Full transcript
0:01Um, welcome everyone.
0:03You have seen me yesterday in case you
0:05attended some of the panels. Here I am
0:09back here.
0:10And what I'm going to talk about is
0:15um
0:16essentially what we have observed, the
0:18patterns we have observed while talking
0:20to some of the our customers on the
0:22training side, on the upscaling side,
0:25and also on the product side. So we have
0:28been talking to a lot of customers,
0:31training a lot of customers,
0:33building solutions for a lot of
0:35customers, and deploying a product that
0:37we have built
0:39on
0:40within a lot of customers. Um,
0:43um, on cloud.
0:45Um,
0:46I don't think I formally introduced
0:49myself. So here I am. Feel free to
0:51connect with me on LinkedIn.
0:53You can look me up.
0:55Um,
0:56and the QR code is there.
0:59I've been practicing AI for more than
1:01half of my life. I long time ago started
1:04in grad school.
1:07Ended up doing machine learning. Um,
1:10worked for Microsoft a bit. Most of my
1:12time was at Bing and Bing Ads. Uh,
1:15and currently I am the founder of
1:18Agentyo AI. I'm also an instructor for
1:20Data Science Dojo, and I am also uh
1:25an adjunct faculty at the University of
1:27Pittsburgh. I teach generative AI.
1:30A few a grad level course and an
1:33undergraduate course.
1:35Um,
1:36let's
1:38dive right in.
1:40So when companies they they [snorts]
1:42start investing in AI when they start
1:45building AI products,
1:47um, there is this core paradox. And what
1:50is that core paradox?
1:52Pretty much every company is investing
1:54in AI right now. I mean, any company
1:56that matters.
1:58You know, everyone is rushing to so that
2:01they are not left behind.
2:03Um,
2:04companies they launch pilots.
2:07Pilots are launched. Um, and at an
2:10individual level
2:12for
2:14for a customer, they
2:17are maybe not customer but actually at
2:20an employee level or maybe at a small
2:22team level, the pilot shows success.
2:26There is some
2:27individual gain that is visible.
2:31Um,
2:32and they try to
2:34and they try to
2:37deploy
2:38these pilots,
2:40but what ends up happening is that when
2:43these pilots are deployed widely, and I
2:46can I can give you examples firsthand
2:48that we have seen a sales co-pilot
2:51deployed that the sales team does not
2:54want to use.
2:55You have,
2:57you know,
2:58uh some marketing CRM type tool
3:03deployed as companion to CRMs. Our CRMs
3:06are already there, but a companion to
3:07CRM deployed, but no one wants to use
3:11it, right? So and there are a lot of
3:13different reasons that we will talk
3:14about, but fundamentally what happens is
3:19companies they do adopt they at least
3:22try to adopt on surface. They try to
3:23adopt AI, but they don't change their
3:26operating model. Um, and they don't
3:30adapt their
3:33standard workflows.
3:34And um
3:36and
3:37the gains that apparently are visible at
3:41an individual or very small team level,
3:44these gains are not visible
3:47um, on the balance sheet.
3:49And as a
3:51AI, the case that we'll make is that
3:54AI especially agentic AI,
3:57it requires your enterprise to be
4:00fundamentally rewired. You have to do
4:03certain things
4:05um, differently. You have to start to do
4:08certain things differently when you
4:10when you
4:11are in that uh
4:13in that AI transformation journey.
4:16And very often
4:19in this case,
4:20the bottleneck is usually not model
4:22quality or data availability. Models are
4:25amazing.
4:28Having
4:29building having built these systems in
4:32the last couple of years,
4:34I can tell you the models have come a
4:36long, long way. It it is incredible how
4:39models are
4:40the innovation has accelerated on the
4:43model side and even the architecture
4:45side. Token limits, inference,
4:49you know, the latencies. Uh,
4:51uh in terms of, you know,
4:55context windows. Context windows
4:57Reasoning has incredibly improved. A lot
4:59of things that we had to prompt for, now
5:02you don't have to. So it's part of the
5:04reasoning architecture.
5:06And
5:07usually the challenges that the
5:09organizations are now facing are not
5:11really
5:13the technical capability only. It is the
5:15technical capability
5:18not able to meet the organizational
5:21design. It is it is the last mile
5:23problem as as we call it.
5:25So here is this this idea
5:29that there are these islands of
5:30productivity. I have this tool. I can
5:32write proposals very well on my laptop
5:36or perhaps
5:37one or two people, but
5:41there is when I try to scale these
5:43pilots
5:44uh
5:46widely,
5:48things don't work. Things don't work as
5:50well. Um,
5:52perhaps
5:54you have this this example here. It is
5:57it is the bottleneck that is shifting.
5:59Previously you were spending more time
6:01in drafting this complex contract, but
6:04now there is a manual legal review that
6:07has to happen. And that manual legal
6:09review is well,
6:13it is still the bottleneck, right? So
6:14you have
6:16you have um
6:18um
6:19uh you have automated part of the your
6:23this contract pipeline, but you've not
6:26fundamentally fundamentally designed the
6:29entire workflow
6:31to move at an agentic pace.
6:35Um,
6:36also there is
6:39what we have observed is
6:41that companies usually they start at the
6:43top. Companies would start scaling
6:46agents
6:47um, right away. Just, you know, just
6:49let's go and build this agent
6:51um, only to realize
6:54that their infrastructure, their tools,
6:57data and tools are not ready. You
6:59[clears throat] may not have the the
7:01right connectors. You don't have the
7:03right access. You don't have the right
7:05permissions and you know, the processes
7:07are not there. Then you have the
7:09security, privacy, and compliance, and
7:11governance problems
7:13that are
7:15that are big challenge. Then you have
7:18this uh
7:19um
7:21uh this interoperability
7:23problem, right? So do do the data
7:25sources and agents can they talk to each
7:27other? And just talking is talking is
7:30not enough. Talking in a manner that it
7:33is safe, right? So
7:34do they talk in a compliant manner? Do
7:36they talk in a
7:38are they maintaining the governance
7:40aspect of it and so on.
7:42Um, then there is the general literacy
7:45in your organization, which is I will
7:47mention in later slide. It is an
7:48underappreciated area. More so, it is
7:53it is incredibly important, more
7:55important than than ever um, that your
7:58entire organization is AI upskilled. And
8:02[snorts]
8:03you know, you will if you are a decision
8:06maker, you will have tons of people who
8:08will come to you, service
8:10service providers and you know, solution
8:12integrators and services companies. They
8:14will come and say, "Hey, we will just
8:16come and deploy things for you." Right?
8:17So we'll just come in and it's going to
8:20be roses.
8:22Um,
8:23and it is going to be happy or you will
8:26you will live happily happily ever
8:27after. But the challenge is
8:30it doesn't work that way. You it doesn't
8:32work that way because the details are
8:35actually quite gory. Look at what they
8:37are.
8:38So first we will look at some key
8:40friction points.
8:42By the way, this is not primarily a
8:44technical talk. This is
8:46maybe for a change. I think most of the
8:48our talks have been technical, but this
8:50is going to be
8:52uh less of a technical talk and more on
8:55the other non-technical aspects of
8:59that are needed for success.
9:01So we'll discuss some canonical friction
9:04organizational friction points. Um,
9:07and these are not technology gaps. I'm
9:09not going to
9:11fret about technology in this talk. I'm
9:13going to talk about really the
9:15organizational design aspect of
9:17AI adoption.
9:19Um, as I said, structural barriers with
9:22structural barriers within an
9:24organization. And what my hope is by the
9:27end of this talk, all of you will be in
9:30a position to
9:33um
9:34to at least assess where the company
9:37stands
9:38and what the what your AI transformation
9:42roadmap must address.
9:45We'll talk about these seven points. The
9:47first one, what we call the pilot
9:50proliferation. There is this thing
9:52called productivity gap. There is
9:55process there's this thing called
9:57process debt. Um then governance is
9:59there, um architectural complexity is
10:02there, the efficiency trap, and the
10:04tribal knowledge is there.
10:07So, the first one.
10:09So,
10:10um this I like this the way
10:13um you know, some of these ideas are
10:15coming from an article um
10:18uh so, I like the the way uh you know,
10:22the the authors actually talk about this
10:24that this is uh
10:26you know, they call it pilot ledge and
10:28transformation pool.
10:30Right? So, and if you look at it,
10:32um
10:33um
10:34this is the reality of many many
10:36companies at the moment. Uh
10:38but the problem happens is the problem
10:41is that many cases there is no
10:44repeatable path, and each each pilot
10:47remains in an island. And the the
10:49question that you should be asking in
10:51your organization is
10:53the this proposal writer that I have
10:55built, or this
10:57this um
10:59maybe CRM automation agent that I have
11:01written, or this automatic automated
11:05email responder that we have created, or
11:07this legal automated legal document
11:09review,
11:10um or this maybe AI paralegal that we
11:13have created, what is the repeatable
11:16path
11:17from this working pilot
11:19uh to
11:21to the point that this becomes a
11:23standard standard company workflow.
11:27Um
11:28the next thing that you need to ask is
11:31um
11:32how would these individual gains, the
11:35respective gains that you
11:37uh that you achieve in respective areas,
11:39how do they show up on your balance
11:41sheet?
11:43Um
11:45um and worse, right? So, uh there is
11:47there's a worse uh something that is
11:49worse. Uh I I don't know if how many of
11:51you have heard of this this term called
11:53AI slop. Uh I mean, they used to call it
11:55work slop. Now they say AI slop.
11:57Basically, instead of AI helping, uh now
12:01there is this AI slop being created, and
12:04now you're spending more time fixing the
12:06problems left by either a sloppy AI
12:09agent or a sloppy use of an AI agent by
12:12uh someone who is uh
12:15you know, has not done their job, right?
12:17So,
12:18um
12:19So, how do you handle with this, right?
12:20So, but this is the best case scenario.
12:22Let's say you have a productive agent
12:24that actually saves you time, how is the
12:27time going to be reabsorbed? Cuz what we
12:30uh have seen is um
12:33you know, basically, uh worst case
12:35scenario is that AI, instead of helping,
12:38uh there is an illusion of AI helping in
12:41productivity, whereas AI is actually
12:43creating more work than uh actually
12:46helping.
12:47Um so, the question uh even if it is a
12:50successful co-pilot is
12:53where specifically would the reclaimed
12:56time go at your company?
12:59And who owns that redesign?
13:02Okay?
13:03Um
13:06Sorry.
13:07Here and process debt.
13:11Um
13:14Broken processes, they cannot be
13:16redesigned. They need a rebuild. And uh
13:19for those of you who have been building
13:21or have been working in data and
13:22analytics for a while now, uh we saw the
13:25same thing when we were going through
13:27this
13:28um the previous wave of predictive
13:30modeling and machine learning and data
13:32science or analytics transformation,
13:34that
13:35uh well, garbage in, garbage out, as we
13:37call it, right? So, if your data is
13:39broken, um
13:42how do you uh can you build models on
13:46it? Um in this case, uh
13:49if your processes are broken, if your
13:50data quality is bad, um data is still
13:53the king.
13:54How do you redesign your processes uh
13:57from scratch so AI can run on the on
14:00them reliably?
14:02Uh does everyone know how to how to use
14:05this AI agent? Because um this
14:07end-to-end automation, it is easier said
14:10than done. Uh in the AMD demo that we
14:12were looking at, this whole open
14:14clothing, it comes with a disclaimer.
14:17Comes with a disclaimer. Um
14:20someone was telling me uh just
14:22yesterday, I think one of the speakers,
14:24or um I'm forgetting who told me, but uh
14:28open claw actually uh
14:30modified its own guardrails.
14:34Uh it found it to be restricting because
14:36when you give autonomy to these tools,
14:38well, they are autonomous.
14:41Um
14:42The blessing in these tools is the
14:43autonomy, and the curse is also that
14:46that that can also become the curse. And
14:48how do you
14:49how do you um make sure that your
14:53process is redesigned such that um
14:56such that it is uh it is AI ready?
15:00And which workflows you need to rebuild
15:03uh before we can uh layer AI on top?
15:07Um
15:09And uh and for that matter, which uh
15:12which workflows can be used as is?
15:15Uh this is a good one.
15:18Now, when we look at when we look at AI,
15:21um
15:23that's
15:24GPT-5.1 GPT-5.1,
15:27uh
15:27uh Lama 4, um
15:30Claude Sonnet, Claude Opus, right? So,
15:32you name any model
15:34and tell me.
15:36Um
15:37These models are incredibly
15:39knowledgeable. They have seen a lot of
15:42data
15:43um
15:44across the board, probably all the books
15:47that are published, uh the entire
15:49internet, all the tweets, all of Reddit
15:52data. You just use your imagination.
15:55But one one piece of knowledge or one
15:59aspect of knowledge
16:01uh that these models do not have is your
16:04organizational
16:06um
16:07um organizational knowledge, right? So,
16:09your whatever is inside your
16:10organization organization, because they
16:12don't have access to it.
16:14Um well, when you build a retrieval
16:17augmented generation [snorts]
16:18application, or you when you build these
16:20data sources and can or you connect
16:23these data sources with your models,
16:27well, you bridge that gap as well.
16:29Now, you these models can now see your
16:33internal enterprise, or your
16:36confidential, or your proprietary
16:39knowledge as well.
16:40But there is still a piece of knowledge
16:42that these models do not have, which is
16:45what is sitting in my head.
16:48I never documented it, right? I'm a
16:50factory worker. Um
16:52well, well, this when this machine
16:54vibrates, I have to kick it, right? And
16:56it starts working.
16:59I've not documented it, right? Or uh
18:04well, you will have to get that
18:05documented. Uh how do you get it
18:07documented? Because the moment you start
18:09talking about it,
18:11uh well, people,
18:13you know, they will feel unsafe.
18:15So, uh a key aspect of your
18:18organizational AI transformation design
18:21has to be that you have to reframe
18:24knowledge capture more uh as well,
18:27legacy building,
18:29uh improving things as opposed to
18:32being a threat to uh you know, not a
18:35cost-cutting exercise. Because the
18:37moment we frame it as a cost-cutting
18:39cutting exercise, well, it is uh you you
18:43will start seeing that defensive
18:45behavior and uh lack of cooperation. I'm
18:48talking about
18:49witnessing this um
18:52in organizations where we help uh we
18:54help them build perfectly fine, very
18:58good AI assistants, but uh adoption was
19:02hard
19:03for one reason or the other, and this is
19:05one of the reasons that was observed.
19:08Then,
19:09uh another aspect, and there is a QR
19:12code. Feel free to actually scan the QR
19:14code for there is a link to this article
19:17um
19:18uh that actually addresses this. There
19:20is this
19:21uh
19:22aspect of
19:25uh governance.
19:27Going back to the open claw example,
19:30autonomy.
19:32If you go forgive autonomy to
19:35uh to systems that are
19:37non-deterministic, they're very
19:38powerful. They can reason on their own.
19:42And um I'm sure um the fact that you
19:45guys are here, you are interested, you
19:46must have seen some of those news that
19:50uh you know, some Replit agent or,
19:52or lovable agent going and
19:55going and deleting a production
19:57database, right? So, we we have seen
19:59some of those.
20:01By design, these tools are meant to be
20:03autonomous, which makes them very
20:05powerful.
20:06What kind of governance framework do you
20:09have on top of it?
20:11What roles and what permissions
20:14these agents or these systems should
20:16have
20:17for them to be able to take actions? Um,
20:21and
20:22when you when you build these systems,
20:24when you start looking at these systems,
20:27these systems are fairly complex.
20:30Uh, when I go and use an assistant, it
20:33is not necessarily a single
20:36large language model call. It's not a
20:38single LLM call. There is a lot that
20:41happens on maybe 10, 20, 30 LLM calls
20:44that are happening. Think about this
20:46that a task that is given to apparently
20:48a single agent, there is several agents
20:50that are running.
20:53How much of can human in the loop even
20:57be
20:58an option there?
21:00Can human in the loop even exist in this
21:02case? So, the question is,
21:05how do you assign permissions to these?
21:07How do you assign roles to these? How do
21:09you
21:09control access? This is a going to be a
21:13challenge. So,
21:15invest in AI governance framework
21:18earlier on.
21:19And understand what are your company's
21:21non-negotiable
21:23human approval gates as agentic
21:25workflows.
21:27So,
21:28um,
21:29these things must be approved by a
21:31human, period.
21:34Because,
21:35for instance, right? So,
21:37you cannot
21:39uh, your triage nurse can never write a
21:43prescription. Your AI triage nurse can
21:46never write a prescription
21:48because there are regulatory concerns.
21:50And now,
21:51it is you, the business owner, who is
21:54going to be um,
21:56uh, who will be in a better position to
21:59decide which of these
22:02human approval gates are non-negotiable.
22:05Um, and in this case,
22:08organizations, they must treat agents
22:10like a managed workforce. Treat them
22:12like your employees. Give them
22:14permissions, onboard them,
22:17do a background check, monitor
22:18performance,
22:20you know,
22:21make sure they're compliant, give them
22:23permissions to for certain actions and
22:25restrict them from
22:27from certain actions. Don't let them
22:29actually don't give them permissions to
22:32elevate their own role and
22:34and or elevate their own access,
22:38things like that. So, treat them like
22:39your
22:40a managed human workforce and that will
22:43actually make a lot of things easier.
22:46Um,
22:47then there is uh,
22:49uh, if you look at this, there is uh,
22:52uh, an enterprise
22:55architecture. Usually, it is actually
22:58quite complex, right? So, how do you um,
23:02what connections and what
23:04interconnections must be in place
23:07before
23:08you can actually layer AI on top of uh,
23:13up of on top of your architecture? Um,
23:16there is going to be
23:18your enterprise is changing constantly.
23:22How do you
23:23how do you
23:25um, make sure that your uh,
23:29your platform evolution, it does not
23:32outpace. I mean, things don't change to
23:34the point that your initial assumptions
23:36are
23:37uh, void.
23:40Um,
23:42then I think this is one once again that
23:44the same idea, the tribal knowledge
23:48part. Um, the efficiency trap
23:51frame the
23:52frame the
23:54this whole idea of AI adoption more not
23:58as a head count reduction exercise, more
24:01frame this as a as a value creation
24:03creation and competitive advantage tool,
24:06not necessarily about head count
24:08elimination.
24:09So, how how can your
24:12your organization actually scale AI
24:14pilots successfully? Um,
24:17I will come back to this slide.
24:20I want to be mindful of the time we
24:22have. I will I promise I will come back
24:24if we have time.
24:25Uh, but let's very quickly go through a
24:28few things. The first and foremost thing
24:30is
24:31that you start with business first.
24:35Um, and some of these
24:37um,
24:38some of these things we just talked
24:40about, uh, take a more business
24:44centric approach. Do not actually assume
24:47that this is merely a technology
24:48problem. It's it's more than just a
24:50technology problem. Identify,
24:53you know, what are the the key KPIs,
24:55what's what is the needle that you need
24:57to move.
24:58Um, where are the opportunities and what
25:00are the bottle what are the bottlenecks
25:03that do to realizing the potential of AI
25:07transformation? So, we finished this
25:09part just now.
25:11But,
25:13there is another aspect of it.
25:16Everything that we talked about,
25:19uh, none of it is going to make sense
25:22unless
25:23your entire organization and is
25:26upskilled. And when I talk about entire
25:29organization, I mean entire
25:31organization, every single person.
25:34And when you look at it,
25:37the this is an underappreciated area
25:40because most of the focus
25:42um, even in analytics, even when we talk
25:46about even when we talk about uh,
25:49the machine learning and predictive
25:50modeling and I've I've been teaching
25:52I've been teaching for quite some time
25:54now.
25:55And majority of the people who attend
25:59attend these
26:00trainings,
26:02I would say significant majority is
26:04people who are
26:06who are builders, who are actually going
26:08to go and implement things.
26:11But, there is no emphasis on the the
26:13rest of the ecosystem. People who are
26:15the enablers of these products, the
26:18project managers, program managers, the
26:19product managers,
26:21uh, the the people who will consume
26:23these products, the the the person who
26:26will create
26:27you will manage your data, the person
26:29who's going to actually use this
26:32or deploy this in the field.
26:36So, we
26:38look at it
26:39if you look at this uh,
26:41uh, you know,
26:42usually companies focus on
26:45on the AI capable tier.
26:49What what I'm saying here is that we
26:52have to also focus on this AI aware
26:56tier, which is um,
26:59uh, you know, some very very basic
27:01literacy in case of AI
27:04and knowledge work, at least some very
27:07basic understanding of how to how to
27:10build how to write effective prompts.
27:13Um, how to get get uh,
27:16um,
27:17you know, understanding of well, if you
27:20phrase the question poorly, um,
27:23it is going to cost you more.
27:25Um, if you if you just
27:27bring in a lot of context, well,
27:30Anthropic is going to be happy because
27:32they will send you a fat bill. But, the
27:34reality is
27:36uh, you have to you have to have people
27:39um, in your entire organization, every
27:42single person who has who is a good
27:45power user of these tools.
27:49Being able to use understand some basic
27:50concepts, how does well,
27:53how does not at the
27:56the architecture level, but generally
27:58how are these tools actually working?
28:00So,
28:01you are able to they are able to
28:02everyone is able to actually get the
28:05most out of these tools.
28:07Um, so an awareness of possibilities and
28:09challenges. Everyone should be aware as
28:11well, if I am a front desk associate and
28:15I'm working at the reception, how can I
28:18use it to make my job more efficient? If
28:22I am a nurse, how can I do it? If I am a
28:24marketing manager, how do I do it? If
28:26I'm a developer, that one is the more
28:28obvious case, but you name it.
28:30Uh, any knowledge worker should be able
28:33to
28:34should have at least a very basic
28:36understanding.
28:37Then,
28:38um, there is this AI ready tier,
28:41where they
28:42where this portion of the workforce that
28:46can connect uh, that can connect um,
28:50the opportunities
28:53um,
28:54in their own industry, in their own
28:56domain
28:58to AI.
29:00Well, if I saw someone
29:03doing an automated proposal writer,
29:05well, what does it mean in my case? If I
29:08see a legal document reviewer in legal
29:11industry, how can I use it in retail?
29:16So,
29:17uh, being able to connect those
29:18opportunities and engage with people
29:23in the AI capable tier to plan some
29:26widespread AI initiatives.
29:28Um,
29:29so, if you look at this,
29:31um, uh, most of the people, they are in
29:35the bottom this the bottom tier.
29:38Um,
29:40uh, the bottom tier and I see some
29:42questions coming up.
29:44I will pause after this slide to
29:46actually take questions. Just
29:48and please type in your questions. I
29:50mean they are coming up on
29:56Um
29:58So, as I said, every single member
30:01in your organization must be AI
30:03literate.
30:04Um
30:05Uh the builders, I don't need to
30:07emphasize. It's builders actually they
30:09want to be very very AI literate. I
30:11don't need to make a case there. But the
30:14ecosystem, people who are consuming or
30:16enabling uh
30:18um uh consuming and guiding. There is a
30:21typo here, right? So,
30:23uh people who are consuming and building
30:26data and AI products, uh they have to be
30:28AI aware.
30:30I mean, so ideally they have to be AI
30:32ready. Um
30:34Um then leadership,
30:36they must be, right? So, there's no no
30:39other way, right? So, uh I know actually
30:42someone
30:44very high up. I mean, I think CXO of a
30:47billion plus dollar company. Um last
30:51last week on-site meeting.
30:53And
30:55the guy actually started building
30:58uh using cloud code, never coded in
31:01entire life ever.
31:03And the guy was so excited about
31:05being able to build
31:07a proof of concept for an application
31:11and giving a tough time to the to the AI
31:14team because well,
31:16uh started building. I mean, said I
31:18mean, is it even possible?
31:20Uh
31:21And yes, I mean, if you look at this,
31:23you're
31:24um now from ideation to actually a proof
31:29of concept, it is quite easy. And now a
31:32lot is possible and this was uh
31:35uh this was actually not possible
31:38uh until a few years ago.
31:41So,
31:42uh let me actually take some of the
31:45questions.
31:47I will go back.
31:49Jane, your question. What role in an
31:51organization fit into the AI ready
31:54stage?
31:56Um
31:58Um
31:59It depends upon the organization. What I
32:01can tell you is AI ready is at least
32:05your product managers, your project
32:06managers, your program managers, people
32:09who are either
32:11um
32:13involved in building products, not the
32:16hands-on part, but at least they are
32:19they are
32:20um
32:21they're the ones who are guiding the
32:22products, right? So, they must be AI
32:24ready.
32:26Um and I actually think that
32:29an AI product manager is going to be one
32:31of the most sought-after roles.
32:33And this is based on
32:35um me hiring a few people here and
32:38there, uh looking at customers
32:42talking about what they need.
32:44Um
32:46And also based on our own products that
32:48we're building, we are
32:51uh
32:52I think that I mean, anyone who can
32:54think as with more customer empathy uh
32:57can
32:58can be more product focused, you know,
33:00can envision, has has an understanding
33:02of
33:04has an understanding of uh the business
33:06and the domain. And if they are AI
33:08ready, that's actually an
33:11an amazing combination. So, but anyone
33:14right now who is actually helping the
33:17builders in a supporting role and think
33:20product program project manager type
33:22people, they have to be AI ready. I
33:24mean, there's no other way in my
33:26opinion. Of course, I mean, depend upon
33:28depending upon what industry what
33:29company you are, you may be able to
33:32survive, but if you want to be, you
33:34know, if you're very very
33:37obsessed or selfish about your career,
33:39you better be AI ready.
33:41Um let me take
33:44uh
33:45um
33:48Okay.
33:50Let me go from the top.
33:54A lot of questions.
33:58What is one practical change that most
34:00improves pilot to production conversion?
34:04Srinath, I mean, you got me. I don't
34:06know. I'm uh I'm not sure if there's one
34:09practical change
34:11that most improves pilot to production
34:14conversion.
34:15That's a tough one.
34:20Basically, let me rephrase the question
34:22and please feel free to type in if you
34:25think that this is
34:26I'm rephrasing it correctly. You can
34:28say, what is the biggest reason the
34:30pilot to production conversion fails?
34:33You know, maybe let me
34:35flip it because if we change it, then
34:37you know, it will successful.
34:38Governance, I would say.
34:41Because a lot of times um
34:44uh the issues are around accidental
34:47exposure of data, accidental
34:49um access, unauthorized access,
34:53accidentally
34:54uh
34:56uh not
34:58uh exposing any PII and all of that. I
35:01think that would be probably if I had to
35:04if you gave me, hey, Yug Raj, you can
35:06fix only one thing, I will fix
35:08governance. And this is assuming that
35:10the pilot of pilot from a technical
35:11standpoint is working well. If it is
35:14working well from a technical
35:15standpoint, I think governance is the
35:16first thing that I will fix. Thanks
35:18thanks for forcing me to think
35:20and come up with a
35:22uh concrete answer. Let me see.
35:25Um
35:28Okay, this is scrolling very quickly.
35:31Um
35:32Do you expect to be
35:35Um
35:39Okay, Jane, the question here is uh
35:43Do you think whether adopting a
35:44generative AI at an enterprise level
35:48will cause human resource redundancy?
35:52Jane, um
35:54I think so.
35:56Uh and uh and this is something
35:59um
36:01and maybe at a very personal level, I
36:03can tell you, right? So, I was just this
36:05morning I was talking to
36:08uh talking to uh a family member a
36:11member about my nephew first year
36:15computer science
36:17um
36:18you know, a freshman or I think
36:20going into sophomore like after summer.
36:24And one piece of advice is um
36:27this is going to happen, right? So, and
36:29I'm I'm not painting a doom and gloom
36:31scenario.
36:34Um
36:35But what is going to happen is people
36:38who are more of generalists, they're
36:40going to thrive and survive.
36:42So, if I if I say, hey, I write
36:45beautiful code and I only write
36:47beautiful code, I don't know anything
36:49else. Well, there is a lot of people who
36:51can do that do it. Uh anyone smart can
36:54write code now.
36:56But not everyone who's smart and can
36:59write code also has customer empathy.
37:02Not everyone can
37:04think about think in terms of you know,
37:06what what makes business sense.
37:08And not everyone can simplify things.
37:11So, um
37:13well, some awesome amazing content
37:16marketer, hey, I write beautiful blogs.
37:18Sorry.
37:20That boat has sailed. I mean, um
37:23uh Claude and Chat GPT and these models,
37:26they can write beautiful blogs, but
37:29not everyone can think
37:32strategically in terms of marketing,
37:33right? So, you cannot
37:35who can execute the end-to-end marketing
37:37strategy? Who can actually
37:39uh [snorts] not only use Claude to cut
37:41down their writing blog writing time by
37:4680%? They still are good writers. They
37:50can still, you know, massage it and you
37:52know, understand I mean, what it takes
37:54to write a better blog.
37:56And they can also
37:57uh have a better content marketing plan.
38:00And also they can step in if there are
38:03other opportunities in marketing. So, a
38:05lot of roles are going to
38:09you know, companies they will need
38:11people who are multi-dimensional.
38:14Um who who can do multiple things at the
38:17same time, right? These very specialized
38:19roles, specially in knowledge work, they
38:22are going to actually become quite
38:25difficult to actually retain.
38:28Okay, let me see. Maybe I will take one
38:30more question and then I'll come back.
38:32Um
38:35Uh let me see.
38:39Can a generative AI ever bridge the gap
38:41between external data and internal human
38:43knowledge? Uh I think so.
38:47And that is a question Mobeen, you are
38:48asking this question. Um I think so,
38:51right? So, um
38:53when
38:54uh back in the day when I started
38:56working in online services, right? So,
38:58from a PhD, right? So, I started
38:59working.
39:01I was fascinated by
39:04how how granular the logging uh is
39:08becoming, right? So,
39:10um
39:11you know, at that time maybe we we
39:13logged every
39:15all the activity in
39:18what that would happen on the on the
39:20service, the search engine.
39:22Then you also are also bringing in logs
39:25from your app and joining them. Then
39:27you're also bringing in the like I think
39:29Google.
39:31Um uh you are gathering data from
39:35your search,
39:37but you're also gathering data from
39:39Chrome. You're also also gathering data
39:41from Google Maps. You're also gathering
39:44data from your home automation, Google
39:47Home. Now, you're bringing in more and
39:49more data sources.
39:51It it is going to be only a matter of
39:53time. The point that I'm trying to make
39:55is when there is economic value in
39:58bringing in
40:00more data,
40:01these the internal human knowledge
40:05organizations, they will start finding
40:07that
40:08finding a way to
40:11to extract that knowledge. Maybe every
40:13single meeting is going to be
40:14transcribed. Right? We'll have
40:18every single interaction, which is
40:19already happening by the way in many
40:21places, but every single interaction
40:23that I'm having, it is going to be
40:24recorded.
40:27Maybe companies might come up with you
40:30know, interview processes. Hey, how do
40:32you do this? Every time you do it, just
40:34talk it out loud. I don't know what that
40:36would look like, but that is going to
40:38happen because there is tremendous this
40:40economic opportunity in
40:42scooping out that internal knowledge,
40:45right? And this is this has
40:47this is how we have evolved
40:49historically.
40:50>> [clears throat]
40:50>> Okay, I'll come back to some questions.
40:52Let me just finish this up.
40:56Um the third the third thing is and
41:00going back to I think there was a
41:02question around this.
41:05Develop a more governance-centric view
41:07of your AI tools.
41:12And what that means is
41:14um that uh
41:17uh
41:19Think about this.
41:21A lot of times um
41:24uh when you go and use SAS or when you
41:26use a third-party tool,
41:29um maybe your industry does not uh
41:33maybe your industry, your regulation,
41:35you are in a country, I mean, you have
41:38to be
41:40your cloud has to be sovereign. Um or
41:42you are in healthcare, there are certain
41:44restrictions.
41:45Um have
41:46proper off off and access controls
41:49for people using these agents and for
41:51agents accessing the the actions and the
41:54actions and the sources of context. Uh
41:57have proper understanding and data
41:59governance and explain
42:01explanation of where the where all the
42:04data set data is, who has access, who
42:07does not have access.
42:09Um observability. In the previous
42:13uh if I may, let me call it the
42:16revolution, the predictive modeling and
42:18machine learning wave that happened,
42:20um a lot of industries, they would say
42:24that they still want to use regression
42:26models because they are more
42:27explainable. And any model that was not
42:30explainable like things like random
42:32forest or um
42:34AdaBoost or XGBoost, they were not cool
42:37because well, we cannot explain
42:40uh
42:41why certain things are happening the
42:44certain way. So, having an observability
42:47layer on top of your infrastructure,
42:50that is going to be important. Then you
42:51have
42:52guardrails.
42:54Incredibly important.
42:55Guardrails are incredibly important
42:58because you may have implications um
43:01um in
43:03if your your chatbot or your assistant
43:06um
43:07um
43:10if your assistant misfires, there can be
43:12a problem. Um
43:14Air Canada recently got sued by a
43:18passenger because
43:20the the chatbot gave some offer some
43:23information that was wrong. Matter went
43:25to court. Court ruled in favor of
43:28of
43:29the customer, not Air Canada. Uh
43:31evaluation. You have
43:34your models rely on sources of context,
43:36your data sources, organizational data
43:38is constantly shifting. How do you make
43:41sure
43:42that your um
43:45your
43:46assistants, they keep performing
43:48correctly and correctly and correctly.
43:50Um so, there is there is a lot of
43:53things. I mean, if you have to take this
43:55organizational design
43:57or your
43:58let's say agentic AI organization, let's
44:00say you have a human org and then you
44:02have your organization of
44:05AI agents. Uh the way you do it, you are
44:08going to actually
44:09um
44:10uh
44:11have similar um
44:13concepts, similar ideas um
44:17that you would be actually implementing.
44:21Um
44:22Okay, let me see what we have further
44:24here. So, uh
44:27just
44:29How can we help as an organization? Me
44:31myself, Data Science Dojo,
44:35We actually are the first boot camp in
44:37the world on agentic AI and large
44:40language models and there are others
44:41that are popping up. We have been
44:43actually building, but the we have been
44:46actually teaching, but the difference is
44:48we are not purely a training company. We
44:51are
44:52trainers who are also builders, right?
44:54So, I'm I'm one of the trainers myself.
44:56A lot of things that you hear are
44:58actually I learned them the hard way.
45:00These are wounds and
45:03scars and scabs that are uh that I was
45:06mentioning. So, we I mean, if you're
45:09looking for a training program, we have
45:11training programs. Um
45:13I personally trained more than 8,000
45:14people in a face-to-face setting. You
45:17know, the company has trained a lot more
45:19than I have. Um any company that matters
45:22on the planet, I mean, someone from that
45:23company has been trained by us. And we
45:26can also do upscaling, organizational
45:29upscaling, enterprise trainings at all
45:31levels. We have trained leadership. We
45:33have trained uh well, the ecosystem and
45:37and then of course builder.
45:40Um we have trained a lot of builders
45:42actually from scratch. Um we all also we
45:46have an AI product which actually
45:49mirrors the philosophy
45:51that you
45:53um that I was mentioning.
45:56A product that is that treats your
45:58agents and agentic workflows as
46:01um
46:02um as a human team
46:05with all the compliance and governance
46:08and other controls.
46:10And
46:11we not only provide the product and
46:14services, we also provide this advisory
46:18how to implement things in a manner
46:21that your AI products are successful.
46:25So, the product that we have, as I said,
46:27right? So, very quickly, it's a it's a
46:30license product that it comes in and
46:32sits in your
46:34cloud.
46:35Um you can use any open source, any
46:38close source LLM.
46:41Um
46:42set up role-based access controls. So,
46:44your agents if they decide to go rogue,
46:47they don't have access uh
46:51um to the wrong things
46:53uh
46:54Uh then guardrails, observability,
46:56evaluation, AI red teaming is built in.
46:59Uh connect to any data. I mean, all the
47:02all the discussion that you were you
47:04were listening to, MCP,
47:07actions and tools and um
47:11you know, security, observability,
47:13everything that you heard in the last
47:15couple of days. So, the the the agenda
47:18was actually based on our experience as
47:22builders and our experience deploying
47:26these
47:27for our customers. Uh as opposed to some
47:30topics that seem to be cool. And this is
47:33a complete platform. You can actually
47:35once deployed, you can actually build
47:38uh build
47:39compre You can there are APIs and you
47:42can build any application on top of it.
47:44We are GDPR, HIPAA, and SOC 2 compliant.
47:48Um and what we do is we take um
47:52um
47:53very similar approach. You know, slogan
47:56is IT is the new HR, right? So, what
47:58that means is that sent there is a
48:00centralized management of AI co-workers
48:03for the entire life cycle from the call
48:06it hiring,
48:08use case identification, deployment,
48:10uh to agent learning and development,
48:13agent ops, observability, monitoring,
48:15feedback, cost monitoring, uh
48:18organizational alignment and compliance,
48:20all of that is actually controlled
48:23centrally.
48:24Um
48:25And I will actually
48:29So, if you So, this is the
48:31link to contact us. If you think that we
48:34can be of any help to you in
48:37for your
48:38um for your training, upskilling, or uh
48:43or product or advisory needs,
48:46you can use this form. Just just send us
48:48a note and we'll reach out to you.
48:50I will go back. I know I am right. It's
48:55right at top of hour. We are here. Let
48:57me just take I will just take two more
49:00questions and then we will
49:04um I'll hand it back to Rebecca.
49:08Yes,
49:10even if you are non-technical,
49:12we have trainings. I'm actually pretty
49:15confident, right? So, I'm pretty
49:17confident that as long as you are
49:18committed, I mean, code is not even a
49:21problem anymore. So, even if you're
49:23non-technical,
49:24you should be able to actually do it.
49:26And we have trainings actually for
49:28non-technical knowledge workers as well.
49:30So, depending upon what you're looking
49:32for, we are
49:34we are going to I mean, that definitely
49:37it should not be a problem. We have
49:38trainings for everyone. Even our
49:40technical trainings, I think the way we
49:41have structured them you should be able
49:44to actually keep up.
49:46Um
49:50Okay, what is the best way to manage
49:53accountability of quality of results and
49:55reliability of the pipeline as there are
49:57a lot of under the hood dependencies?
49:59Have
50:00take a more observability evaluation
50:04and governance first approach, right?
50:06So, everything that you're talking
50:07about, you have to have
50:10uh
50:11um all those controls in place before
50:14you even get started.
50:16Uh
50:18Okay, and I think I will hand it back to
50:24uh Rebecca.
50:26Um
50:28thank you so much Roger for that, you
50:29know, really useful
50:31um discussion, really important one,
50:32too. And I really like to the point
50:34about, you know, what about the wisdom
50:36sitting kind of exclusively inside a
50:38human's mind that's not specifically
50:40injected into a knowledge base um for an
50:43AI agent. And
50:44you know, organizational wisdom often
50:46goes undocumented, so it's an important
50:48piece of missing data. And
50:50and who knows, maybe in the future um
50:52it'll start to become more of a
50:53requirement for some documentation
50:55coming out of people's uh deeper thought
50:57process that goes into a task.
50:59How maybe a direct link, right? So, just
51:03a neural link, right? So, yeah.
51:04>> Nothing's nothing's far-fetched. Um so,
51:06we have about 7 minutes until our next
51:08session. So, we'll take a seven a short
51:107-minute break, and then we'll be back
51:11soon for the next session.
51:13Thank you, Rebecca.