
right where it's like what is all the
data what are all the different triggers
what like what is the information that
we need and then what do you want to do
with it well you want to be able to talk
to it you want to be able to ask
questions and you want to be able to get
some insight ites and those insights are
going to pull from certain data and
they're going to be structured in a
certain way. And and so it would
literally be structured the same exact
way. We would think about it the same
exact way. What what custom building
would give you is just a little bit more
or a lot of bit more customization, but
it's also going to it also comes with
some negatives, right? or you need to um
it's just a lot more work up front, you
know, honestly to
>> and there and there's maintenance you
have to do like Zapier like they are
doing all the maintenance on the back
end. So yeah, there's a ton of
trade-offs. The the last question
because I know we're at time is just to
answer Mark your question, can you show
us the production zap that you made and
walk us through the actions? Um yes, the
the cool thing about Zapier is after you
build an agent or an automation uh you
can actually share it uh for people to
like just customize on top of. So when
we send the um the postrecording email,
we'll include links to both the report
generator and to the email generator
that connects to the report so that you
can again it's going to take some
finagling to connect the right data
sources for you and your customers, but
it's at least probably 75% of the work
is there for you to build on top of.
Sweet. Um I think that is it. I want to
be respectful of everyone's time. Thank
you all as always for joining uh the
show and we will catch you next week on
Human in the Loop. And feel free to
email us if you have any questions at
all. Alex alex10x.co. Arman 10x.co.
Thanks everyone. Have a good night.
Click on any text or timestamp to jump to that moment in the video
Share:
Most transcripts ready in under 5 seconds
One-Click Copy125+ LanguagesSearch ContentJump to Timestamps
Paste YouTube URL
Enter any YouTube video link to get the full transcript
Transcript Extraction Form
Most transcripts ready in under 5 seconds
Get Our Chrome Extension
Get transcripts instantly without leaving YouTube. Install our Chrome extension for one-click access to any video's transcript directly on the watch page.
