Building an AI Startup: What's Different in 2025? | #871 - CXOTalk

Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
We're discussing AI startups today on CXO Talk episode 871 in
a conversation with Arvind Jane,the founder and CEO of Glean.
He was a distinguished engineer at Google before starting and
taking public cybersecurity company Rubric.

(00:22):
Glean has raised $600 million and is valued at almost 5
billion. Glean is an enterprise AI
company. Think of it as Google or Chat
GPD but inside your company. So that's what we do.
And you are a, shall we say, born in AI or AI native company.

(00:44):
So you're the perfect person to help us understand what's going
on with start-ups, AI start-ups today.
Absolutely. We are in fact I think the first
company to bring the transformertechnology to the enterprises
and, and and say, yeah, it's a lot of great learnings, you
know, starting as a native Gen. AI company six years back.

(01:05):
Gen. AI company for six years.
That's well before the marketinghype surrounding ChatGPT and
Open AI. That's true.
In fact, the term Gen. AI did not exist at that time.
But the core technology that powers Gen.
AI, the Transformers, they were there and we were using them,
you know, in early 2019 to, you know, understand enterprise

(01:29):
data, knowledge, information at a really deep level using AI and
then making it all searchable for people inside a company.
Can you describe the differencesbetween AI startups and
traditional technology startups?Every company like, you know,
first of all, whether you're a startup or you are a mature

(01:51):
company, I think AI is becoming such a core fundamental tool
that you've got to use in, you know, in products that you
build, because that's, that's the way to sort of stay ahead.
That's the way to actually buildnew amazing things.
So, so I actually feel like, youknow, like this.
My view is that probably all startups that you think of
today, you know, that you can call them all AI startups.

(02:13):
The but, but maybe the other wayto look into look into this is
that there are some companies, you know, that are actually
building core foundational technologies.
For example, they're building models, they're building
infrastructure to actually trainmodels.
And maybe, you know, that's whatI'm like, you know, you know, I
would say like, you know, one set of companies, but then the
vast majority of AI startups that are getting started today

(02:37):
are companies that are thinking about solving a business
problem, a consumer problem where they feel that AI and you
know, the new reasoning and generation capabilities is going
to play a big role in, in the product that they're going to
build. The but like, you know, like us
living in the AI world, you know, all the, all the startups

(02:59):
that I, you know, engage with, like I'm not, I've actually not
seen a single startup that is not, you know, that that's not
actually making AIA big part of the core tech stack.
As AI has matured, there's become a greater clarity that AI
as you just described must support the solving of some
business problem. For AI startups where the the

(03:25):
centerpiece is that AI technology, what are the
differences from again traditional software companies
where yes, they're solving a business problem, but then but
the technology is very differentthe underlying foundations.
AI technology, you know, moves very fast, it changes, you know,

(03:46):
at A at a very rapid pace. So when you think about product
development, the the new AI startups, you know, they're
they're able to you know, the first of all, they're very lean.
They they are able to actually do a lot of things, you know,
because software programming, building systems with AI, you
know, these things are becoming,you know, much, much easier than
than you know, than in the pre AI world.

(04:09):
So, so I think, I think the like, again, like, you know, I,
I, I, I struggled with your question because I, you know,
like, like my, from my vantage point, like when you think about
investment that is happening, you know, most of the investment
from venture capital actually iscoming into companies that are
making AIA big part of their story.

(04:31):
And, and so I can compare more with like the, let's say that
startups, you know, that we're getting started two years back
versus today. The key difference between them
is, well, like thinking about this new model where you can
truly build products, you know, which are a lot more powerful,
you know, a lot more capable than what would what you would

(04:53):
have even, you know, you know, thought of like, you know, a
couple couple years back. But but but I search like, you
know, I think from a company building perspective, like, you
know, startup is a startup. You know, we are AI company and
like you know, when I compared this with my previous startup, I
would say that like, you know, most of what we do doesn't
change. Like, you know, we, we still
have to think about the core business problem.

(05:15):
We have to figure out how we're going to actually solve that,
how we're going to build great technology and and like, you
know, the teams to to then sort of like, you know, bring that
technology to our customers. So so personally, like I have
this feeling that yes, like, youknow, AI is becoming a big part
of our technology stack, but fundamentally how we build and
design companies are not changing except for maybe one

(05:36):
thing like, you know, you do here.
Sometimes companies saying that,well, like, you know, now with
AI, you can actually have a one person company that can generate
a billion dollars because, and then so you see a little bit of
that, like, you know, like, you know, in the new generation of
companies, you are seeing some of these, you know, AI companies
like scaling up revenue at the pace that, you know, the

(05:58):
previous industry of SAS companies couldn't.
Like, you know, we have, we saw some, you know, I, I did these
examples all the time, you know,like start up actually reaching
$20 million, you know, in revenue two months, you know,
like after it's got started or acompany that's, you know, at 100
million rate, like, you know, within the first year, these
things are not possible. Like in, you know, in many ways,

(06:19):
you know, in the Pai world, But given like you know how fast,
like you know, you can actually build products and, and how
different your products can be compared to the, you know, the
current products in the market. It's allowing people to actually
just fundamentally scale at a different pace and level than
than the previous generation of start-ups.

(06:40):
I just want everybody to know that you can ask your questions
right now on Twitter. There's a tweet chat, X on XI
should say there's a tweet chat and X chat taking place.
Use the hashtag Cxotalk if you want to ask your questions.
If you're watching on LinkedIn, just O your questions into the

(07:04):
chat. Arvind, you raise a very
interesting point just now, how startups, AI startups can
generate so much revenue with relatively few people.
What is it about AI, this technology and the nature of the
startups that are happening around this to enable this

(07:26):
phenomenon? So first you can actually build
products that are very differentin their capabilities.
Then like, you know, then thingsthat are out there, when you
think about, for example, software development, there are
companies, you know, that are offering new development

(07:49):
environments, which allows developers to go five times
faster than what they could before.
And, and so there's a, there is such a big leap in, in the
capabilities, you know, that these, you know, that these
products bring, which creates that instant, you know,
excitement in the market for them.
Like, you know, like, you know, these, these new coding tools,

(08:10):
like, you know, even even for us, like, you know, as a native
Gen. AI company, like we, we see it
everyday. Like, you know, people just
like, you know, they, they want to use these tools.
Like, you know, there is a, like, you know, there's, there
is a big, you know, demand, you know, from the ground up, like
there's love for these new AI products.
That is, that's incredible. And that is what's allowing them

(08:30):
to, to actually like launch a product in the market, like, you
know, have a PSG motion and, andlike, you know, there you go.
Like, you know, people come in, they want to use these products.
So it's, it's, I think largely it's driven by these brand new
amazing capabilities and software doing things that
people didn't expect, you know, it could do.
So I think I think that's the biggest part.

(08:52):
Like that's what plays that excitement and then instant
demand. But the other thing which which
is actually allowing these companies to have this success,
you know, you know, in a short duration of time is that, you
know, the software development itself has, you know, gotten
very accelerated. When you use these new tools,
you can build products which area lot more capable than, you

(09:16):
know, current products that we have in the market.
And, and you can do that in a short period of time.
Like you actually put something amazing, you know, in a month,
you know, in two months, you know, something that used to
take you like a year or two years before.
So that's that's the other factor that is actually, you
know, shrinking these these cycles for these companies.
So you have this combination of tremendous interest and market

(09:41):
demand combined with significantdeveloper productivity increases
as a result of the tools. Is that a correct way of saying
it? That's right, yes.
We have an interesting question coming from LinkedIn and this is
from Santosh Sirivo, who says who is actually using and paying

(10:06):
for AI services. AI has captured everybody's
imagination. There is no enterprise in the
world today that feels that well, like, you know, we can, we
can, we can look into AI like, you know, a few years from now.
Everybody knows that they have to act, they have to embrace AI

(10:29):
today if, if you want to be, youknow, if you want to stay
relevant. And so, so starting from that,
starting from that, like high level desire to actually, you
know, bring AI into your enterprise.
Now what we're seeing is that there are two different ways our
people are bringing AI into the company. 1 is the CIO, you know

(10:52):
that, you know, CIOs like, you know, obviously hold the, the,
the responsibility to bring the right set of technologies, you
know, to, to the entire working population of your, of your
enterprise. So they are taking the
leadership in terms of bringing,you know, bringing godly,
applicable, useful products thatare AI based to their

(11:16):
enterprise. And that's for, for us, for
example, at Glean, you know, youknow, they are our primary
personas too, because Glean as atool, you know, it's, it's a
knowledge access tool. Like people go to glean, they
ask questions, you know, we quickly answer those questions
for them using all of the enterprise context and data and
knowledge. And so that's the tool.
That's actually every knowledge worker, like, you know, whether

(11:38):
you're engineer, the support person, somebody in HR or IT or
sales, all of you have the need for that.
All of us have questions, all ofus have tasks that we think AI
can do for us now. So it's a broad tool and CI OS
are the ones who actually typically like, you know, will,
you know, purchase a a company by two like that.

(11:58):
But then you also have every individual functional leader.
You know, if you are head of support, if you are the CTO and
you're trying to actually make sure that you know, your
engineers, your developers are productive with AI.
If you're, if you are a sales person, like you know, the sales
leader and you want to make surethat you're using modern ways

(12:20):
to, to prospect, to reach out toyour customers, to actually have
powerful engagements with them. Function by function.
You're seeing every functional leader in the enterprises
looking into and evaluating AI tools and bringing them on
board. So this is a, this is a very,
very broad phenomena, industry wide vertical, wide
geographical. Like, you know, I've been

(12:42):
travelling a lot across the world.
I don't, I don't, I don't see any company that is not paying
attention to AI, any country that's not paying attention to
AI either. So this is, you know, like is,
is everyone really is the answerlike you know, is bringing AI
into their into their enterprises.
Subscribe to our newsletter Joinour community Go to cxotalk.com

(13:02):
Subscribe to the newsletter. Check it out.
So it's following a traditional enterprise software, the
purchase process in a way because as you just described,
you've got IT and the CIO and atthe same time you have a
functional line of business leaders, HR, whatever,

(13:24):
marketing, whatever it might be.Both of these groups are looking
at AI products. And I'm assuming this comes
right back to what you were saying earlier, which is the
core issue. It's not actually the
technology, it's the business problem that's being solved.
We have a very interesting question from LinkedIn and keep

(13:47):
your questions coming in. We have some questions on
Twitter as well, and we're goingto get to all of these.
Risha Varshney asks, how do yourofferings address the AI ethics
jailbreak, prompt leakage and AItransparency slash
interpretability risks? And I'll just mention that Risha

(14:11):
is Senior Director of Risk Systems Development at Freddie
Mac. And of course, Financial
Services is keenly interested inthese topics.
So Lean is an enterprise AI solution.
We work with the largest enterprises out there in
different industry verticals. We work, we work a lot with

(14:32):
financial services and these risks are all very real.
Like, you know, I think with AI,the, you know, it's a very
powerful technology And, and it's also like a, it's, it's,
it's, you know, it's grounds for, you know, innovation and
like folks who are trying to actually create security
problems and, and you have to bevery, very careful in terms of

(14:52):
rolling the technologies in the right way, in a secure way.
And also like, you know, in a way which basically ensures
that, you know, the technology is unbiased and, and, and it's,
you know, it's doing stuff that you think is, you know, safe,
you know, for your, you know, inyour context, in your
enterprise. So some, I'll give you some

(15:12):
examples of like, you know, things, you know, problems that
you have to solve on the security front.
One of the key, you know, thingsthat AI is that if you're going
to make it useful for your enterprise, your, for your
organization, you know, you haveto take these language models
that are built using the public web data.
They don't really know much about your enterprise and your,

(15:35):
your, you know, you know, your information, your data.
And so somehow you have to figure out how you're going to
actually connect that enterprisecontext with the power of these
language models and, and do it in a safe and secure way.
For example, if you train the model and you connected all of
your enterprise data and knowledge, you know, with that

(15:57):
model and now anybody can go inside your company and ask
questions, well, it's going to leak a lot of sensitive data to
people who should not have actually, you know, have access
to that information. Because enterprise data and
knowledge is very, you know, it's it's private in nature,
like, you know, like a given document, like, you know, only a
few people may have access to itinside the company.

(16:19):
So you can't take like, you know, our content wholesale,
like you know, in your enterprise and just train or
build any I system with it. Any AI system that you build, it
has to understand who's using, you know, that that particular
system or software. And how do you make sure that
you know, AI only uses information that you know, this

(16:39):
person, you know, is entitled tois, you know, allowed to use
inside the company and create those safe and secure AI
experiences. So that's one of the problems
that we solve. Like in Glean, any AI usage, you
know, any agent that you build on Glean, you have to sort of
use that as an employee and you have to be signed in.
And then we will actually ensurethat whatever we do for you is

(17:02):
done with knowledge and information that you could
access to. I would like you to join the
Cxotalk community, so go to cxotalk.com and sign up for our
mailing list so we can notify you about upcoming shows because
we have amazing discussions likethis.
All right, let's go to Twitter and to Arsalan Khan who says, do

(17:27):
you think AI will become a simple plug and will become
simply plug and play even for non-technical employees in the
enterprise, so the the broader ease of use for the rest of us?
Well, absolutely. I'm both hopeful and confident,

(17:49):
you know, that that's what's going to happen.
I think the true power of AI will be realized when it's, you
know, so easy to use. You know, you can actually do
things with AI as as a business user, you know, you should not
be required to understand, like you know, how to code, how to
build systems. You know, be an engineer, you

(18:10):
know, you like, you know, as youknow, for example, like let's
say you are, you are an, you know, you are an, an employee in
the legal department and you review contracts everyday.
And you know, you know, it's a, it's a process, you know, that
takes you a lot of time, you know, and you should be able to
just, you know, ask AI. You can, you know, like, you
know, and say that like, look, this is, this is, this is how I

(18:30):
review a contract. This is the process that I
follow. And you should be able to just
say that, you know, to an AI system that actually then goes
and identifies, you know, that particular business process for
you, automates it for you. So AI has to truly become
accessible like that with our Asian PIC platform.
That's what we do. You know, we are actually really
thinking about how to make this technology super accessible

(18:53):
like, you know, make, you know, make sure that it doesn't matter
who you are. You could be in HR and, you
know, in finance and legal and you, you, you know, you don't
need to know anything about AI, how it works.
Like, you know what, what are language models?
Like that's, that's not like, you know, that's not what you
need to worry about. You need to just like work with
a smart. You're like working with AI
should be like, you know, working with a really smart

(19:15):
person who you could actually goand get some work to, you know,
you just tell them, you know, like how you know, like how this
work needs to happen and then AIjust makes it happen for you.
That's, that's the, that's the model that you're going to see,
you know, with any like the agent, you know, AI agent
platforms that are going to succeed in the market are going
to be of that nature. Like, you know, you have to
elevate, you know, like the, youknow, the capabilities of these

(19:37):
systems, you have to make them more accessible to non-technical
users. And the other thing that that
that also add to it like to go one more step, you know, beyond
that people are not seeking, people are not seeking help, you
know, from AI as much as you would expect.
Like, you know, we all have habits, you know, we do things
the way we do. Like, you know, like there's a
lot of inertia in terms of like,you know, thinking about like,

(20:00):
well, should I do this task differently?
Like people don't think that way.
Like, you know, you, you don't have time often.
Like, you know, most of the times you yeah, you don't even
have time to think about, well, should I change?
Like, you know, the way I work and so.
It's not only that AI has to be easy and you need to be able to
summon it and, you know, get it to do work for you.
You know, AI also has to follow you.

(20:21):
And it has to come to you and say that, look, yeah, I'm, I'm
observing that you're doing thiswork every day and spending, you
know, two hours trying to get this work done.
And I can help you with it. So, you know, it has to come to
you. And and that is when, like, you
know, you really, you know, see,people will embrace the
technology at a large scale. I'm looking forward to that day
happening because I can tell youI use large language models

(20:44):
every day and I use multiple language models and different
modes, research, non research and so forth.
And so much depends on the modelyou're using.
It seems the time of day, the phases of the moon, how you
construct your prompt, and then the whole thing's just a pain in
the butt. Yeah, it is hard.

(21:04):
Like it's not easy today. And then by the way, you are
using like what I would say probably one of the most, you
know, one of the most accessibletools, like, you know, you're
just talking to a system like innatural language.
Yes, you have a few knobs to select hidden there.
So it's, but it's going to get easier like, you know, and
that's what like our job is, youknow, at clean, like one of the

(21:24):
things that we do is so we're not an LLM company, We're not,
you know, building and training these foundation models.
But we are like, our goal is to see that you as a business user,
how do we make sure that all this innovation that's happening
in the industry, how do we make it more accessible to you and,
and make things easy, you know, as seamless as like, again,

(21:45):
like, you know, the working model for me with AI is, well,
AI is like a smart human that's been in your company from day
one. They know everything about your
company. They know all the people,
they've read all the documents, all the, all they've been part
of every single meeting and they're ready to now help you
24/7. Just ask them what you want and
they do it for you. Like that's, that's the, that's

(22:05):
the right model for when you know AI is going to deliver true
value in enterprise. Greg Walter says glean looks to
be the glue or connector betweendisparate databases.
Do you see a future where this function is no longer needed?
Enterprise environments are so complex today.

(22:26):
You know, any large enterprise, you'll get like, you know,
thousands of systems, you know, the data depositories,
databases, you know, unstructured, you know, data
depositories, you know, documents and the like.
The true magic of AI happens when you have access to all of

(22:47):
that information across all of these different systems.
Maybe like, I think things are actually going to get simpler in
the sense that, you know, AI is going to get smarter and
smarter, like in terms of, you know, being able to connect with
all of those different systems over time.
And, and a lot of things that wedo today at lean, like, you
know, which we have to actually do a lot of hard work for to

(23:08):
actually connect with these different enterprise systems.
Like we do expect, you know, it to get easier over time.
So yes, like, you know, some like, I think that's the, you
know, if we're, if we, if the answer is that like, you know,
the complexities are not going to go away, you know, then I
think he is not doing his job. Dave Brace on LinkedIn asks, is

(23:29):
it likely that non deterministicagentic systems can honestly
prove that they are truly trustworthy in the enterprise
and that business leaders can trust products like Glean to
make thousands of business decisions every day?
AI is non deterministic. It can also be wrong.

(23:53):
It can hallucinate and to some degree, like, you know, humans
are also like that, like, you know, if you, you know, if you
have questions and you go and ask somebody, like sometimes,
you know, they don't have the full context, they're going to
give you an answer and it may bewrong or it may be incomplete.
And, and so I think the, it's, that's, that's, that's the
fundamental thing to remember that, you know, these AI systems

(24:17):
are actually more like humans, you know, and, and less like
machines, you know, that they used to and, and now you have to
figure out like, well, how do I use this technology?
Like this is not perfect, is going to make mistakes.
So how do I trust it, you know, with, you know, my mission
critical processes where precision is actually, you know,
absolutely required. And so, so there are a few

(24:40):
different strategies that I would say that like, you know,
you know, as an enterprise, you can think about there's a lot of
work where you don't need precision, you know, you know,
work where you need creativity. And that's where like this
technology is already really, really well suited to do.
But then when you think about like your task with that require
precision, AI can be actually used in a few different ways. 1

(25:02):
is that, let's say that there's a business process and now
you're going to agentify, agentify that business process,
you're going to automate it. You're going to ask AI to, to
understand that business processand come up with a plan with the
workflow to, to actually executethat business process from now
on. And so when you, when you use AI

(25:22):
in this fashion, like, you know,you will go and you know, ask AI
to actually build that agent foryou and like, maybe to make
mistakes and but you know, you should, you should be there.
Go and supervise it like, you know, go and like ask it to
tweak, you know, it's work, you know, go manually fix, edit it.
And ultimately in that collaboration with AI, you
actually encode and build that agent.

(25:45):
You know that workflow. But now this workflow is
deterministic. Like you, you, you put some
investment in it. AI helped you like, you know,
build this workload very quickly, but you were totally in
control, you were monitoring it.You've gotten into a coordinate
into a place where now, now as Isaid, it's deterministic.
And now this business process can actually run and it is OK.
Like you know, you don't, you don't need full automation.

(26:06):
Like you can actually work with AI, you know, and spend like you
know that, you know, initial time like an hour or two, but
now you're going to have like, you know, this automation for
years because and it's no longernon deterministic.
So, so you don't think about it like the technology allows you
to to know great things and you don't have to rely on AI to
actually make complex decisions.You know, behind the scenes you

(26:29):
can actually work with it that initially and and build
deterministic systems with it aswell.
So you're describing essentiallythe role of that phrase.
We often hear the human in the loop and what's the appropriate
relationship between the the theperson and the AI system that at

(26:50):
this stage of development is a tool rather than I'm looking at
LinkedIn and Greg, Greg Walters says the true magic of AI
replacing old standard applications.
Let's take some examples. You know, we initially were
talking about a legal person that reviews contracts.
Now, if you get 100 page, you know, you know, agreement,

(27:14):
customer agreement, and it's going to take you like, you
know, a week or two weeks to actually go and review this and
make sure that you know, all theterms and conditions, you know,
meet, you know, your enterprisesrequirements and you're to
redline this document. And and you didn't say that
like, you know, well, I don't trust AI to do this for me.
Like this is pretty sensitive stuff, but well, you know, get

(27:36):
AI to do the first version of the redlining and you know, it's
going to do a great job. Like, you know, if you just tell
it like how you do it, you tell that to AI, it's going to get
90% of the way there. And if it gets 90% of the way
there and now you can actually, you know, like fine tune that
and finish that work, you know, with the context and know how
well, like, you know, that two week task, now it's actually,

(27:57):
you know, a one day task for you.
And that's, that's, that's big. Like, you know, you don't have
to, you don't have to actually like aim for 100% automation and
like, you know, remove yourself from the task completely.
Like, you know, there's, you know, we can, you know, I'm very
happy if I get 90%. You know, it's a big impact to
the business. This is from Gersheron S on
LinkedIn who says which agents are driving the most value in

(28:23):
large enterprise functions. Even better if you have examples
in an industrial AI context. And I should mention that
Gershwan is an AI product manager in metals mining.
So which agents are driving the most value in enterprise
functions? Top use cases for AI today in

(28:46):
the enterprise are the followingthree.
Number one is general knowledge access and assistance.
So you know, you are a knowledgeworker, you, you may be in, you
know, healthcare or financial services or you know, industrial
sector and you have questions, you know, you have questions

(29:07):
that you need answers to. You have information that you
need to do your tasks and, and you use AI to actually, you
know, help you with that. So like tools like ChatGPT or
tools like clean, like inside your company that are just
general purpose, they're not meant for a specific use case.
They're basically knowledge tools like, you know, they help
you, you know, they make knowledge accessible, you know,

(29:29):
from the world, from your enterprise to you so that you
can move faster with, you know, with your tasks.
That's the number one use case for AI today.
And it's not surprising because like this whole revolution was
catalyzed by chat GPD. And so that's, that's the,
that's the application that comes to everybody's mind when
they think about AI. The the second use case is, I
would say for, for software development, which is around

(29:52):
code generation, like developerstasks, like how you build
systems, you know, right, you know, develop technology that's
fundamentally changing with AI. So that's, that's a very
powerful use case. Lot of correction there.
And then the third one I would say is, you know, is around
service. When I say service, I mean like,
you know, I know folks who are actually like taking questions,

(30:13):
taking, you know, complaints tasks from their customers or
their internal employees. So these are like customer
service teams, IT, you know, internal IT help desk, you know,
HR help desk, folks like who areservicing other people's
requests and demands, like, you know, you know, taking, taking
your company's data and knowledge and automating a lot

(30:34):
of that interaction with AI. That's the, that's the big use
case. So functionally those are the,
those are the three top applications and, and, and I
think like across industry verticals like those, that's
what we're seeing. So like I'll give you some
examples on customer service. So we have like these large
telcos, you know, that have like, you know, 50,000 or even

(30:54):
100,000 customer care agents dayin day out.
Like, you know, they are gettingquestions from their customers
that they need to quickly answer.
And you know, if you, if you make those transactions twice as
fast, like, you know, that's, those are hundreds of millions
of dollars of savings, you know,for those teams.
So like that's a very powerful use case for AI today for for
glean, you know, we already talked about software

(31:17):
development, like, you know, like, you know, we have large
enterprises in industrial sector, in retail, you know,
where you are fundamentally changing how, how you use AI to
build systems faster, to test them to, to review, like, you
know, code to troubleshoot. These are some of the, you know,
key use cases as well, like, youknow, for, you know, across the

(31:40):
industry. This is, again, from Arsalan
Khan on Twitter. He says AI might be able to
automate standard operating procedures, but what about the
institutional knowledge that resides in people's heads?
And he says, isn't this a significant job security issue?

(32:03):
Today AI like even with Glean, we are able to actually tap into
a lot of their institutional knowledge to then create these
powerful like, you know, GPT like experiences where you can
come and ask questions and you can answer that using, you know,
that institutional knowledge. Then think about institutional
knowledge like it's actually present in a few different form

(32:25):
factors. You know, you have documents
inside your company, you know, where people like, you know,
write stuff, you have ticketing systems, you have like
databases, you know, CRM systems.
So there's a lot of wealth of knowledge inside each one of
these different systems. But then there's also a lot of
knowledge and communication tools like, you know, e-mail or,

(32:47):
or Slack. And there's an, an increasingly
like, you know, enterprises are changing behavior so that they
can capture more and more of that institutional knowledge.
One example of that is that, well, like, you know, if, if two
people are going to actually talk and have a meeting, you
know, record that meeting or at least, you know, like capture
the summary of, you know, that what happened in that meeting

(33:09):
and make it available to AI so that, you know, it can be, you
know, tapped into in the future.So, so I, so I think like so, so
that's, you know, these are all the things that you can do
today. Like, you know, we, we capture
all these forms of institutionalknowledge to then actually make
that knowledge work for you as an individual and help you.
And, and like, you know, is driving that.

(33:30):
Like now people are also more motivated to, to actually
capture this information more inour company, for example, like,
you know, we now record like every non, you know, one-on-one
meeting. Like, you know, I could say, you
know, confidential meeting with between the employee and a
manager. We won't.
But like, you know, if it's about, if it's about, you know,
a technical discussion, it's about like getting some work

(33:52):
done, you know, we'll typically record those meetings so that
like all of that data is available to AI in the future to
help us. So that's but but now coming
back to your second question on,well, does it actually, you
know, create an issue with job security?
Like, you know, we, you know, I believe that, you know, the like
as an individual, the strategy for you to make sure that you

(34:14):
stay relevant is you don't go and learn AI like you don't like
this. These tools are amazing and
don't think of AI as something that's going to take a job away
for you from you. I don't think, you know, AI is
that powerful. I don't think it's going to do
it for most people, but you'll certainly like, you know, lose
edge against somebody else who knows how to use these AI tools

(34:35):
and work faster and better than you.
So, so I think that the, the, I,I think that what, what as
individuals, what we need to do is like, you know, like, like,
like, like, like this have been always like that.
Like you know, when new technologies come, folks who
embrace them, like you know who quickly try to learn them, are
the ones who come out ahead. I think this is a very important
point and I certainly give people that same advice that

(34:58):
you, you must be learning how touse AI to make yourself faster,
better, more efficient. And there are so many jobs that
will get displaced. So if you're listening and
you're not doing that, you know,I'm assuming if you're listening
to CXO talk, you are doing that.But so tell, tell the folks you

(35:19):
work with who maybe are not so far on the cutting edge, give
them that, that advice. It's good advice.
So let's talk, come back to startups and we have a question
from Twitter directly on startups.
And this is from Gus Bechdash, who says, Arvind, what advice do
you have for people forming AI ventures?

(35:42):
Many make the mistake of choosing problems that the
platforms will take over, makingtheir companies irrelevant.
I actually don't believe, first of all, that, you know, if you
start a company, if you start tosolve the problem that you're
going to fail because of somebody else, because of like,

(36:02):
for example, an incumbent or a large company actually solving
that problem and solving them before you.
You're going to have a firm belief like as an entrepreneur,
that you can solve the problem faster than a large company.
A large company always has lots and lots of things to actually
worry about. Do you know, they're
structurally not designed in a way that they can move faster

(36:23):
than you like as a really, you know, as a nimble early startup?
So so I don't actually agree with that premise of like people
making that mistake. Most of them startups fail
because people give up like, youknow, because you lose
confidence in your own idea. You don't have enough conviction
because like, you know, if there's a real problem, well,
you have the right to go and solve it and and you will be and

(36:45):
you know, competition is not something that's going to
matter. So, so with that, with that
said, I will add to what's the right strategy for you to choose
as a founder? Well, like, you know, pick, I
always like to pick problems which are first of all, they're
obvious. Like, you know, like you have to
go and talk to five people. They will not argue with you

(37:07):
that, hey, is this a problem or not?
I think is it all like, if you talk to the first five people
and there and you you don't havethat clarity, but you know, from
them, then there's something wrong and you got to like work a
little bit more on it idea. So, so get to that, get to that
level where like, you know, whoever you talk to actually
agree with you that yes, you know, you're solving a real
like, you know, important problem.

(37:28):
And and also like, you know, like I like to work on problems
which have broader impact. So big problems, you know, that
a lot of people are going to have because you know, that's
going to actually create more opportunities for you.
Like even if there are a few other companies that also solve
the same problem where there's alarge market that you can tap
into and you'll have your own success, you know, along with
them. And then over time, you know,
like, you know, if you do the best that you're going to win

(37:50):
like over everybody else. And then, and then the, the, the
last thing I would add is don't like come up with an idea where
you feel like, well, like all I need to do is use AI and it's
gonna actually solve this particular task for me.
Like, think about, you know, if it is easy to build, if it is

(38:13):
super easy to build, then everybody else can also build
it. And then you're not really
adding value. So, you know, like AI should be
no more than one of the tools inyour tool kit to solve that
problem. But like, you know, make sure
that you know there's something substantial that you're
building. There are so many companies that
are building wrappers around themodels.

(38:37):
What about consolidation among these types of companies as the
models advanced their capabilities?
If you are a startup and you area thin dropper over the core
capabilities of an LLN, well, you'll be irrelevant quite soon.
Yeah, you have to like in the mindset, like I'll tell you what

(38:59):
we do again, you know, you know,we, we know that like, you know,
so the way our product works is that, you know, we actually work
with all the LMS, you know, thatare out there in the market.
You know, whether it's, you know, LMS from Open AI or in
Tropic or Google or Meta or, youknow, like and all like, you
know, so many LMS from open source, we work with all of
them. And our model for like what we

(39:21):
do is, well, we're going to use all the capabilities that these
LLMS, you know, provide to us and they will actually bring
those capabilities to our customers.
But we'll also be, you know, ensuring that we're building a
very deep technology stack on top of that platform.
And as the LLMS advance, as theyactually, you know, add some of

(39:43):
those capabilities that they've built ourselves that actually
throw our, we have to throw whatwe built like, you know, that
the LLM providers can do alreadyfor us and actually keep going
up the value chain. Like, you know, and, and so
that's, that's the model that wechoose.
Like, you know, like to get out your space and maximally use
like the innovation that's happening in the industry, but

(40:04):
then you don't build build a significant layer on top of of
that so that you can make that technology accessible to your to
your customers. Given the importance of the
foundation models to your business, how do you manage the
fact that AI is evolving at sucha rapid pace?
And how do you balance the stability of your business and

(40:29):
product direction while these underlying capabilities are just
shifting all the time? We have no choice.
You have to like take advantage of the rapid innovation that is
happening in the industry, otherwise you're going to be
left behind and you have to fundamentally change like, you
know, your execution model. Like, you know, we have to like,

(40:49):
you know, like we, when we started, you know, when we
started glean the we sort of very understood that very, you
know, at a fundamental level, like, you know, what
technologies, you know, we had available from open source and
cloud. And, and then you get to sort of
build your road map and you build a, you build a one year
road map. And and that's, that's not no

(41:11):
longer how things work today. Right now.
Technology changes on a monthly basis.
And, and so you have to fundamentally change your
architecture. Like, you know, like one of the
things we change, as an example,is that we don't have the annual
like, you know, or a quarterly planning process.
We switch to a monthly planning process in terms of like, like
how we're going to actually, youknow, build our technology
because every month, like, you know, there's new things to look

(41:33):
at and you have to quickly adapt.
The other thing that we, you know, also added is this concept
of the, well, figure out what you're going to throw like every
month, you know, in technology that you've built.
Like this is a new fundamental way of, you know, building tech
startups is that, you know, you will become obsolete if you

(41:53):
actually hold on to technology that you build for multiple
years because all of that technology that you built two
years back is most likely obsolete at this moment.
And so you have to constantly like, think, you know, of this
execution model where not only are you thinking about like new
things to build, you're also very actively thinking about
like, you know, things that you need to actually throw away and

(42:15):
actually leverage like, you know, the innovation from the
industry to replace that. We have another question from
Twitter from Arsalan Khan again who says since most AI and data
is created in English or Chinese, do you think start-ups
should also focus on non-englishor Chinese?

(42:36):
And does this create a digital divide?
1st, the content is created in many, many different languages.
Yes, you know, I, you know, English is dominant in some
ways, but there is, there is plenty of like, you know,
content, plenty of systems, you know, in different languages.
And, and I think what AI makes it possible today is it actually

(42:58):
helps you build products that work like, you know, that, you
know, that are that, that are global in nature.
Like it's much easier today to build a product, you know, that
you can actually bring to customers in, you know, all, all
parts of the world. You can localize your products
much easier with AI. You can make it work in Japanese

(43:18):
and in Korean and like Hindi andall the different languages at
at a much faster pace. Like, you know, so you know,
this is, this is one of the, youknow, a, you know, a core
capability of AI. But then in terms of biases and
the digital divide, it is true that the, and it's been true

(43:40):
like, you know, forever, like, you know, even on the Internet,
even Pai, when you go on Google and search for information,
there's always that bias that creeps in because like, you
know, English dominates as, as the source of knowledge in the
world, right? And, and so the, that, that's
it, that's, that's a good one. Like there's a problem I don't
have a good answer on, like how,how to sort of resolve that as

(44:04):
as as you know, AI becomes more and more capable.
How do you make sure that it's taking everybody's point of view
and taking, you know, all the knowledge that's out there?
I think I think like one thing maybe I should add is that we
have more capability today with,you know, with AI to process
like even non digitized, you know, content in a much, much

(44:24):
easier way. So hopefully, like, you know,
there's, there's some silver lining on, on on that side that,
you know, you can actually make use of, you know, data
knowledge, you know, content andlike different languages more
than ever before with AI. Let's talk about start-ups now
from the perspective of enterprise buyers, which is a a

(44:48):
very important part of course ofany start-ups life cycle.
If you're, if you're an enterprise start up, do you have
advice or a a framework that enterprise buyers can use for
evaluating AI startups? Especially given the fact that,
as you pointed out, every startup these days is using AI

(45:11):
and the hype is so intense, veryoften it's, it's hard to sort
through the claims to find out what's real and what's not.
And I'll just mention one one thing here that I remember
traditional software companies, ERP vendors and other enterprise
software companies. And I have to say this

(45:32):
situation, you know, 20 years ago was no different from that
standpoint. Now.
I mean, software companies make outlandish claims.
And So what should IT buyers andline of business buyers do about
it? This is one of the toughest
problems like, you know, for, for an enterprise buyer today.
I think the AI industry has donea bit of disservice making bold

(45:55):
claims, but then not not being able to follow through.
You know, like it's very easy tocreate really, really amazing
demos and visuals, you know, forwhat, what AI can do for you.
And enterprise buyers have like,you know, and they've realized
that like, you know, as they trythese products out that like,
you know, the claims often, you know, like, you know, are are

(46:16):
much larger than like, you know,the reality.
So I think like I would say like, you know, maybe maybe let
me take a step backwards and then first talk about like,
well, what should we even do? Like I think like there should
be a plan for like how you're going to roll out AI inside your
enterprise. And I feel like, you know,

(46:38):
centralizing that, you know, youknow, and having a core, you
know, AI strategy for enterprise, first of all, is the
right way to start. Like now think about all the
things you'd like to do this year with AI.
What are the different areas that you would like to actually
see AI make an impact? You can pick a few departments.
You can say that, well, you know, for our engineering teams,

(47:00):
for our customer service team, you know, like these are the top
two or three priorities, you know, where we, you know, want
AI to actually make an impact. So first build that road map and
build that together in your enterprise, you know, the CIO or
if you're just, you know, if you've chosen to have a, a head
of AI, chief AI officer type of,you know, role in your
enterprise, like, you know, let them, you know, give them that

(47:22):
charter of like making sure they're working with all the
different, you know, functional teams.
And you come up with a, a desired road map for AI for you
for this year. So you start, so you start
there. Now in terms of like the number
of vendors is not just like, youknow, Michael, as you said, like
it's not just the startups, it'sactually every existing software

(47:44):
company is also an AI company. They all have, you know, AI
things to sell to you. And I think you to make some
decisions there, like what's theright strategy for you?
You know, we feel that like, like, no, like, you know, you
have to be, you have to control like how many AI products you're
going to actually bring in. It's very, very hard to
evaluate. It's not easy, by the way, like,

(48:05):
you know, it's just like, you know, the, the time that you
have to spend time to evaluate every single AI product is
enormous. Like, you know, you like, you
know, like setting those systemsup, getting them up and running
in your environment and testing them.
And like, you know, you sort of like companies have gone through
this exercise where they spend 6months, you know, and just to
find out that, you know, that this thing didn't work.

(48:27):
So like, what are the right strategies?
So, so like 2, So, so, so my suggestions, like, you know, I
have two pieces of advice. Like #1 like, basically work
with fewer, you know, number. Like don't have too many POC's.
Like, you know, start with a fewso that you can do justice to
them. And, and 2nd, instead of like

(48:48):
relying on demos and, you know, getting excited by a
presentation. Well, you can have a safer
strategy, which is like, we'll go and what every vendor that
you work with, like see if they have proof points like go and,
you know, make them, you know, connect, you know, connect you
with, you know, customers that they were able to create success
for. So I think that is like, you

(49:09):
know, it's a lot more helpful like, instead of evaluating
like, you know, like actually had those conversations with
your peers in the industry and see like, you know, where they
are achieving success. Like if you tried 4 things,
you'll see one thing that's successful, share that story
with others and, and, and they will share with you.
And that's probably the way to sort of scale and, and maybe
maybe this is a plug for Glean, but the, the way we think about

(49:33):
AI is that, well, like fundamentally all AI in your
enterprise is about making, you know, working with some data
that's in your enterprise using the reasoning and intelligence
powers of the language models. And then after that doing some
work, which you're again going to save, you know, in your
enterprise systems, you know, that work is going to get

(49:55):
recorded and saved in your enterprise systems.
So, so fundamentally, when you think about all AI use cases,
they're about working with your enterprise information and, and,
and then you're making and applying AI on it to do some,
you know, to, to actually make some magic happen.
And so we chose a different strategy with Glean, which is
that, well, why don't we actually build a system that's

(50:17):
connected to all of the enterprise data, right?
That's what Glean is. And, and now we're giving you
this platform where you can go and build like many, many of
these agents, many of these applications yourself.
And, and, and that way, like, you know, with this horizontal
strategy, you can, you know, youcan do a much better job at
security governance, you know, also like not having to buy

(50:38):
many, many different products. Like it's more cost effective
and, and and and and allows you to sort of like, you know, get
more value through like, you know, through 1, like, you know,
11 product. All right, we're almost out of
time. So Arvind, I'm gonna ask you a
bunch of questions, some from me, some from the audience
listening and I'll ask you to respond very quickly from an

(50:59):
enterprise perspective, the build versus buy decision, how
should enterprises make that choice?
Very quickly please. It's a build plus buy.
You have to make sure that you get as much turnkey technology
as you can, but you have to alsoremember that it's not going to
be enough like, you know, to addtrue value.
You will have to, you know, you know, build on top of those

(51:21):
systems. From Twitter, a big hurdle for
AI and LLM platforms is that they are not integrated with
work flows. How do you see the market
evolving? That's exactly right.
And I think the you you have to,you know, build that layer.
One of the key techniques for that has been rag like so you

(51:43):
take all of enterprise data and knowledge systems for flows and
you actually build the, you knowthis, you know, a retrieval
system and an action system on top of all of your enterprise
data and systems. And then you connect with AI
like you know, that's exactly what we, you know, the problem
that we solve with you. Another one from Twitter who the
CTOCFOCOO chief digital officer?Who creates AI strategy?

(52:08):
And should the chief AI officer report directly to the CICEO?
That's a good idea. Like so if you, if you actually
have a chief AI officer, having them report to the the CIO or
the CEO could be a good strategybecause this is a company wide
effort. So like the, if you don't
associate it with a function that's actually better in my

(52:29):
opinion. And but otherwise, like, you
know, I think like, like, you know, it's not don't be rigid.
Like, you know, like whoever is motivated, who's excited and
like, you know, like and has thecapability, let them drive their
efforts initially And like if you can always figure out how to
reorganize later, like, you know, in a more scalable way.
Here we have another question from Twitter, and this is about

(52:51):
pain. If AI startups are getting more
lean, are older AI startups at adisadvantage?
And as these older startups become more lean because more
efficient using AI, what is the implication for employment and
their workforce? So on, on the 1st question,
there is indeed a like a first mover disadvantage in AI because

(53:15):
when technology changes so fast and you, you have some of that
legacy code base in your systems, like it makes it like,
makes you a little bit less agile.
So well, that's, that's part of like, you know how it always is.
Like, you know, as a new startup, you always have the
agility advantage. And as a start, that's a little
bit older. Well, like, you know, work hard,
like work hard on modernizing, you know, your systems, your

(53:35):
stacks so that you can, so you don't fall behind for those
reasons. And then in terms of employment,
I think the like, I'll, I'll just say one thing.
I've not seen people losing employment because of AI.
Yeah, like, you know, we work with so many large enterprises.
They're bringing so much automation and efficiency for
them. But like, you know, every, like,

(53:57):
you know, every enterprise that we work with, you know, they
actually are concerned with bothbottom line and top line.
And so when they get some bins, nobody's, you know, giving up
their team members. Like, you know, they're not
actually, you know, they're justactually thinking about like,
well, I can do more. And, and so companies are
actually building products at a faster pace.
They're getting more things done.

(54:17):
So, so I'm not, I'm not super worried like, you know, I think
I guess like, like I will come back to the point that well,
like just stay relevant. You know, that's that's the
thing that matters for you as anindividual.
If you if you learn how to use AI you know you will have no
problems you know in the future.You're advocating maintaining
intellectual curiosity. I'm not trying to put words in
your mouth because as AI drives efficiency gains, if you're

(54:42):
intellectually current on what'sgoing on, you can adapt and work
around that the changes that aretaking place.
Yeah, and companies are really hungry.
They're really hungry for AI experts, for AI talents or like,
so become one. I think it's going to be good
for you. What advice do you have for
entrepreneurs considering starting an AI startup?

(55:04):
Number one conviction like you know, we know.
Stay, stick with your idea, don't give up and, and #2 well,
like, you know, if you're, if you want to succeed, like, you
know, be ready to, to work very,very hard.
And, and, and then, then the last thing I would say is that,
you know, a company is all aboutpeople that you, you know, build
it with. So like, you know, focus on like

(55:26):
having, you know, having a greatfollowing team and, and, and
spend a lot of time, you know, trying to get the right people
that employees, you know, in your organization.
And like you know, when you get the right people, you know they
will do the you know they will build great products.
They will make you succeed. Final thoughts or advice for
CIOs who are making AI investment decisions very

(55:46):
quickly, please. Focus on security, focus on the
like, you know, centralizing, you know, like your AI software
stack and, and, and, and I thinkthe, the, the last thing is, you
know, that I would say is that like we don't try to sort of get

(56:08):
small wins. You know, this is my advice.
Like, you know, don't like, you know, don't like, you know, like
create projects where you know, if it doesn't deliver 50%
success, it's a failure. Like, you know, like, you know,
force every, you know, team, every function in your team,
like in your organization to, topick, you know, one or two wins
for AI, like, you know, every quarter and see how that goes.
All right. And with that, we are out of

(56:30):
time. A huge thank you to Arvind Jane
from Glean. Arvind, thank you so much for
taking your time to be with us today.
Thank you so much, it was a lot of fun.
And thanks to everybody who watched and especially you folks
who asked amazing questions. You guys are truly awesome.
Before you go, subscribe to our newsletter.

(56:51):
Join our community. Go to cxotalk.com.
Subscribe to the newsletter, check it out and we'll see you
again next time everybody. We have awesome shows coming.
U talking about AI? See you later.

All Episodes

Building an AI Startup: What's Different in 2025? | #871

Episode Transcript

Popular Podcasts

Bookmarked by Reese's Book Club

On Purpose with Jay Shetty

Dateline NBC

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

.css-14f5ked{margin:0;word-break:break-word;display:-webkit-box;-webkit-box-orient:vertical;box-orient:vertical;-webkit-line-clamp:2;overflow:hidden;}Building an AI Startup: What's Different in 2025? | #871