Why Most AI Projects Fail in Enterprises? Rohan Murty Explains

0:00If you and I are competing, [music]

0:02um, and you have the same model that I

0:04do,

0:05well, since you've democratized

0:06intelligence, [music]

0:07we used to compete in who has better

0:10access to market with better people,

0:12with better product, better IP, and so

0:14on. Well, now if all that gets

0:16normalized, right? Then what remains is

0:18difference. How are you dividing your

0:19optimism versus, uh,

0:21concern versus pessimism? There are

0:23probably two fundamental questions. The

0:26first of which is the role [music] of

0:28human beings going to continue to be

0:31what it has been so far. That's the

0:32heart of everything. And the second

0:34[music] is there also intermediaries,

0:35and is the role of intermediaries going

0:37to continue in many industries? I think

0:39both of these are being put to the test

0:40in [music] a way that we never could

0:42have done before.

0:54Ron, thank you so much for joining me

0:56today on NASSCOM conversations. So, uh,

0:59you know, the big challenge is as

1:01enterprises try and adopt and adapt AI

1:04into their workflows, into their

1:06processes, how do they do it? And what

1:08is the way to do it in a way that it's

1:10genuinely outcome-based, so that they

1:12can realize the benefits of AI also

1:15while competing in most cases, in a

1:17pretty intense marketplace. So, two

1:20questions to start off. What are they

1:22doing right and what are they doing

1:24wrong? See, the the greatest strength of

1:26all these models uh, is also

1:29simultaneously the greatest weakness,

1:31and that is these models generalize very

1:33well because they also their

1:35architectures are such that they can

1:36generalize. They're trained on data on

1:38the internet, which is quite diverse and

1:41large. They do lots of different things.

1:43And that's wonderful. That's why you

1:45can, you know, span a whole gamut of of

1:47different things.

1:48But at the same time, they understand

1:50very little about the nuances and

1:52specifics of how an organization works.

1:56Of the local

1:58uh,

1:59uh, sort of knowledge that an

2:01organization has accumulated in serving

2:03its market, its customers, its

2:05hard-earned sort of wisdom and

2:07experience of its people, etc., and so

2:09on. And so, when these models don't

2:11really have that information or don't

2:13understand it, the problem is you'll end

2:15up with models telling you things that

2:17are very generic. And often then people

2:19saying, "Hey, this isn't actually really

2:21moving the needle for me." And perhaps

2:22the best evidence of this is by large

2:25today, the truth, despite all this hype

2:28of agents and agent AI and all of this

2:30stuff, the core fundamental truth is

2:33simply this, that you only have models

2:36operating or being successful in few

2:39areas. One example is code generation.

2:42Right? Perhaps maybe a little bit in

2:44legal, uh, but beyond that by and large,

2:47barring some exceptions, most functions,

2:50there are no real successful agents. And

2:53by successful I mean agents that work,

2:55that worked in organization

2:56satisfaction, that produced some

2:57positive outcome, that produced an ROI,

3:00etc., and so on. And the reason for

3:02that, or rather the biggest reason for

3:03that, is these models understand these

3:05agents are powered by models, and these

3:07models don't understand organizations

3:09that they're trying to serve. Mhm.

3:11Right? So, they're missing the context

3:14that is unique or local or specific to

3:16that organization.

3:18Um, and that's why you're, you know,

3:20we're not yet seeing agents rise up and

3:24sort of take over the enterprise in the

3:25way that we hope. Uh, you're seeing it

3:27more perhaps in the consumer space and

3:29so on, but not in the enterprise, not

3:30yet. Right. So, you know, uh, you've

3:32written a series of articles on this,

3:34including for the Harvard Business

3:36Review. So, one of the things you've

3:37talked about is contextual computing,

3:39and I'll come to that. But give us a

3:41contrast between, uh, one of the

3:43examples that you've used for in

3:44insurance or hospitals, how things used

3:47to be done earlier versus how things can

3:49be done now if you were to genuinely or

3:52efficiently use AI. So, a very specific

3:55example is Let me actually take an

3:57example of sales, perhaps maybe your

3:58audience or It's a It's a broader

4:00example, right?

4:02Um, if you think about

4:04the accumulated knowledge of a customer

4:07in an organization, it's actually first

4:09takes in the sales team, but it's not

4:11just in the sales team. It's also in,

4:13let's say, your customer supporting

4:15because your customer support team is

4:16interacting with customers. They're

4:18engaging with customers, they're

4:19learning from customers. It could be in

4:20your customer success team. It could be

4:22in your product team. Can be in your

4:24delivery team.

4:25Each one of these teams talks to

4:27different parts of a customer stack, and

4:29therefore they're accumulating unique

4:31context about the customer. But none of

4:34this is really captured or stored

4:36anywhere. Right? It's these You you you

4:39kind of learn bits about the customer,

4:40but it's highly siloed and fragmented,

4:42and it melts away.

4:43Um, and then the sales team has to go

4:45back to the customer and figure out how

4:47do they serve them better and what might

4:48be the issues, etc., and so on. But if

4:50you could capture this context within

4:52your organization of how people are all

4:54interacting with different parts of the

4:56customer, the intelligence that they're

4:58gathering about different parts of the

4:59customer, you can put it all together

5:01and get a real 360-degree view of the

5:03customer. And what your sales people can

5:04now do is they can say, "Hey, here here

5:06are some proactive ideas or pitches that

5:08I can go make to the customer." Or when

5:10they walk to the customer, they already

5:11know what are the things that are broken

5:13because they know a different part of

5:14your organization has learned this. This

5:16is perhaps an example of what I mean by

5:18context that's local to an organization.

5:21But in this case, it's siloed across

5:22different parts of the company. Okay,

5:24that makes sense. So, and that's a good

5:25example to start now building further

5:27on. So, what can an enterprise do? So,

5:30you talked about all these, uh, you

5:32know, maybe I don't know, terabytes of

5:34information in the form of many things.

5:36It could be conversations, mails,

5:38exchanges. Anything you do in

5:39applications, yep. Anything you do in

5:41applications. So, all of it there all of

5:42that is sitting there. Yep. So, what's

5:44next? So, now, okay. So, let's say

5:46building on this example, right? Um, you

5:49you break open these silos, you capture

5:52all of these different contexts that

5:54different parts of your organization

5:55Your organization has a large surface

5:57area interfacing with customers, and you

5:59start capturing context about the

6:01customer but through your people.

6:03Now, suddenly you seem you will have a

6:05more complete picture of what is

6:06happening with this customer. You get

6:08intelligence of what is the customer

6:09thinking, what are their priorities,

6:11what is working, what's not, what are

6:13their failure scenarios, etc., etc., and

6:14so on.

6:15And and so now you can do things like,

6:19"Can I go and proactively suggest

6:20something to this customer because I

6:22have intelligence to know where they

6:23might be going?"

6:25Or to put it a different way, you

6:27already have a whole bunch of leads

6:28about the customer that you don't know

6:31you have. Mhm.

6:32Right? So, you can actually change how

6:34you, you know, your top line. Right? Um,

6:37uh,

6:38you know, you can extend this further,

6:40yeah, and so on, and so forth. But this

6:42is just one example. I I can give you

6:44more examples of context and so on, but

6:46but you kind of get a sense. Yeah, and

6:48so let's take the example of an

6:49automobile dealership, uh, which is part

6:51of a larger, uh, you know, chain of

6:53dealerships. How would they, uh, you

6:55know, and they're always collecting

6:56information because customers are

6:58walking in through the door and leaving.

7:01And and there are nuances as to how you

7:03deal with a customer in different parts

7:04of the country, in different, maybe,

7:06weather conditions because you're

7:07selling different propositions for the

7:08same car. So, how would all of this come

7:11together? So, and ac- actually it's also

7:12a compute question. So, let's say the

7:14data is being collected, as you said,

7:16uh, through emails and through maybe,

7:18telecaller, conversations, and so on.

7:20How does that What is the computer

7:22required to, you know, convert all of

7:23this into and what does that? In IT

7:26services itself, your delivery team is

7:28talking to a customer regularly. Mhm.

7:30And they have a whole bunch of context

7:32but closer to the ground of what's

7:34happening with the customer's reality.

7:35Then your sales team is talking to

7:37perhaps the economic buyers who are

7:38higher up the stack. Perhaps there are

7:40other parts of your company, legal and

7:41etc., who are also talking to the

7:42customer.

7:43But all of these folks are learning

7:46different For example, this delivery

7:48team is learning, "Oh, you know, this

7:50rollout in this geography is not going

7:51well." Mhm. Or their priorities in this

7:53other They they hear rumors, they hear

7:56gossip, they hear all kinds of different

7:58things that may be happening at the

8:00customer. Right? And there is valuable

8:02information. They may attend

8:03presentations about the customer with

8:04the customer where they learn, "This is

8:06how the customer is thinking. This is

8:08the kind of vector on which they want to

8:11follow." Now, how do you capture all

8:13that intelligence and then share that to

8:15the sales team which is trying to figure

8:16out, "Hey, how do I actually serve this

8:18customer better? Where do I take them

8:19next? How do I grow revenues?" etc.,

8:21etc., and so on. Today, all this

8:23information is in pure and silos. Now,

8:26if you could magically give everyone

8:27access to CRM, and more than that, you

8:29could magically get everyone to in a

8:31disciplined fashion to go enter data in

8:34the CRM, and good luck with that. Then,

8:36yes, you'll truly capture all the

8:37intelligence. But instead, if you could

8:39actually just capture the context that

8:42arises when people interface with their

8:43customers, you now have you'll take

8:46intelligence from multiple different

8:47teams in your organization and start

8:48putting it together and saying to the

8:50salesperson, "Here are 10 possible

8:52proactive pitches that you can go to the

8:54customer. How do we know them? Because

8:56some other team across the globe that

8:57was talking to the customer, talking to

8:59junior person, got some intelligence

9:00that these things might be important for

9:02them."

9:03Does that kind of make sense? Yeah.

9:04Yeah. So, and and I guess the logical

9:07next question is how do you What's the

9:08kind of compute that's working here to

9:10put all of this together in in the way

9:13things are today or the kind of what

9:15what you're seeing around us right now?

9:17So, I'll give an example of, for

9:18example, how we built it. We built a

9:20platform called Context Fabric, a

9:22non-compete Work Fabric. The compute is

9:25it's a combination, actually. It's it's

9:27a sort of different models. There are

9:28models that run at the edge, Mhm. uh, on

9:31edge devices because they capture how

9:32humans interact with their devices. Uh,

9:35and then you have some models that run,

9:38uh, on GPUs in the cloud. And so, it's

9:40kind of this it's shared between the

9:42two. You capture, you know, these

9:45interactions at one point, filter them,

9:48um, and so on, and then assemble, um,

9:51put it all together, make sense of it in

9:53the cloud. Mhm. Backed by GPUs and so

9:55on. So, let's go back to the sales uh,

9:58sales model that you talked about. And

10:00if that company today an IT services

10:02company is trying to move to a situation

10:06where let's say all this data is being

10:07captured and so on. So, who does it for

10:10them? I mean of course an IT services

10:12company will do it itself, but who does

10:13suppose this was not an IT services

10:15company or it was an automotive company

10:16or what I'm talking about is not is a

10:20fundamental shift.

10:21>> Mhm. The reason it's a fundamental shift

10:23is see for the last 40 years we have

10:25lived in a computing paradigm mhm

10:27where you've had a microprocessor

10:29and the world is built around the

10:31microprocessor, right? You build

10:32products and services around it. And now

10:34what did that mean? That means that the

10:36microprocessor is really good at the

10:37following. You give it a specific

10:39detailed

10:41instruction.

10:42Like add this number to this number and

10:44it does exactly that.

10:46If you if you get an error, it's because

10:49you made an error. You the human said

10:51something that didn't make sense. So,

10:53this is how we wrote code, we did

10:55migrations, they've done, we built all

10:57kinds of things on this. But what has

10:59changed now is you now have different

11:00fabric of computing where

11:03um it's no longer about giving it a

11:05specific instruction. You give it an

11:07intent. Yeah. And it gives you a range

11:09of possibilities of outcomes.

11:11We've never really dealt with computing

11:13devices or or or abstractions like this

11:16in enterprise. And that is a whole new

11:18world.

11:19Because in this world you no longer will

11:21spend the same kind of time and effort

11:23giving detailed specific instructions

11:26writing reams of code in the same way

11:28etc. etc. and so on. So, everything

11:30changes with this, right? And so,

11:32whether it is IT services or any of

11:34these kind any company in any of these

11:35industries

11:37now you have to operate in a world where

11:39fundamentally your computer is

11:40stochastic or non-deterministic. Mhm.

11:43You say something and you I

11:45you know

11:46you get possibilities of different

11:48answers and now you can make sense of

11:49them as you wish as you wish. Mhm. We've

11:51never had an opportunity like this

11:53before.

11:54Right. And so it is in this world that

11:56you can situate sort of these kinds of

11:58things. What does context mean? Context

12:01is literally means what it means in the

12:02English language, right? Which is

12:04you have these non-deterministic

12:05computing machines and you tell them

12:07these are models, you tell them hey this

12:09is the context in which I'm operating,

12:10this is my goal

12:12and that's an intent.

12:14And now the machine tries to achieve it

12:16it thinks it understands or you know

12:18some part of your goal and tries to

12:19achieve it within that and so on.

12:21Now that is the big shift or that's the

12:23big change. And and therefore what it

12:25means, right? Is if you take IT services

12:27companies as an example for 40 years

12:29they've mastered writing code. Mhm.

12:33Um

12:33in my opinion

12:35uh I think the future for them now is

12:39you instead of placing the micro the

12:40microprocessor at the center and saying

12:42we do services based on the

12:44microprocessor, they place the model at

12:46the center. And therefore all of them

12:48are about engineering context

12:51into these models.

12:52And so the future of IT services, at

12:54least according to me

12:55uh or at least the future of the

12:58traditional notion of IT services I and

13:00I'm not including BPO in that

13:02>> Mhm.

13:03is shifting from writing code to

13:05microprocessors to shipping context to

13:08models.

13:09Next. And and it's interesting because

13:11you've also talked about how, you know,

13:13just using models is not going to solve

13:15your problem unless you marry it in a

13:17deep way with the context of your

13:19organization because everyone will have

13:20the models. Right. Right. So, walk us

13:22through

13:23why we should be careful in in this

13:25context. The great question. Think of it

13:27this way, right? If you and I competing

13:30um and you have the same model that I do

13:33in some sense you've democratized

13:35intelligence

13:36um

13:37and if that's the case, then how do you

13:40and I still differentiate? Because

13:41technology has been a big part of how we

13:42differentiate.

13:44Uh well, it turns out that

13:46you know,

13:47what we have done all along is we have

13:49gathered maybe I serve a certain market,

13:52you serve a different market um

13:55or you serve a set of customers and I

13:56serve a different set of customers etc.

13:58and so on. Or my response to those

14:00customers is different from yours. All

14:02of that is unique to me or unique to

14:03yourself, right? And so there is this

14:05context in which your organization works

14:08or my organization works.

14:10Um and that is the key thing that makes

14:13us different in a world where

14:14intelligence is democratized. Mhm.

14:17And so the only argument that I have

14:18there is okay, so how do you compete in

14:21a world where because earlier on

14:22intelligence was not democratized.

14:24Right? What would happen is you would

14:26hire smarter people, you'd pay better

14:27salaries and I'd pay less salaries or

14:29I'd I'd different people, right? Or etc.

14:31and so on. And so we used to compete in

14:34who has better access to market with

14:36better people with better product,

14:38better IP and so on.

14:39Well, now if all of that gets

14:41normalized, right? Then what remains is

14:43difference for us to exist as two

14:45separate organizations. The only thing

14:47is the context that our each individual

14:49company has built up over the over the

14:51past. And we if you take that context

14:54and marry it with models, then that is

14:55how you continue to differentiate in a

14:57world where intelligence is

15:00democratized. Right. It's so it's it's

15:02really saying that your analog skills

15:03matter more because the digital layer is

15:05something that's going to really be

15:07uniformized in some way. I think no, we

15:09can digitize the analog skills too as

15:10well. And that is part of the context,

15:12yeah, right? But but yes, along the

15:14lines of what you're saying, Bobby.

15:15Right. So, as you look ahead to Rohan,

15:17what are you seeing? I mean we're in 26,

15:19we've been seeing all these

15:21you know, announcements and reactions.

15:25I'm not just just talking about the

15:26stock markets worldwide who are

15:28currently in a grip of fear because of

15:30AI. How do you see this? What's how are

15:32you dividing your optimism versus

15:35concern versus pessimism? You know, the

15:37truth is it's

15:39it's

15:40I think there are probably two

15:42fundamental questions. The first of

15:44which is

15:46is the role of human beings going to

15:49continue to be what it has been so far.

15:51That's the heart of everything. And the

15:52second is there also intermediaries and

15:55is the role of intermediary is going to

15:56continue in many industries. I think

15:58both of these are being put to the test

16:00in a way that we never could have done

16:02before.

16:03Um on the first on the first thread

16:06um the way I view it is let me give sort

16:08of this this example.

16:10Um for a long time I privileged the

16:13importance of creativity in my life.

16:15Now at least to some in some areas

16:18I don't want to sound overly sort of

16:21optimistic, but in some areas

16:23this these stochastic machines

16:27just by guessing the next word, it's

16:30literally it's it has within a certain

16:31set of words it guesses the next one.

16:34I'm I'm very loosely describing.

16:36It seems to approximate

16:38the kind of output that suggests

16:40intelligence or suggests creativity more

16:42than anything else.

16:44Um I never could have imagined that's

16:46how you approximate or simulate

16:47creativity.

16:49And so what is creativity? So, I've been

16:51asking myself then is

16:52creativity intelligence really like some

16:54random function? Like so when I have an

16:56idea it's maybe that too is random and I

16:57just don't know enough neuroscience to

16:59know if it's random or not.

17:01And if that's the case, that poses a

17:02very fundamental question saying maybe

17:05you know, it's not even about having

17:06machines understand or any of these

17:08fancy words as long as machines can

17:10simulate creativity

17:11then there's a serious threat to

17:14human work in many many different areas.

17:17And I think increasingly that's going to

17:18be true.

17:19Okay. Um uh at least for certain kinds

17:22of work. Mhm. Um now it comes coming to

17:25the second the disintermediation

17:28part, right? You have

17:30analyst firms

17:31as I deliberately use them as an example

17:33and it's true for many industries or you

17:35have all these exchanges, right? Where

17:37you know, Visa, MasterCard etc. and so

17:39on.

17:40You know, a lot a lot of these things I

17:42think AI changes the game. Why is the

17:43middleman there and you have to pay the

17:44middleman's tax?

17:46Right? Because as an example in analyst

17:47firm, what do they do? They take their

17:49information arbiters. They would take

17:51information one place and and give it to

17:53another place where people want this

17:55information. But yeah, I can do this

17:56without any of its cost in in the

17:58middle. So, I think these are the two

17:59things that it it will disrupt humans

18:02human labor certain kinds of human labor

18:04and then the and then these

18:07intermediaries.

18:08Both of which I think will happen. Last

18:10question. So, as you

18:11I mean as someone younger and you refer

18:14to let's say your father's generation

18:15which set up

18:17a company like Infosys

18:19oh

18:19the younger generation is all is do you

18:21see it as really disrupting almost

18:23everything that uh

18:25previous two generations have done? Not

18:26necessarily in in this particular case,

18:28but in general

18:29computing computing technology

18:31or do you see it

18:33changing the very nature of the world in

18:36again in the context of technology or

18:37something else?

18:39No, I I don't think there is this

18:40reverence or not. I don't.

18:42Instead I

18:45I think the only question is are we able

18:47to do something with computing now that

18:48we couldn't do before? I think that's

18:49the only relevant question in my

18:51opinion.

18:52Whether that erases the past or not or

18:54builds on top of it to me that's a

18:56tertiary question.

18:57Um and absolutely that's the case that

19:01you know, what we what is happening now

19:04um moves computing forward in a way that

19:07we didn't think was possible before. I

19:09mean the fact that like my mother uses

19:11chat GPT, right? A 75-year-old lady in

19:13chat GPT. That to me is it's it's beyond

19:16just an industry now. It is uh

19:19it's societal.

19:20Right?

19:21And the fact that there's a machine that

19:23can simulate creativity in a way that we

19:24couldn't do 5 years ago, 10 years ago.

19:27This is just the needle is move forward

19:28and this will move forward even more and

19:30more and more.

19:31That's kind of how I see it. Right. Good

19:33note to end on. Thank you so much for

19:35joining me. Great. Thank you. Thank you.

19:42>> [music]