E07: Company-Scale Agentic AI

Company-Scale Agentic AI: Why AI Isn’t Stupid, It’s Blind (and How to Fix It)

In this episode of Old School / New Tech, Ran Aroussi explains what “company-scale agentic AI” really requires and why most AI rollouts fail: the model isn’t stupid, it’s blind to how the business actually operates. He shares why he wrote a new executive-focused book, Company-Scale Agentic AI (a free follow-up to Production Grade Agentic AI), and outlines a five-part loop—observe, understand, build, run, compound—built around an always-on “company brain” that absorbs emails, chats, calls, and files to map processes, relationships, and bottlenecks. Ran emphasizes the difference between deterministic automations and true agents, argues for starting with human-gated execution to build trust, and highlights role-based access control via middleware as essential for organization-wide deployments. He also describes a browser “morning brief” workflow that keeps tasks from falling through the cracks and urges teams to adapt AI to existing tools instead of forcing employees to change how they work.

00:00 Welcome and Topic
00:29 Why I Wrote It
02:20 AI Is Blind
06:45 Building Company Brain
07:28 The Five Step Loop
11:15 Automation vs Agents
13:53 Gated Readiness Dial
17:01 Role Based Access
20:58 Morning Brief Extension
22:34 Meet People Where They Work
24:57 Loop Recap and Wrap

-----

Old School / New Tech is produced by VarOps.

• Subscribe to the newsletter – the written version of this show, weekly: varops.com
• Read the books – Company-Scale Agentic AI (for operators and the C-suite) and Production-Grade Agentic AI (for engineers): varops.com/books
• Work with us – VarOps installs resident AI inside mid-market companies. It runs where your work already happens, sees the whole operation, and what it learns belongs to you: varops.com/transform

Creators and Guests

Host

Ran Aroussi

Building software and AI for companies that can't afford to guess | Automaze •VarOps

What is Old School / New Tech?

Old school thinking meets new tech — an unfiltered podcast, hosted by Ran Aroussi and a co-host you won’t see coming.

​Live every episode. No production lag, no editing. Business, AI, tech, startup ideas, and whatever's worth talking about — through the lens of someone who's been building production systems for 35+ years

SPEAKER_00: Hello, hello.

Welcome to another episode of
Old School New Tech.

Thank you for joining me.

My name is Rana Russi, and today
I want to talk to you about

company scale agentic AI and
kind of the different things

that I've been seeing and I've
been experiencing in part of our

work with different companies
when it comes to uh really

agentic scale.

So the reason for this is I
wrote a book, another one.

It's called Company Scale
Agentic AI.

It's I would say that it's an
executive version of my previous

book, which is Production Grade
Agentic AI.

And the previous book was for
engineers.

It was about 650 pages long, and
full code, the whole thing.

And it did its job.

But the message that stuck with
me the most when I wrote it and

kind of received feedback was
actually from uh the people who

signed the engineers' paychecks,
which are the executives and all

kinds of uh founders, CEOs,
CFOs, uh those types of people.

And they also admitted that I
could have written in more plain

English for the non-techie
crowd, so that's why I went

ahead and created and and wrote
this new book, which will be

start to be available in the
next few weeks, probably first

weeks of August, I assume.

And it's about how the company
from the company perspective

should look at agenda
corporations and the different

things that they need to do.

So that this is kind of what I
wanted to talk about.

The this is not a pitch for my
book, just like my previous

book, it's gonna be free.

Uh, you'll be able to download
it and get it for free.

So I'm I'm not trying to get one
on you.

Obviously, if you want to, the
printed version is gonna be

available on Amazon, but PDF and
uh EPUB are gonna be available

for free on my website.

So sorry.

So here it is.

The main thing that I've noticed
with different companies that

I've both worked with and
advised with, and kind of came

on board to try to analyze what
the problem is, is uh the AI is

not stupid.

Okay, it's blind.

And uh that that's kind of the
reframe of the whole thing.

A lot of the time, a lot of the
times executives and even

engineers tend to blame the
intelligence.

They you know they come up with
things like, oh, the the model

is no good, or it's not strong
enough, or the workflow is not

smart enough, it's not
optimized.

And I think that yeah, it could
be the case where it's not smart

enough or it's not uh optimized
enough, but more often than not,

it's blind rather than stupid.

It means that it doesn't have
access to how the company

actually operates, and the
whoever implements the AI in the

organization, it usually happens
in the form of training or in

the form of some sort of let's
teach you how to use cloud code

or sorry, not cloud cloud, but
cloud desktop or chat GPT.

So this is where it usually
ends.

But the problem is that it's
very disconnected.

There's no obviously there's no
company brain, even if they they

think they have one, it's not a
company brain, it's an

assortment of information that
different people have at their

favorite chat companion.

So it's not really a genic.

Most of uh, you know, if if
you're seeing this and if you're

looking at the listening or
seeing this episode, you are

probably more belong more to the
tech crowd, so you're probably a

bit more advanced, or you think
that everyone and their mothers

are using AI and workflows and
loops and graphs and all that

cool pipe words that are going
on at the moment.

But the fact of the matter is
that in most organizations and

in most large-scale companies,
especially companies that are

not startups, they're not new
companies that have been around

like 10, 20, 15 years, not to
say 30 or 40.

The AI, if it's even being used,
it's being used in the scope of

firing up Chat GPT, asking a
question, and moving on.

So don't assume that everybody
is using AI just yet.

So what I'm noticed happening is
that you know, the more advanced

people of those companies that
are not using just chat or cloud

desktop, what they do is they
tend to start a project which is

already a step in the right
direction within those chat

applications.

They load it with all sorts of
information about how the

company works and different
policies of the company,

different types of call
transcripts that they had, which

is great, but it's not enough.

You know, every company has that
one person that knows everything

and it's already like a second
nature, it's not really a

well-thought-out process.

You know, Jane knows how to do
XYZ.

And if you ask Jane to kind of
lay out how she's actually doing

that or what helps her make the
decisions, she's gonna write it

down.

But I bet that in most cases
it's not gonna be accurate

because some of the things, if
not most of the things that

she's doing, uh she's not even
thinking about them.

She's just doing that because
she's been doing it for 10 years

or 15 years, she's she's been
doing that.

So some of the things, you know,
they they only live in the COO's

head, or they're a decision that
happened on a WhatsApp thread,

or uh a decision that happened
on a call, on a phone call, and

nobody bothered to log it
anywhere.

So that's what I mean when I say
the AI is blind, it's not

stupid.

And the road to fix it starts
with fitting it all the

information in a loop.

Yeah, I know it's a hype word
right now, and that's not even

the type of loop that I mean.

What I mean is an ever-present
and ever growing company brain.

So let me show you an example of
what I mean.

So this is uh the book, but
that's not what I wanted to show

you, so let me start sharing the
screen.

Yeah, so that's the loop that
I'm mostly interested in showing

you.

It's a loop that has five pieces
to it.

One is the observe part, and
this is kind of a snippet from

from my book, but let's see if I
can make it even larger so

everybody can see it.

All right, yeah, yeah, that's
much better.

So this is this is the loop.

You start with observe, meaning
you have to install a ghost in

the machine, you have to install
an AI agent that lives on Teams

or Slack, obviously, not in
private messages unless it was

invited.

It's BCC'd to email that are
going on in the company, or even

better, hooks up to the
company's email server.

It's also a resident on the
company's WhatsApp group or

iMessage groups.

It's there whenever there's a
call transfer, sorry, a

conference call that's
happening, or even a phone call

that's happening using tools
like Plot or Pocket or these

types of uh widgets that clip
onto your phone.

So it needs to know everything.

Obviously, it needs to sit on
the company's Google Drive, and

it just needs to absorb
everything that the company is

doing.

This loop on a smaller scale
needs to happen on a per

employee as well level and on a
per team level.

But right now, let's talk about
the company as a whole.

So it needs to observe
everything and it needs to

understand and distill how the
company actually works.

So if you go to a manager and
those companies and you're gonna

ask, you know, how where are the
bottlenecks?

What's the process of doing X,
Y, and Z?

How long is something taking?

They're gonna answer wrong most
of the time because there's a

difference between where you
think the bottlenecks are and

knowing where they really are,
where time is being spent the

most.

So that's the second part of
that loop is to understand

everything that's going into the
company and to build its own

company brain and its own
knowledge graph of how the

company works, its processes,
the relationship between people.

You know, Jane usually calls
upon Mark to do XYZ, and then

Mark sends an email to whoever
it is, and so on and so forth,

and just build that uh knowledge
map of information.

The next thing that the next
part of the loop is to build.

So now that we understand
everything, that system needs to

come up with its own suggestions
on what needs to be automated,

at the very least, a report of
where the bottlenecks are, what

needs to be automated, where
things usually kind of fall uh

through the cracks, you know,
stuff like that, and either have

someone on the company build
that workflow or start building

its own skills, and then the
next phase obviously is to run

it, and then this stays in a
loop that the final step is the

compound.

But if you notice here on the
run part, you can see that

there's a dial for readiness.

So, first off, even though I'm
kind of an AI guy and I'm all

in, and this is uh the topic of
this conversation, I want to

kind of make sure that we
understand, and this is the I'm

looking for a graph here.

I'm gonna find it, it's
somewhere here.

Yeah, all right.

I want to first make sure that
we all have the distinction

between an automation and an
agent, and this is another graph

that that is important.

So too often I see kind of the
run to fully autonomous agents,

and in most company and business
environments, what you need is

is an automation, you which
means it's fixed path.

You know, whenever a support
email comes in, the next step is

to gather all the information
about the client or the past who

the account representative is,
the path, the past

correspondence with that with
that client, what products they

bought, so on and so forth.

So that's you don't need an AI
for that.

It's just gathering the
information.

You can go directly to your CRM
with an API and build this.

And then you need to understand
the problem and answer the

client.

Or if it's a refund, just make
decide on whether or not you're

making a refund.

If it's a very deterministic
step, you don't need AI at all.

And but in most cases, you do
have an AI block within an

automation.

So writing the email, making a
sense of what the customer wrote

in the support request.

So that those are the types that
are more fuzzy, and they can

benefit from using AI.

So it's it's not like here's a
customer request, let's throw it

into an LLM and hopefully can
figure things out because I've

loaded it with a bunch of MCP
tools and I'm just gonna pray

that it's gonna work.

No, there are some
deterministic, mostly

deterministic steps before the
AI gets involved.

An AI, on the other hand, is for
the fuzzy parts for the things

that are non-deterministic.

So if an urgent email arrives in
that regard, so the first thing

that we might want to do is if
there's an urgent in the

subject, or maybe this entire
block can be part of this step,

is who should we redirect the
email to?

We need to understand from the
context of the email who is

responsible, and do we need to
get back to the user with more

information, or can we
re-resolve it without even

escalating it to a human, and so
on.

So that's a very important
distinction.

And the reason why I wanted to
cover this before going on to

the next step is to talk about
this style.

Most of the operations,
especially in the business

environment, need to need to
start off as gated, meaning the

AI and the agentic stuff is
going to do, or the agentic or

the automation is going to do
most of the work, but it's gonna

leave you with the final step.

So it's gonna leave you with a
draft email, for example, or

it's going to leave you with an
order that's already filled and

waiting for you to review it.

So this is a very important step
because in a business

environment, obviously, one you
know, wrong move can cost the

company from tens of thousands
of dollars to millions of

dollars, depends on the company
size and scale, etc.

So that's it that that's a very
important thing that you you do

not rush to full automation when
um during that time where the

company, you and other
stakeholders in the company need

to build trust in the fact that
the system can actually perform

well on its own without any
human in the loop or human in

the gate.

These are two different things.

Here I'm talking specifically
about human in the gate.

So that's probably one of the
more important things is not to

rush into the automation because
if and the goal, by the way, of

the gate is not to get a draft,
see that it's not exactly what

you wanted, then edit it
yourself and then send the

email.

No, the goal is for you to see
that if two drafts in a row are

kind of not correct or not in
the spirit of the company or not

according to the policy or
whatever it is, then you go back

to either the skill or the
process or whatever automation

or genetic automation that
you've created, then you go back

and fix that.

You fix it as a root cause and
not at the um not at the output

level.

Okay, so that's that's kind of
something that's that's very

important.

And as the company learns more,
or the company's uh brain, this

entire loop keeps learning from
how you're fixing the drafts and

how how many times you had to
cancel an operation, it keeps on

learning and keeps adjusting.

Now, another thing that is very
important, and and I won't don't

don't want to go into too much
technical detail because you'll

as I mentioned you're more than
welcome to go and uh download

the book.

You know, in fact, I think I
created like a widget here.

Here you go.

So here's a widget.

Just scan that, you can go
download the book, and obviously

I'll add the link in the show
notes.

But the the issue of role-based
access management.

Another problem that is
occurring quite often with uh

companies when they implement a
genetic process or let's just

call it the implement AI into
the organization, is they are

bound to not implement it at the
organization level, but they're

doing it more on the department
level or on a personal level.

And the reason is that if I'm
connecting to the CRM or the ERP

or the accounting software, if
I'm connecting to that software

with privileges of the CFO, I
should be able to see

everything.

But someone from I don't know
from the development team or the

HR team should not have access
to the company's financials or

the company's legal team, and
building those gated systems

where every person in the
company have possibly access to

the same tools but in different
access levels, that is something

that is very hard to implement.

And one of the things that I
recommend that companies do is,

and there's really no way around
it, is to build a custom

solution, which is something
that I've done with uh with

Moxie, my own open source tool,
is to build a mechanism for a

middleware, that's the word I
was looking for.

So to implement a middleware
that automatically removes

different tools or adds
different tools that are

available to the agent based on
who's talking to it.

So some people will have access
to rough numbers.

What are the growth or the trend
in the company's financial?

So they're gonna get it in
percentage-wise all only because

they need to build their uh
sales prediction or sales

pipelines or whatever it may be.

Some will have access to the
full information on a per line

item basis, and some will have
no access to this tool.

So by implementing that
middleware that sits between the

user and the brain, and kind of
filters down the different tools

and levels of of what can be
done with the system based on

who you are, and it's also
recommended that that you we

know who you are, where whether
you're contacting the system

using email, using Slack, using
uh chat-based interface, or

anything like that, that is also
an opportunity for the

middleware to take to take
advantage of.

But filter the access level
right then and there before the

the request even gets to the
system.

So it's very hard to do.

It usually takes you know
integration with tools like

WorkOS, which in turn needs
integration with the company's

Active Directory.

And yeah, it's it's it's not an
easy task, but if you have any

chance of implementing an AI
system that you can feel

comfortable with, and on an
organizational level, I'm

talking about hundreds of
employees, if not thousands of

employees, then you have to have
that role-based access

management ability.

Let's see if I can think of
another thing.

Oh, yeah.

One other cool thing that we uh
we had an opportunity to

implement with one of our
clients is to build, they wanted

it as a new tab in the browser,
just a simple Chrome extension,

but the implementation is the
same wherever it is.

So whenever they open a new tab,
it automatically connects to the

system which pulls information
like a morning brief, but that

gets updated every time they
open a new tab.

Obviously, with some sort of a
caching, we don't need to

literally fetch new information
every second, but every few

minutes they're gonna get that
information refreshed and they

don't have to worry about you
know forgetting to get back to

Sarah on Slack or that they have
an email waiting.

It's gonna be right then and
there.

And unlike a to do app, because
that's what basically that

that's how it kind of looks
like, unlike a to do app, when

you click on it, it
automatically takes you to the

right email or to the right
Slack thread.

And the only way This can get
checked off is by you taking the

action on the Slack email,
WhatsApp, whatever it is, or

delegate it to an agent saying,
okay, go back to get back to

Sarah and tell her that blah,
blah, blah.

So that's the only way to take
off that task.

And it's and it's nice because
you don't have to worry about

things falling through the
cracks because you always have

that being that collects
information about everything

that's on your plate at any
given moment, and you tend to be

a lot more efficient.

And that may, you know, that
slightly takes me back to my

final point before I close on,
is most AI kind of

transformations or consultants
that I noticed, they tend to

force the company to reshape
itself or its workflows around

how the AI works.

Yeah, you need to move this
workflow into cloud co-work or

you need to start using this
tool or that tool.

But you know, that 60-year-old
HR person, they're not going to

change a workflow that they've
had and they've been using for

the past 20 years.

So you can either fire them,
which I don't know if it's the

right thing to do if they're
there for 20 years.

I don't know if it's in your
best interest.

You can force them to adopt new
tools, which will probably cost

you in time wasted, morale, and
not at all sure that they're

actually going to stick to it.

Or you can shape the AI to work
where they work.

So if they like to keep working
with their outlook, then make

sure that they can email the
company's AI if that's their

preferred way of talking, or if
they're already on Slack or

Microsoft Teams, that that's
where they're gonna have that

interaction with AAI.

So they wouldn't have to
download new software and get

used to new things because the
time to adoption, yeah, you can

with new hires, with the newer
crowd, or with the younger crowd

in the company, you you can
probably say, hey, that's uh a

better tool.

Let's let's start using Buzz,
for example, over Slack.

But most of the company is still
on Slack, so you'll need to

build some sort of a bridge
where the cool kids are hanging

in in Buzz and using agents to
do all sorts of cool stuff, but

the rest of the company keeps
using whatever it is that

they've been using for the past
10, 20, 30 years.

So, yeah, that's kind of what I
wanted to cover today.

And again, it's not a pitch to
my book, it's it's completely

free.

Where is that graph that this is
the loop I want to leave you

with?

And I think that this is the
most important part of the whole

thing is we need to start with
observing, make sure that

there's always an observant and
observation process that's going

on and lives inside the company,
and it needs to understand.

And when I say design the
company, it's I mean everywhere.

Because if one decision was made
over a phone call and that was

an important decision, and
nobody followed up on it on

email or whatever, the that
ghost listens on the the it will

take the wrong decisions.

So you need to make sure that it
knows everything, and that's

basically the the process.

You need to observe where where
did it go?

You need to observe, understand,
make sure there's uh

distillation, uh distillation
process, I think I'm saying it

right, and then start building
around the conclusions that the

and the reports that the AI has
been giving you, and then it

just compounds every time you
learn and you make a new

decision.

The company the company brain
and the company's ghost, for

lack of a better word, need to
adjust, need to adapt, need to

update its its own processes,
skills, or whatever it's

decision making.

And this might be a good
opportunity to kind of make a

mental note to have another
episode where we can probably

talk about a company-specific
LLM that you fine-tune that is

built on top of all of that.

So decisions are happening on
the LLM level rather than on the

harness level, or even better,
both.

But yeah, it's out of scope for
what I wanted to talk about in

this episode.

So, yeah, that's that's pretty
much it.

And if anyone has any questions
or feedback about this thing,

you're more than welcome to hit
me up on LinkedIn, X, email, you

know where I am.

And yeah, thank you for spending
this time with me.

And I'll see you on the next
episode.

Bye-bye, everyone.