Agentic AI and the End of Traditional SaaS

[00:00:04] Today my guest is Sir Ram. He’s the
[00:00:06] group CTO of SP global where he runs a
[00:00:10] 15 person technology organization and
[00:00:12] leads Agentic AI. He spent over 30 years
[00:00:15] at the frontier of mobile cloud and AI
[00:00:17] and also has spent a decade at IBM as a
[00:00:19] distinguished engineer and master
[00:00:20] inventor with more than 150 patents to
[00:00:23] his name. It’s great to have you on the
[00:00:24] podcast.
[00:00:25] >> It’s great to meet you Ashar. Thank you
[00:00:26] for having me on.
[00:00:27] >> All right. So starting off you built a
[00:00:29] conversational AI at Oracle years before
[00:00:33] you know it was fashionable and
[00:00:34] everybody wanted a chatbot. So how was
[00:00:36] that experience and and you know what
[00:00:38] has changed back then and today where we
[00:00:41] are.
[00:00:41] >> You know I think those were the early
[00:00:43] days of AI where I think there was
[00:00:46] recognition that the nature of the human
[00:00:49] interface was changing and we were
[00:00:51] moving from you know the mouse and
[00:00:53] keyboard and maybe a touch screen to to
[00:00:56] having a natural language conversation
[00:00:58] with intelligence or with a software
[00:01:00] system. And so the the layer was all
[00:01:03] about, you know, the human interface,
[00:01:05] you know, replacing a traditional user
[00:01:07] interface with a voice- based or a
[00:01:09] textbased interface. And, you know,
[00:01:11] those were the days when when teenagers
[00:01:14] and a lot of young people did not want
[00:01:16] to download an app for everything. You
[00:01:18] know, they wanted to stay in their
[00:01:19] current messaging channels, you know,
[00:01:21] like WhatsApp or or Telegram, and they
[00:01:24] wanted to be able to talk with
[00:01:25] enterprise systems, you know, on on the
[00:01:27] other end. and and so the conversational
[00:01:29] AI layer was all about that was about
[00:01:31] understanding intents and entities. It
[00:01:33] was early days. I think what’s changed
[00:01:35] now is that intelligence has moved is
[00:01:38] moving in into the software system
[00:01:40] itself. So interactions over time will
[00:01:43] be less and less initiated by humans.
[00:01:46] you know these systems will decide when
[00:01:48] to reach out to you and tell you when
[00:01:49] you need to know something you know and
[00:01:51] ask you for permission to do something
[00:01:53] as opposed to the human initiating you
[00:01:55] know the interaction with the software
[00:01:57] system which is still typically the case
[00:01:58] but that’s what’s changing
[00:01:59] >> right one of the things that I typically
[00:02:01] tell to people is that there used to be
[00:02:03] some real a IML work and then there’s
[00:02:06] LLM and building rappers around them
[00:02:08] right so back in that day I think from a
[00:02:10] technology or implementation perspective
[00:02:12] I guess that was the biggest difference
[00:02:14] that you can see that a lot inference
[00:02:16] engines had to be written down. Data had
[00:02:18] to be trained and it was much much much
[00:02:20] harder to build a chatbot or natural
[00:02:23] language processing versus as of today.
[00:02:25] >> It’s a it’s a great observation. I I
[00:02:27] think you know we we used to spend time
[00:02:29] you know 8 years ago on on building
[00:02:31] these very complex pipelines of of data.
[00:02:34] We would label data. We would curate
[00:02:37] these data sets. We had different data
[00:02:39] sets that drove the creation of a model.
[00:02:42] Whether it was a machine learning model
[00:02:43] or whether it was a neural networking
[00:02:45] model, we still had the concept of
[00:02:46] labelled data sets. That is still there.
[00:02:49] I think the two things that have changed
[00:02:50] is you don’t need to be a data scientist
[00:02:52] to build world-class AI applications
[00:02:54] anymore. You you you can be an engineer
[00:02:57] and do that. And I think that’s what LLM
[00:02:59] have changed. They’ve changed not just
[00:03:01] the fact that you can innovate around
[00:03:02] them. They’ve also changed the fact that
[00:03:04] you don’t need to be a data scientist.
[00:03:06] But the process of creating a custom
[00:03:08] neural networking model or a custom
[00:03:10] machine learning model. They make that
[00:03:11] dramatically easier. You know, the LLM
[00:03:13] help you with that with all of these
[00:03:14] programming tools like cloud code that
[00:03:17] we use heavily. You know, they’ve really
[00:03:19] changed the game in terms of how fast
[00:03:21] and how efficiently you can put out
[00:03:23] business functionality and business
[00:03:25] applications. You know, leveraging AI.
[00:03:27] >> I think it’s more of more becoming a
[00:03:30] level playing field, right? So a lot of
[00:03:32] people can now do a lot of complex
[00:03:34] technical stuff that only a certain
[00:03:36] faction of the people used to do or you
[00:03:38] could do and now all of that knowledge
[00:03:40] is available to everybody to be able to
[00:03:42] build it easy and it’s becoming much and
[00:03:44] more easier. And once that kind of you
[00:03:46] know kind of takes off then you’ll see a
[00:03:48] lot of other innovations also coming in
[00:03:50] because again it’s a now since it’s a
[00:03:52] level playing a field anybody everybody
[00:03:54] can come up and start doing it. So but
[00:03:56] that was interesting. I mean uh I just
[00:03:58] you know I wondered what you would be
[00:04:00] thinking that hey we used to do this
[00:04:02] like five 10 years back and it was so
[00:04:04] complex and now it’s so anybody can just
[00:04:07] you know turn out the you know the thing
[00:04:09] that’s changed the the big thing that’s
[00:04:11] changed in in my view is that you know
[00:04:14] you you had AI being democratized so
[00:04:17] that everybody could use it as you were
[00:04:18] saying but you also had you have
[00:04:20] knowledge that’s being democratized and
[00:04:22] made available to you in the palm of
[00:04:24] your hand via your mobile device but if
[00:04:26] you’re a startup today, if you’re a
[00:04:28] founder, you know, four or five years
[00:04:29] ago, your biggest challenge was to go
[00:04:31] build your MVP, you know, was to figure
[00:04:33] out what you were going to build and and
[00:04:35] you you were potentially non-technical
[00:04:38] or you were not technical enough and you
[00:04:40] needed help. You had to go find that
[00:04:41] engineering leader. You had to go find
[00:04:43] engineering resources. You had to go and
[00:04:45] do that work. And and that’s what’s
[00:04:47] fundamentally changed now because a
[00:04:48] smart founder who who has domain
[00:04:50] expertise who knows and understands the
[00:04:53] problem they’re trying to solve better
[00:04:55] than anyone else now has the tools to go
[00:04:58] build an MVP all by themsel and that’s
[00:05:01] what it changed. So this notion of a
[00:05:03] one-man founder, you know, who’s using
[00:05:04] AI to, you know, do product management,
[00:05:07] you know, engineering, you know, DevOps,
[00:05:09] but also his marketing and his insight
[00:05:12] sales and lead genen, all of those
[00:05:14] things that you needed people for, you
[00:05:16] can now do with AI. And I think, you
[00:05:17] know, those were some of the sentiments
[00:05:18] you were you were expressing and I just
[00:05:20] use different words to describe them.
[00:05:21] >> I that’s fine. I agree. Yes, that’s
[00:05:24] that’s where it is. So that’s good. Now
[00:05:26] coming back to a couple of things where
[00:05:28] you know we talk about so a lot of
[00:05:30] organizations and people that we’re
[00:05:32] working in today they throw around these
[00:05:34] different terms right so BPA business
[00:05:36] process automation RPS journal
[00:05:38] automations AI agents and and I feel
[00:05:41] that a lot of those people just assume
[00:05:43] that’s they’re all similar right so can
[00:05:46] you kind of help draw the actual lines
[00:05:48] what what’s generally different about an
[00:05:50] AI agent versus you know RPA or BPS or
[00:05:53] >> Absolutely. Absolutely. I I think you
[00:05:55] know when the biggest innovation you
[00:05:57] know if you look at you know like 2018
[00:05:59] and 2019 you know there was a lot of
[00:06:01] talk of robotic process automation where
[00:06:03] you would have a piece of software
[00:06:05] essentially log into most of the time a
[00:06:08] web browser based application where it
[00:06:10] would go and try to simulate in exact
[00:06:14] great detail what a human did to fulfill
[00:06:17] a specific workflow you know so if I had
[00:06:19] to go and for example log into a
[00:06:22] software system you know and approve a
[00:06:23] purchase order based on some criteria
[00:06:26] like comparing terms and conditions or
[00:06:28] dollar amounts or dates or verification,
[00:06:32] approving an invoice, verifying it with
[00:06:34] a PO on an ERP system. You know, there
[00:06:37] were there were a set of flows that were
[00:06:39] exact and and the software would follow
[00:06:41] that exact path. And if the developer of
[00:06:44] the ERP changed the UI by a few pixels,
[00:06:47] the entire RPA automation would break.
[00:06:50] it would fundamentally break you know
[00:06:51] and nowadays an agent is especially if
[00:06:54] you look at tools like cloud co-work or
[00:06:56] if you look at openclaw for example
[00:06:59] they’re able to manipulate browsers in
[00:07:00] the way that a human is so if you if you
[00:07:03] for example want to file your expense
[00:07:04] reports and when we do this at work in
[00:07:06] in at my workplace you simply point
[00:07:08] cloud co-work at your receipts and you
[00:07:11] tell it the system to go log into to
[00:07:13] file the expense reports without telling
[00:07:14] it anything about how that system
[00:07:16] actually works it just figures it out it
[00:07:18] it’s able to start a browser session.
[00:07:20] It’s able to log in. It’s able to, you
[00:07:22] know, navigate the pages of the expense
[00:07:24] reporting application and just do the
[00:07:26] job, you know, without any human
[00:07:27] intervention. And it’s able to do this
[00:07:28] without training, without doing anything
[00:07:30] because I think, you know, it’s just
[00:07:32] smart enough to understand context and
[00:07:34] it’s been trained on how to use these
[00:07:35] systems before, right? And so I think
[00:07:38] that’s the big change is a software
[00:07:40] agent works like a human who is learning
[00:07:43] and figuring things out and it’s able to
[00:07:46] handle a lot of variations. It’s not a
[00:07:49] fixed flow you know where correct which
[00:07:51] is very brittle and that has to be
[00:07:53] monitored. So I think that’s the big
[00:07:55] change and and you can the other big
[00:07:57] change is that to build an RPI you you
[00:07:59] you needed some developer who understood
[00:08:02] a specific toolkit from UiPath or
[00:08:04] whatever. You don’t need any of that.
[00:08:06] All you need now is a business user who
[00:08:08] can describe a specific flow build
[00:08:11] something called a skill which is
[00:08:13] nothing but a mark very simple markdown
[00:08:15] and the agent reads that learns it
[00:08:17] executes it. you can schedule the skill
[00:08:19] you it it is able to do all of that
[00:08:21] without requiring a technical person. So
[00:08:23] the process of designing and building an
[00:08:25] agent and deploying it to work for you
[00:08:28] on your desktop is is dramatically
[00:08:30] simple. Now you still need engineers to
[00:08:32] make sure that the agents work properly
[00:08:33] on the cloud, secure it, manage it and
[00:08:35] do the agent life cycle management. You
[00:08:37] know, just like application life cycle
[00:08:38] management, you have this concept of
[00:08:40] agent life cycle management where you
[00:08:41] have to manage the life cycle of the
[00:08:43] agent. You have to scale it, you have to
[00:08:44] secure it, you have to manage it, you
[00:08:45] have to do all of these very important
[00:08:46] things, you know, and ideally you’re
[00:08:48] doing it in the cloud or you’re doing it
[00:08:49] in in some place that you’re hosting in
[00:08:51] a central way and the agent is not
[00:08:53] running on desktops. But I think those
[00:08:55] are the big differences between the old
[00:08:57] RPA world and these new and agentic AI,
[00:09:00] you know,
[00:09:01] >> right? So autonomous decision making,
[00:09:03] not getting stuck, figuring out what to
[00:09:04] do, what not to do. So yeah, absolutely.
[00:09:07] So that’s what AI agents is versus RPA
[00:09:10] or your standard good old you know
[00:09:12] previous workflow automation was a set
[00:09:14] of decisions and everything.
[00:09:15] >> But haven’t you also notice that in some
[00:09:18] orgs or in some processes you might not
[00:09:20] even need an AI agent right? So a lot of
[00:09:22] organizations they come up and and they
[00:09:24] describe what they’re doing maybe you
[00:09:26] know proper just a good old workflow
[00:09:29] automation sequence of where the
[00:09:30] decisions are fixed. you don’t probably
[00:09:32] need an agent because then installing an
[00:09:34] agent also means spending tokens and
[00:09:36] money and hosting and all of that as
[00:09:38] well.
[00:09:38] >> Has that been an experience or no?
[00:09:40] >> That’s a brilliant question. I think you
[00:09:42] asked like two or three separate things
[00:09:44] there and I and I want to address each
[00:09:45] of those important things. First is I
[00:09:47] think you’re absolutely right when you
[00:09:48] say that the biggest value today in
[00:09:51] automation whether you use AI or not the
[00:09:53] lowest hanging fruit in terms of
[00:09:55] eliminating repetitive actions without a
[00:09:58] whole bunch of variability. So in every
[00:10:00] enterprise you’ve got this lowhanging
[00:10:02] fruit. Now the question is are you going
[00:10:03] to automate it? And the answer has to be
[00:10:05] absolutely you have to answer you have
[00:10:07] to automate it. Now do you need AI to
[00:10:08] automated it? I think in a lot of cases
[00:10:11] you may not. You may not need AI. I
[00:10:13] think if you use a traditional workflow
[00:10:15] automation tool if especially if you if
[00:10:17] you’ve got the maturity to be API based
[00:10:19] you know and you have a set of APIs that
[00:10:21] helps you automate. So you don’t need to
[00:10:23] go the RPA route. You know, you you can
[00:10:25] do a lot of things in in many
[00:10:27] organizations where you’re where you’re
[00:10:28] automating a lot of repetitive business
[00:10:30] workflows. The problem I think comes in
[00:10:32] in in a lot of organizations where they
[00:10:34] don’t have the maturity. They haven’t
[00:10:36] built you know this enterprise service
[00:10:38] bus. They don’t have APIs for
[00:10:40] everything. They’ve got a lot of humans
[00:10:42] who are manipulating browsers who are
[00:10:44] basically logging in and doing a lot of
[00:10:46] things using the browser. They’re
[00:10:48] reading documents. You know, you have
[00:10:49] document related workflows. they’re
[00:10:51] reading documents that they have to go
[00:10:52] extract something out of and then they
[00:10:54] have to automate that workflow as a
[00:10:55] result of what they read from the
[00:10:57] document. So you’ve got this need in
[00:10:59] many cases to use AI. But even there,
[00:11:02] you know, I think you still need a lot
[00:11:04] of traditional tools. You need a
[00:11:06] workflow engine. You know, you need you
[00:11:07] obviously need an AI based document
[00:11:09] recognition platform or a document
[00:11:10] understanding platform. You have that.
[00:11:11] You plug in, you know, your business
[00:11:13] rules and you have what you would need
[00:11:15] for a huge amount of automation, right?
[00:11:18] So I don’t think you need AI. you may
[00:11:20] need AI depending on your use case and
[00:11:22] if you do it’s become dramatically
[00:11:24] easier to do now. So leading on to that
[00:11:26] question, you’re running Agentic AI in
[00:11:28] production for several Fortune 100
[00:11:30] clients, right? So what are the things
[00:11:32] or what are the use cases maybe that
[00:11:34] you’ve noticed that agents are genuinely
[00:11:36] doing very reliably today versus where
[00:11:39] people are just you know there but
[00:11:41] they’re not and people just overestimate
[00:11:43] something that they agents can do and
[00:11:45] >> one is you know how we leverage AI and
[00:11:47] ATKI use cases internally within the
[00:11:49] company and then how we’re leveraging
[00:11:51] infusing aki into our into the products
[00:11:54] and software that we build that we
[00:11:56] deliver you know customer solutions
[00:11:57] based on to to a lot of our customers
[00:11:59] worldwide. So if you look at and I’ll
[00:12:02] also touch on the token question which I
[00:12:03] don’t think I I answered you know the
[00:12:05] previous question that you asked I don’t
[00:12:06] think I answered that I I will touch on
[00:12:08] that what is happening is that
[00:12:10] internally we’ve got a number of
[00:12:12] repetitive use cases whether you look at
[00:12:14] our marketing functions whether you look
[00:12:16] at our security operations center
[00:12:18] functions whether you look at number of
[00:12:20] other functions that happen within the
[00:12:21] enterprise you you’ve got human beings
[00:12:23] doing the same thing over and over again
[00:12:25] and we we’ve identified the highest
[00:12:27] value workflows and we’ve automated them
[00:12:29] used a combination of claw hardcore work
[00:12:30] you know to do that we’ve completely
[00:12:32] redefined our software SDLC you know we
[00:12:34] we now call it our own internal AI SDLC
[00:12:37] and we have you know a cluster of nine
[00:12:39] agents that we use in the software
[00:12:42] development life cycle that actually do
[00:12:44] a lot of work all the way starting from
[00:12:46] design and requirements all the way down
[00:12:48] to you know DevOps deployment managing
[00:12:51] testing all of that right so we’ve got a
[00:12:53] we’ve got this these agents helping us
[00:12:56] and these agents are are superstars you
[00:12:58] know they they’ve helped us you know get
[00:12:59] a lot of efficiency and increase the
[00:13:02] throughput of every engineer
[00:13:03] dramatically. You know, so we use use
[00:13:05] that extensively internally. On the
[00:13:07] outbound side, if you look at the
[00:13:08] products that we sell, for many of our
[00:13:11] products, we now sell agent that come
[00:13:13] with those product that are able to use
[00:13:14] those products on your behalf if
[00:13:16] necessary. Our agents understand our
[00:13:19] software very well. They’re already like
[00:13:21] they have the context that they need to
[00:13:23] go operate on our software. And we’ve
[00:13:24] also gone and built MCPs for all of the
[00:13:28] systems we care about in the back end in
[00:13:30] the sense that you know to access a
[00:13:31] backend system oh I don’t have an API
[00:13:33] for that system you know oh I have to go
[00:13:35] build an API you don’t need to because
[00:13:36] in this world of of AI you know the
[00:13:38] integration pattern has moved from being
[00:13:41] completely API based and that being the
[00:13:43] recommended method and it’s still an
[00:13:44] amazingly effective method and we should
[00:13:46] still make and build headless systems
[00:13:48] especially if you’re building new
[00:13:49] systems it has to be built in the right
[00:13:51] way and you have to expose an API for
[00:13:53] everything and you have to have the user
[00:13:55] interface using the same APIs that you
[00:13:57] expose to other machines. But if you
[00:13:59] don’t have an API, you have this notion
[00:14:01] of an MCP now where an agent can
[00:14:03] communicate with other agents, you know,
[00:14:04] delegate work to other agents, co access
[00:14:06] data that’s sitting in backend systems,
[00:14:08] apply logic that may need to be applied
[00:14:10] to that data. And so all of those
[00:14:12] things, the key technology components
[00:14:14] that you need are are in place for us to
[00:14:16] bypass the API economy, you know, and
[00:14:19] you bypass the API economy and it takes
[00:14:21] us into what you’re talking about, the
[00:14:22] token economy, right? and and what are
[00:14:24] we looking at right for us? We are LLM
[00:14:27] agnostic. You know, we we don’t rely on
[00:14:29] a specific LLM. We look at multiple LLMs
[00:14:32] to go do a specific job and if I can get
[00:14:35] better quality. I’ve got certain
[00:14:36] workloads like like in in the security
[00:14:38] case, you know, I I’ve got an agent that
[00:14:40] behaves like a security operations
[00:14:42] center level one agent. Okay. So, it’s
[00:14:44] triaging security problems and when it’s
[00:14:46] doing that, it’s going to virus total.
[00:14:48] It’s checking an alert. It’s doing all
[00:14:50] these things that a human would do. But
[00:14:51] we don’t just trust one model. We use an
[00:14:53] ensemble of LLMs there in that case. And
[00:14:56] in other cases that are of lower value,
[00:14:58] we may not need to use, you know, the
[00:15:00] Frontier, the latest Frontier model. We
[00:15:03] can use a version of the model where I
[00:15:04] can get a million tokens for six bucks.
[00:15:06] But I also think there’s a version of
[00:15:08] this where you have your own hosted
[00:15:09] model. These models that are coming out,
[00:15:11] especially from the Chinese, are really
[00:15:13] good. They are good enough to be good
[00:15:16] enough. So, so it’s moving from a
[00:15:18] token-based economy to one where tokens
[00:15:21] don’t matter if you have a model that
[00:15:23] you can use on your own, you know, lower
[00:15:25] scale GPU farm in your own
[00:15:27] infrastructure or in your own private
[00:15:29] cloud. And when you do that, not only
[00:15:31] are you out of the token game, but you
[00:15:33] get results that may be good enough for
[00:15:35] a specific use case. So, it all depends
[00:15:37] on the use case. If I’m doing some deep
[00:15:38] researchy thing, I may need like the
[00:15:40] latest models, but the reality is that
[00:15:42] in an enterprise, you don’t. have most
[00:15:44] of the decisions you make are much
[00:15:46] simpler. If you’re trying to automate
[00:15:48] repetitive workflows, there’s not this
[00:15:50] massive like complex decision making
[00:15:52] going on. It’s not happening. So you
[00:15:54] you’re better off with a Quen model.
[00:15:56] You’re better off with a Kimmy model
[00:15:58] that just came out a few weeks ago.
[00:15:59] These are openweight models that you can
[00:16:01] train to your needs and use in your own
[00:16:03] infrastructure, right? and and get out
[00:16:05] of what I call token jet that I think
[00:16:07] over time the next 6 8 12 18 months I
[00:16:10] think this token issue will get fixed
[00:16:12] you know because every single
[00:16:13] participant in this economy is working
[00:16:15] on bringing token costs down anyway have
[00:16:17] you found organizations they are willing
[00:16:19] to so are they more inclined toward
[00:16:21] on-prem systems or they are okay with
[00:16:23] cloud-based systems because generally in
[00:16:26] the past that’s been a major hurdle with
[00:16:28] odds right they wanted data to stay with
[00:16:30] them but they they got okay over time
[00:16:32] with AWS
[00:16:34] and all of these models as well. But
[00:16:36] then you now we going back to setting up
[00:16:38] GPU farms and have disaster recovery GPU
[00:16:41] farms and all of that hardware
[00:16:43] procurement because that’s a that’s a
[00:16:45] whole lot of challenge in its own self,
[00:16:47] right? It’s like going back to where we
[00:16:50] transformed from 50 years back.
[00:16:51] >> Absolutely. You’re right. Look, I think
[00:16:53] you know it’s it’s a traditional capex
[00:16:54] versus appex is it
[00:16:55] >> correct? Are you going to invest in
[00:16:57] something and you need it longterm? Are
[00:16:58] you going to manage or is it much easier
[00:17:00] to go to the cloud? almost every every
[00:17:01] single time, especially for the small
[00:17:03] player. It’s easier to go to the cloud
[00:17:05] to get what you need and to do what you
[00:17:07] need to. And the cloud gives you amazing
[00:17:08] benefit. You know, it it gives you
[00:17:10] scalability. Arguably, it gives you
[00:17:12] better security. It gives you all of
[00:17:14] these things like, you know, disaster
[00:17:15] recovery, failover, better reliability,
[00:17:17] redundancy, all of those things, right?
[00:17:19] And you can manage the the experience
[00:17:21] that you provide to your customers
[00:17:22] because all of the tools are there. You
[00:17:23] know I I I think if you are in a
[00:17:25] regulated industry like healthcare or
[00:17:27] financial services and you have certain
[00:17:29] unique needs you know that’s where some
[00:17:31] of this debate becomes you know a little
[00:17:33] more relevant. I already have
[00:17:34] investments in my data center. You know,
[00:17:36] I I’ve got a private cloud which is very
[00:17:39] important. You know, I can I can do
[00:17:40] things in the context of my VPC or my
[00:17:42] private cloud. Even though Amazon, you
[00:17:43] know, or Azure or GCP is who manages
[00:17:46] that private cloud, it’s still my
[00:17:47] private cloud. It’s like my virtual data
[00:17:49] center in the cloud. So, you can do it
[00:17:51] there. But the recent uh you know, OBBB
[00:17:53] that President Trump passed, there are
[00:17:55] some massive depreciation, you know,
[00:17:57] benefits there. You’re able to take 100%
[00:18:00] depreciation on any capex that you
[00:18:01] invest, which I think is is huge. You
[00:18:03] know, I think you can do some very
[00:18:04] interesting things with that. But it is
[00:18:06] a trade-off. You know, I think it’s
[00:18:07] always the cost of entry is much lower
[00:18:10] the bar of entry when you go with the
[00:18:12] cloud. That’s the key. You know, you and
[00:18:13] you can manage your cost, you know, much
[00:18:15] better on a cloud. You have the right
[00:18:17] controls, you know, whereas you have to
[00:18:18] think about other considerations when
[00:18:20] you have your own data center like we
[00:18:21] do.
[00:18:21] >> I think and I we we kind of addressed it
[00:18:23] building MCP servers, getting API free,
[00:18:26] but a lot of AI or applied AI can only
[00:18:29] be done once the foundational work is
[00:18:31] right. Right. So a lot of these orgs are
[00:18:34] it’s sitting on decades old systems and
[00:18:36] and uh what we’ve felt as well and what
[00:18:39] I’ve seen in the I mean and I’d love to
[00:18:41] hear from you how much of the real work
[00:18:43] before really getting to applying AI
[00:18:45] because data previously used to be
[00:18:46] garbage in garbage out now it’s garbage
[00:18:48] in garbage square multiplied out right
[00:18:51] because you’re actually relying on
[00:18:53] decisions that that impact a lot of
[00:18:55] areas right so how important do you
[00:18:57] think or how much of the real work gets
[00:18:59] really done in modernization or
[00:19:01] modernizing that foundational work first
[00:19:03] and what kind of modernization is really
[00:19:04] needed? Do we just bear layers on top
[00:19:07] that organizes the data? Do we put in
[00:19:09] ESBs? Do we put in technology changes?
[00:19:12] What are your thoughts?
[00:19:13] >> I think there are multiple very
[00:19:14] important questions here actually uh you
[00:19:16] know if I can you know break those out.
[00:19:18] One is you know I I I think you’ve got a
[00:19:20] traditional consulting type approach
[00:19:23] where you go into you know an existing
[00:19:25] customer environment and and we don’t do
[00:19:27] professional services. We don’t do IT
[00:19:28] service consulting at all. We don’t sell
[00:19:30] that as a service. What we do sell is
[00:19:32] FTEEs, you know, we we we’ve got an FD
[00:19:35] team and we go deploy a team of FDE to
[00:19:37] figure out a specific AI related problem
[00:19:39] for you to see if AI can shortcut a lot
[00:19:42] of things. We typically don’t go and we
[00:19:45] we’re not going to build APIs for you.
[00:19:46] We’re not going to API enable the
[00:19:48] backend system. We don’t we don’t do any
[00:19:50] of that. I think the consulting game has
[00:19:51] morphed into an FD game where you’re
[00:19:54] going in and trying to in 30 60 90 days
[00:19:58] leverage you know a frontier model or
[00:20:00] any model you know to go deliver
[00:20:02] automation for a list of use cases that
[00:20:05] you identify and to build the governance
[00:20:06] around it to have a light LM to have
[00:20:08] guardrails at an enterprise level to
[00:20:11] have guardrails at a division or
[00:20:12] business unit level you know to enforce
[00:20:14] those guardrails at runtime to figure
[00:20:16] out what backend model to use based on
[00:20:19] what you know And and so you you’ve got
[00:20:21] this proxy infrastructure where you have
[00:20:23] agent AI authorization. You’ve got human
[00:20:26] in the loop built in for all of your
[00:20:27] important use cases. So you’ve got so
[00:20:30] that what a consulting business means is
[00:20:32] is is changing a little bit. You know,
[00:20:34] it’s not the way it used to be. What
[00:20:36] makes all this interesting is that the
[00:20:38] value of SAS is reducing. So if you are
[00:20:41] a software company today, you had this
[00:20:43] IP based mode, you know, and and you’re
[00:20:45] like, well, I’m going to charge you, you
[00:20:46] know, a thousand employees per month,
[00:20:48] you know, and 70% of these employees
[00:20:50] logged into Salesforce or Service Now
[00:20:52] two or three times a month. 10% of the
[00:20:54] employees logged in, you know, 50 times
[00:20:56] a day, but you were charging for a,000
[00:20:58] regardless of who used what. So that
[00:21:00] software pricing model is fundamentally
[00:21:02] broken. It’ll take a while to change and
[00:21:04] it’s changing. But you you can see all
[00:21:06] of this reflected in the SAS evaluation.
[00:21:08] in the valuation you you’re seeing
[00:21:10] massive drops and and it’s a one-way
[00:21:12] street. It’s not coming back because the
[00:21:13] IP mode that these companies have,
[00:21:15] they’re not that great anymore because
[00:21:17] AI
[00:21:18] >> no IP is anymore. There’s no IP anymore.
[00:21:20] >> Exactly. The the value of the IP that
[00:21:23] they were claiming that they have is not
[00:21:24] all that. Right. So they are realizing
[00:21:27] that they’re going through you know big
[00:21:29] changes in their own road map but
[00:21:31] ultimately enterprises are going to be
[00:21:33] this hybrid of agents and humans working
[00:21:35] together. There’s always going to be a
[00:21:37] human in the loop for critical tasks in
[00:21:39] regulated industries. you know, you are
[00:21:41] going to use some SAS software, but how
[00:21:43] you get charged for it is going to
[00:21:45] fundamentally change. And there’s a huge
[00:21:46] opportunity for this FTE concept to go
[00:21:48] in and if you’re a forward deployed
[00:21:50] engineer to go and very quickly leverage
[00:21:53] AI to bypass an entire generation. You
[00:21:55] know, like I’m not going to build APIs.
[00:21:57] I don’t need APIs. You know, I’m going
[00:21:58] to take what you have as is today. I’m
[00:22:01] not going to ask you to forklift upgrade
[00:22:02] your backend ERPs. None of that. I’m
[00:22:04] just going to put this AI middleware
[00:22:06] operating system, so to speak, in your
[00:22:08] enterprise. that’s just going to manage
[00:22:10] the mess that you already have for a lot
[00:22:12] less. You know, I’m just going to manage
[00:22:14] your mess and automate your mess as is
[00:22:16] with minimum work because I have these
[00:22:18] agents who are smart enough to do a lot
[00:22:20] of repetitive things and you have to
[00:22:21] build the governance framework and this
[00:22:23] management framework around it which is
[00:22:25] where I think a lot of interesting you
[00:22:26] know consulting opportunity now lies.
[00:22:28] >> What you’re saying is that we don’t need
[00:22:30] to modernize the foundation. we maybe
[00:22:32] connect agents directly to the data and
[00:22:35] they can make sense out of that mess
[00:22:38] that’s there.
[00:22:39] >> In many cases that that may absolutely
[00:22:41] be the case and because customers don’t
[00:22:43] have time or patience for some one year
[00:22:46] like forklift you know this big big
[00:22:49] thing anymore. You know they just they
[00:22:50] just don’t have time and then they’re
[00:22:52] not going to wait. They want everything
[00:22:53] right away. And you know they have
[00:22:55] business cases that operate in quarters
[00:22:56] you know 30 60 90 days and and they have
[00:22:59] these business cases that operate and AI
[00:23:01] now gives you this unique set of tools
[00:23:03] that if you life cycle manage and govern
[00:23:05] properly and you have the right human in
[00:23:06] the loop policies for critical things
[00:23:08] you’re in good shape.
[00:23:09] >> Yeah. No that’s uh that does make sense
[00:23:11] on the SAS side of things. I do I do
[00:23:14] have a different point of view at times.
[00:23:15] I feel that I know people say that seed
[00:23:17] press pricing is going away and it
[00:23:20] doesn’t work in that way. you should
[00:23:21] move to outcomes or usage based pricing.
[00:23:24] I do feel that usage based pricing or
[00:23:27] token based pricing would actually be
[00:23:29] detrimental for the for the customers
[00:23:31] because imagine you have Salesforce
[00:23:33] right so you have 100 seats and you can
[00:23:35] add in unlimited leads and in your CRM
[00:23:38] and work around them versus this flips
[00:23:40] where now you have to pay per lead or
[00:23:42] something of those because that’s where
[00:23:44] usage is right so your cost increases
[00:23:47] could potentially increase substantially
[00:23:49] you might be better off with a seat
[00:23:51] usage pricing or or something on those
[00:23:53] lines because it’s like for tasks that
[00:23:55] were being get done for free within that
[00:23:57] seat, the seat is removed and all of a
[00:23:59] sudden you are paying for usage. So
[00:24:01] maybe that’s uh I mean I do think that
[00:24:03] that’s the case because this token is I
[00:24:05] don’t know if you’ve played arcade games
[00:24:06] when you were a kid. Uh you used to put
[00:24:08] in those tokens and you used to keep on
[00:24:10] playing and playing and playing, right?
[00:24:11] I do feel that that that could be that
[00:24:13] could be actually not so good for the
[00:24:15] customers. But you know time will tell.
[00:24:17] >> Yeah. Let’s let’s let’s talk about this.
[00:24:19] Think about a world and I I I’m making
[00:24:21] this prediction. Think about a world
[00:24:22] where the cost of a token tends to zero.
[00:24:25] How how would that change your thinking?
[00:24:27] So for example, if you look at cloud, if
[00:24:30] I use an older model, it’s six bucks for
[00:24:32] a million tokens now. You know,
[00:24:33] >> that’s what changes right now because
[00:24:34] they’re paying for it, right? So
[00:24:36] somebody is paying for those tokens
[00:24:37] right now.
[00:24:38] >> That’s right. That’s right. And and
[00:24:39] you’re right, there’s a small set of
[00:24:41] people who who are, you know, who are
[00:24:43] consuming most of the compute and they
[00:24:45] may be subsidizing the cost of a token.
[00:24:47] You you could be right. you also have
[00:24:49] have these localized models you know but
[00:24:50] I but I think over time cost of a token
[00:24:53] will will continue to plummet and I
[00:24:55] think you’re going to have many open
[00:24:56] source models right including you know
[00:24:59] meta meta is looking at open sourcing a
[00:25:00] lot of key models so you’ve got people
[00:25:02] that are looking at open sourcing the
[00:25:04] existing model right open and and
[00:25:06] sharing the weights of the model so you
[00:25:07] can train the model to your enterprise
[00:25:09] use cases there’s a second angle there’s
[00:25:11] a lot of concern in the industry that by
[00:25:14] using these frontier models a lot of
[00:25:16] enterprise related IP gets leaked to the
[00:25:19] frontier model vendors huge concern
[00:25:21] they’re worried about. So they’re
[00:25:22] looking for app. So I think this notion
[00:25:24] of these open weight models equal weight
[00:25:26] models that you are like you know
[00:25:27] understanding and training for your
[00:25:29] enterprise use cases I think I think
[00:25:31] it’s going to be an ongoing thing you
[00:25:33] know I I think that’s going to be there
[00:25:35] but if you look at a world where token
[00:25:36] costs reduce you know I think it becomes
[00:25:39] less of an issue do see things evolving
[00:25:42] you know along those firms. So
[00:25:44] previously it was consumer data that was
[00:25:45] being consumed now it’s business data
[00:25:48] that is being in trouble right because
[00:25:50] otherwise people always had their data
[00:25:52] in AWS and Azour the difference was it
[00:25:54] was in form of their database and
[00:25:56] everything they always had that data
[00:25:57] over there I don’t know why the first is
[00:26:00] that you know now they will be using it
[00:26:02] to train it and whatnot I mean they
[00:26:04] already had access so it’s you used to
[00:26:07] click this off button they do not use my
[00:26:09] data you do the same in cloud or or open
[00:26:12] AI or somewhere else that hey you know
[00:26:14] so
[00:26:15] >> it’s it’s a good observation yeah it’s
[00:26:17] it’s a very good observation I think you
[00:26:18] you you’ve got most industries and then
[00:26:20] you have some of these regulated
[00:26:21] industries you know you’ve got laws and
[00:26:23] standards like if you look at the the
[00:26:24] European AI act which we have to deal
[00:26:26] with because we have so many customers
[00:26:27] in Europe you know they they are they’ve
[00:26:30] always been much more careful you know
[00:26:32] so if you look at state of California
[00:26:33] and the United States they’ve kind of
[00:26:35] been on the front end of some of these
[00:26:36] regulations when it comes to data and
[00:26:38] privacy and things like that we have
[00:26:40] customers in Europe you know this is
[00:26:42] public public information. We have
[00:26:43] customers in Europe where we’ve got
[00:26:45] localized LLMs deployed in their private
[00:26:47] cloud for these reasons. So, so it’s
[00:26:49] interesting and I and I do see more
[00:26:50] movement towards this. One of the main
[00:26:52] pitches or one of the main ROI that’s
[00:26:56] that usually is around automation and
[00:26:58] agents is that they save money and time.
[00:27:01] So, from what you’ve actually seen
[00:27:03] deployed, where do you think is the real
[00:27:06] measurable return and where companies
[00:27:08] just waste money and it really doesn’t
[00:27:10] pay off much? I think it’s fair to say
[00:27:12] that for us we’ve seen huge improvements
[00:27:14] in in our in our technology division
[00:27:16] when it comes to the the software
[00:27:18] development life cycle. So that that is
[00:27:20] like a slam dunk you know we we’ve
[00:27:22] gotten significant benefits from that
[00:27:25] you know and and we’ll continue to you
[00:27:27] know improve you know we’ve been able to
[00:27:29] rewrite some systems we we’ve just done
[00:27:31] amazing work right in in that area. Then
[00:27:34] if you look at things like inside sales,
[00:27:36] you know, marketing, uh if you look at,
[00:27:38] you know, the way we respond to RFPs, if
[00:27:41] you look at our security function, if
[00:27:43] you look at our operations functions,
[00:27:45] each of these business units, you know,
[00:27:47] who run these divisions, you know,
[00:27:48] they’ve got a set of core repetitive use
[00:27:50] cases that they’ve identified and that
[00:27:52] we’ve almost completely automated.
[00:27:54] They’re in the process of automating all
[00:27:55] of that. And I have no doubt in my mind
[00:27:57] that we will continue to make progress
[00:28:00] on each of these fields. you know, we so
[00:28:02] we’ve already done a ton of it. You
[00:28:03] know, we’ve got a lot of use cases that
[00:28:05] we’ve now go and gone and built skills
[00:28:07] and workflows for and that are on
[00:28:09] production. They’re going on every day.
[00:28:10] So, I think our org is is massively
[00:28:13] benefiting, you know, from the use of AI
[00:28:14] and and we we are very business case
[00:28:17] focused. So, so the question that you’re
[00:28:19] really asking in many ways is how do you
[00:28:21] determine the business case? How do you
[00:28:23] measure the business case? How do you
[00:28:24] measure the effectiveness of AI enabling
[00:28:27] something? And so, we’ve got the hooks
[00:28:29] to do that today. You know, we we
[00:28:31] already measured human productivity.
[00:28:32] We’re now measuring agent productivity.
[00:28:35] We’re measuring how much work a human
[00:28:37] does. We’re measuring how much work the
[00:28:39] AI agent does. And the agent works 24
[00:28:41] by7. It doesn’t take vacation. So I
[00:28:43] think there are a lot of things that
[00:28:45] still need to be done from an agent
[00:28:47] perspective that are completed for a
[00:28:49] human. So for example, you know, you’ve
[00:28:50] got an HCM system today in most
[00:28:52] enterprises that measures the life cycle
[00:28:54] of an employee. employee is onboarded,
[00:28:56] an employee is like assigned roles, an
[00:28:58] employee is measured on productivity,
[00:29:00] you know, and and and it goes on. So the
[00:29:02] same thing we’re now putting in place
[00:29:04] for an agent, you know, you have an
[00:29:06] agent life cycle management system,
[00:29:08] right? You we’re going to have this
[00:29:09] hybrid workforce of humans and agents
[00:29:11] working together in tandem to deliver
[00:29:13] value and you have to continuously
[00:29:15] measure that value. So the way similar
[00:29:18] way we onboard a human, we measure his
[00:29:20] performance or his performance. We
[00:29:22] figure out what are the good things, bad
[00:29:24] things, what needs to be improved. We
[00:29:26] follow the same cycle for the agent as
[00:29:28] well. I think coding software
[00:29:30] development that’s like the biggest use
[00:29:31] case especially since we are in the tech
[00:29:33] world, right? So we constantly see it.
[00:29:35] So we have this inflow of information as
[00:29:37] well that what’s really going on and
[00:29:39] we’re able to experience it. But let’s
[00:29:40] say if we talk about some organization
[00:29:43] or some business like an insurance or I
[00:29:45] don’t know healthcare, have you seen
[00:29:47] some use cases there? You’ve been able
[00:29:49] to see a lot of real measurable return.
[00:29:51] I think we we have several uh use cases
[00:29:54] like that. So so you know if you look at
[00:29:55] look at an insurance company for example
[00:29:57] you know they’ve got claims that they
[00:29:59] need to manage you know and these claims
[00:30:01] are structured and unstructured data. So
[00:30:03] a traditional forte for us has been AI
[00:30:06] that converts unstructured to structured
[00:30:08] information. Now, we apply it
[00:30:09] differently in different vertical
[00:30:11] industries that we operate in and in
[00:30:13] different regulated industries that we
[00:30:15] operate in, but the core premise is
[00:30:17] still that you get a ton of unstructured
[00:30:19] and semistructured information in and
[00:30:21] you have to be able to handle that at
[00:30:23] scale, you know, and so if you’re an
[00:30:24] insurance company, claims processing is
[00:30:26] is an easy one. You know, you’re able
[00:30:27] to, you know, to process claims and and
[00:30:29] you have structured input there too.
[00:30:31] Many insurance companies have have given
[00:30:33] you given you an app or a web interface
[00:30:34] where you upload what you need but then
[00:30:36] understanding if I if I upload a picture
[00:30:38] of a damaged automobile or if I you know
[00:30:40] upload some complex information about a
[00:30:42] claim a health insurance claim I still
[00:30:44] have to understand it make sense of it I
[00:30:46] have to convert that into something
[00:30:48] that’s billable in the healthcare case
[00:30:49] it has to be converted directly from
[00:30:51] this unstructured format using a lot of
[00:30:53] processing rules it has to be converted
[00:30:55] into a specific EDI format payer the
[00:30:57] insurance company understands when the
[00:30:59] claim comes in from the provider and
[00:31:00] therefore the pay the provider gets paid
[00:31:02] for it. So I think this this health
[00:31:04] insurance space obviously is one that we
[00:31:06] we spend a lot of time in and that we
[00:31:08] handle claims for we do adjudication for
[00:31:11] we do coding for these claims. It’s it’s
[00:31:13] a key area for us but but you have this
[00:31:15] type of use case validity in almost
[00:31:17] every type of insurance. Coming back to
[00:31:20] what we were talking about software AI
[00:31:22] writing you know a lot of software and
[00:31:24] agents kind of spinning up custom
[00:31:26] software and demand very easily for
[00:31:28] professional services and generally how
[00:31:30] do you think what does mean from a
[00:31:32] economic perspective of how enterprise
[00:31:34] software not get spilled rightly used to
[00:31:36] sell bodies or hours or something on
[00:31:38] those lines. Now this entire shift has
[00:31:41] happened. How do you think economic
[00:31:43] perspective of things has has impacted?
[00:31:45] >> Yeah, absolutely. I think there are two
[00:31:46] or three you know questions here. I I’ll
[00:31:49] I’ll try to answer them. I I think the
[00:31:50] first one is if you’re an enterprise and
[00:31:52] you’re buying SAS systems, right? And
[00:31:53] and this is the case with our team. You
[00:31:55] go look at the system and it’s some
[00:31:58] system for example that manages and
[00:32:00] validates firewall rules. You’re paying
[00:32:02] $100,000 subscription a year for the
[00:32:04] system. Algo sec for example, it’s it’s
[00:32:07] handling and validating all your
[00:32:08] firewall rules. Now you could on paper
[00:32:10] if you had the domain expertise build
[00:32:12] that software yourself. Now what’s the
[00:32:14] hardest problem associated with building
[00:32:17] your own system or application? You have
[00:32:19] to life cycle manage it. So the cost of
[00:32:21] maintaining patching securing a
[00:32:24] customuilt application is 3 4x the cost
[00:32:28] of building that application. So you
[00:32:30] build the application but then you have
[00:32:31] to patch it, you have to upgrade it, you
[00:32:32] have to secure it, you have to manage
[00:32:34] it, you have to scale it and you do this
[00:32:35] continuously. there are bugs that come,
[00:32:37] you know, from any open source component
[00:32:39] that you’re using. So, a lot of work
[00:32:41] associated with maintaining and doing
[00:32:42] application life cycle management, the
[00:32:44] costs are much higher. So, but that’s
[00:32:46] what AI solves. AI helps you automate
[00:32:49] that piece. So, you’ve got a class of
[00:32:52] applications in my view that enterprises
[00:32:54] are just going to rebuild themselves.
[00:32:55] It’s an inescapable conclusion. They’re
[00:32:57] going to do this with their tech teams
[00:32:59] inhouse or with their consulting
[00:33:01] partners. They’re going to do that. The
[00:33:03] second is for the applications that they
[00:33:04] do use, they’re going to be a class of
[00:33:06] applications that they have to maintain
[00:33:07] subscriptions for. What they pay per
[00:33:09] subscription and the economics behind
[00:33:10] that are going to change. Instead of
[00:33:12] just paying per employee per month,
[00:33:14] you’re going to pay as you said earlier
[00:33:16] based on outcome. You know, you’re going
[00:33:17] to you’re actually going to pay for
[00:33:18] outcomes. And we XPP, we only sell
[00:33:22] outcome. While we sell software as well,
[00:33:24] we’re mostly paid based on outcomes that
[00:33:25] we deliver. And so, you know, we’re
[00:33:27] actually in a good spot, you know, in a
[00:33:29] better spot than most software companies
[00:33:31] as far as I’m concerned. Our molt is
[00:33:33] still selling outcomes, right? So, I
[00:33:35] think the companies that sell outcomes
[00:33:36] versus the companies that sell licenses,
[00:33:39] that’s the big dichotomy. You know, the
[00:33:40] companies that sell licenses are the
[00:33:41] ones in trouble. You know, it may not
[00:33:43] manifest itself this year, but it’s
[00:33:45] happening. It’s just happening. It’s
[00:33:46] irreversible, right? And people are
[00:33:48] calling calling it by different names.
[00:33:49] Some people call it the SAS apocalypse.
[00:33:51] You know, they’ve got they’ve got names
[00:33:52] like that, but it’s it’s just an
[00:33:54] irreversible trend. Where is the
[00:33:55] opportunity if you’re a consulting
[00:33:57] company? I really think it’s around
[00:33:59] these FDE teams. I think you know if you
[00:34:01] have people that are able to AI enable
[00:34:04] high value use cases on behalf of
[00:34:07] customers and be able to provide a
[00:34:09] governance framework for AI within their
[00:34:10] organization. I think that’s the
[00:34:12] near-term value. I think you’ve got a
[00:34:14] situation you now have the ability to
[00:34:18] very rapidly improve the productivity of
[00:34:21] your workforce. That’s what AI gives
[00:34:23] you. And if you’re a consultant who can
[00:34:25] accelerate that and continuously help
[00:34:27] with that, I think there’s huge value
[00:34:29] you know as opposed to the traditional
[00:34:31] consulting you know world I think if you
[00:34:33] have a team of FDE that you can very
[00:34:35] rapidly deploy you know and there are
[00:34:36] companies like Palanteer doing this at
[00:34:38] scale but I think that’s the concept I
[00:34:40] mean if you look at if you look at claw
[00:34:42] if you look at anthropic or you look at
[00:34:44] openai you know they’re all building
[00:34:46] these FD teams now the problem is that
[00:34:49] they’re all about enabling dhman in your
[00:34:50] enterprise and if you are an actual
[00:34:52] consultant you’re like well you I think
[00:34:53] you need to be using an ensemble of
[00:34:55] models. I think you need to be managing
[00:34:57] your token cost. I think you need to be
[00:34:59] frontier model agnostic. I think you
[00:35:01] need the right governance structure
[00:35:03] around this. I think you need to think
[00:35:05] through policy management for agents. I
[00:35:07] think you need to think about this whole
[00:35:08] agent life cycle management thing and
[00:35:10] manage that whole thing. And and if you
[00:35:11] if you have these this FD team, I think
[00:35:14] you can design a set of market offerings
[00:35:16] right around that because I think that’s
[00:35:18] where the practical value is. Couple of
[00:35:20] questions on AI taking over the job
[00:35:22] world. Which white collar jobs do you
[00:35:25] think are genuinely at risk and which
[00:35:27] ones just feel at risk but are just
[00:35:30] going anywhere?
[00:35:30] >> Yeah. Here’s what I’ve observed, you
[00:35:32] know, actual numbers. So I think if you
[00:35:34] look at our product teams, right, if you
[00:35:36] you’re on product management, you’re in
[00:35:37] engineering, you’re on DevOps, you’re on
[00:35:39] QA, you know, the the trend across all
[00:35:42] of these is that the senior people’s
[00:35:44] value has increased. So if you’re a
[00:35:46] senior developer, you’re an engineering
[00:35:47] leader, you know, or you’re you’re a
[00:35:49] technical lead, right? You have IC’s
[00:35:51] reporting leads, leads, a team of leads
[00:35:54] report to an engineering manager, you
[00:35:55] know, that’s kind of the structure that
[00:35:56] we have. So you know, I I think the
[00:35:59] senior people is where all the values at
[00:36:02] today, you know, because they’re able to
[00:36:04] they already know the codebase. They
[00:36:05] understand what needs to be done and
[00:36:07] they’re the best at therefore at
[00:36:08] prompting and instructing the AI as to
[00:36:10] exactly what they want and how they want
[00:36:12] it done. I’m seeing this play out in a
[00:36:14] very fascinating way. the value of our,
[00:36:16] you know, our mature and seasoned people
[00:36:17] has gone up a lot. If you’re an entry-
[00:36:19] level person in any field, you are
[00:36:21] learning to use AI to become better,
[00:36:23] faster at your job and dramatically more
[00:36:25] productive. I view AI ultimately as
[00:36:29] something that will increase jobs, will
[00:36:30] result in the creation of business
[00:36:32] models that we still don’t know or
[00:36:34] understand. That results in net new
[00:36:36] opportunity, willing to take risks,
[00:36:38] willing to learn, willing to understand
[00:36:40] the AI. So I think I think the number
[00:36:42] one skill that I always tell people is
[00:36:44] I’m like look learn how to prompt become
[00:36:45] a master be an expert at your domain
[00:36:47] whatever your domain is. So if you’re an
[00:36:49] expert at your domain, you know, and
[00:36:51] you’re an expert, you you build the
[00:36:52] prompting skills you need, you’re king,
[00:36:54] you know, you you you get to write, you
[00:36:56] get to do really well at an amazing
[00:36:57] career. If you’re a startup founder, I
[00:37:00] think AI has unleashed a spectrum of
[00:37:02] possibilities that didn’t even exist
[00:37:04] because you now with your idea, you can
[00:37:06] be a oneman startup and actually go way
[00:37:09] past MVP, raise your series, A, do all
[00:37:11] of that. So I think I think we live in a
[00:37:13] remarkable world where AI is I think
[00:37:16] going to allow you to be more productive
[00:37:17] than you’ve ever been before. Are there
[00:37:19] going to be you know impacts on the
[00:37:20] economy? Are people going to lose jobs
[00:37:22] and you know are some jobs going to
[00:37:23] morph into other jobs and all of those
[00:37:25] things will happen as part of any you
[00:37:28] know productivity software that we put
[00:37:30] out anywhere right and AI is the
[00:37:32] ultimate productivity software so it’s
[00:37:34] going to have some impacts but I think
[00:37:36] you have to measure things in five and
[00:37:38] 10ear cycles you know so I think when
[00:37:40] you look at this cycle when when we
[00:37:42] emerge out of this we’re going to see
[00:37:44] massive productivity improvements we’re
[00:37:46] going to see massive you know economic
[00:37:49] progress and AI will be by far you know
[00:37:52] a net plus for for the economies of the
[00:37:55] entire planet. Yeah, I think u
[00:37:57] absolutely. So this is the first there
[00:37:59] are two things that have changed right
[00:38:01] uh from previous how do you put it
[00:38:03] previous innovations like internet
[00:38:05] socials and if you go back like 30 years
[00:38:07] or so but I think there are two main
[00:38:09] things that have changed one is that
[00:38:11] it’s the first time that any software or
[00:38:14] technology has really come for the
[00:38:16] knowledge worker previously that was not
[00:38:18] the case right so it was more on mundane
[00:38:20] task with everything else so it’s the
[00:38:21] first time that a knowledge worker kind
[00:38:23] of feels threatened that hey what’s
[00:38:25] really going on I can now do basic ling
[00:38:27] ing contracts myself. I don’t need to
[00:38:30] really talk to a lawyer. I can just, you
[00:38:32] know, plot it out and and figure it out.
[00:38:33] So, so those kind of threat and the
[00:38:34] second is the rate of change I think is
[00:38:37] much faster, right? So, you can’t really
[00:38:40] think 10 years now. You you have to
[00:38:41] think in 2 years, 3 years, 5 years, the
[00:38:44] rate of change or rate of adoption
[00:38:46] itself is increasingly
[00:38:49] is very has become very fast. And while
[00:38:51] I I strongly agree that you know senior
[00:38:54] engineers, senior people, they have that
[00:38:56] knowledge, they have that context, they
[00:38:58] have everything, but they’re also also
[00:39:00] kind of averse to there there’s a trust
[00:39:03] deficit in adopting AI which younger
[00:39:06] generation will just right so they’ll
[00:39:08] just do it. I mean I noticed my daughter
[00:39:10] using Snapchat or Insta like I don’t
[00:39:12] know I can’t I can’t imagine myself
[00:39:14] using it the way she is able to do
[00:39:16] though right. So I mean just the swipes,
[00:39:18] the gestures, the everything and I’m not
[00:39:20] that old. I’m still I mean I had Insta
[00:39:22] like 5 10 years back and this is fine
[00:39:24] but still the way she is able to adopt
[00:39:27] there’s going to be a major difference
[00:39:28] between seniors and unless what let’s
[00:39:30] say I’d say to my team as well that you
[00:39:32] have to unlearn the way you’ve been
[00:39:34] doing things right so you really have to
[00:39:36] forget the way you have done your work
[00:39:39] so far and really go on to and try this
[00:39:42] new work. Um and it’s not that easy. Uh
[00:39:45] but yes, time will tell and then you
[00:39:48] know I think uh I also agree that this
[00:39:50] will open up doors that we don’t know
[00:39:52] yet and that’s where there’s a lot of
[00:39:55] opportunity that can come.
[00:39:56] >> All right. Absolutely. Absolutely.
[00:39:58] >> Closing this out. Um sorry do two
[00:40:00] questions. What do you think is the next
[00:40:02] wave that most CTOs, IT directors,
[00:40:06] everybody and I’m talking from an
[00:40:08] enterprises, business perspective, not
[00:40:10] from a services for company perspective
[00:40:12] are not really maybe thinking about and
[00:40:14] and they should be thinking about it. So
[00:40:16] that’s one.
[00:40:17] >> I think the biggest thing that’s going
[00:40:20] to happen is and we started the
[00:40:21] conversation with this is becoming AI
[00:40:24] native and that that means that doesn’t
[00:40:26] what does it not mean? It doesn’t mean
[00:40:28] that you put a conversational AI or a
[00:40:30] voice agent on top of your user
[00:40:32] interface to be an alternative to your
[00:40:34] user interface. It doesn’t mean that.
[00:40:35] That’s good. That may help you with
[00:40:37] productivity and you know helping with
[00:40:38] customer service and may improve the
[00:40:40] quality of the human human interaction
[00:40:42] or human interface all of that. But I
[00:40:44] think it’s the guts of these software
[00:40:45] systems that enterprises rely on that’s
[00:40:48] going to change. It’s going to be the
[00:40:49] infusion of agi technology into the core
[00:40:52] of these systems. How you build a
[00:40:54] software system has changed. I think how
[00:40:57] you think about what software can do for
[00:40:59] you has changed. Taking cognitive
[00:41:01] ability, taking the power of these LLMs
[00:41:03] and making them a core part of what the
[00:41:06] ERP actually does or what these
[00:41:07] enterprise systems actually do is going
[00:41:09] to be huge. It’s going to change the way
[00:41:12] we think about automation because these
[00:41:14] systems increasingly not going to, you
[00:41:17] know, request inputs from humans. Even
[00:41:19] today, you know, most software systems
[00:41:20] is humans initiating interactions. I
[00:41:23] think the big shift is going to be these
[00:41:25] AI agentic AI enabled systems that do
[00:41:27] things for you and they ask your
[00:41:29] permission and then they learn from that
[00:41:31] and then they perform actions and then
[00:41:32] they inform you and if there’s a human
[00:41:34] in the loop step that needs to be done
[00:41:36] by you like you have to approve a
[00:41:37] massive payment you approve it and it
[00:41:38] happens but the before and after work is
[00:41:41] handled end to end and you now have I
[00:41:44] think we’re on a path where you’re going
[00:41:45] to have more and more intelligence
[00:41:47] embedded in these systems that’s going
[00:41:50] to change the way that humans interact
[00:41:51] with these systems. these systems are
[00:41:53] going just going to do a lot more stuff
[00:41:54] and we will gradually move to a point
[00:41:56] where you’re just informed when things
[00:41:58] happen as a courtesy as the system is
[00:42:00] designed to do and these systems are
[00:42:01] doing everything themselves you know so
[00:42:03] I think I see that as as probably the
[00:42:06] big longerterm trend and I think I think
[00:42:08] the ERPs are to be rewritten for that I
[00:42:10] mean I don’t think this is what they’re
[00:42:12] doing currently they’re slapping a layer
[00:42:14] on top they’re not reimagining their
[00:42:16] system and I think I think that’s the
[00:42:17] big change that needs to happen whether
[00:42:20] it’s Salesforce or whether it’s you know
[00:42:21] service now or whether it’s SAP the
[00:42:24] systems have to be reimagined with agent
[00:42:26] AI at the core and I think that that
[00:42:27] requires in many ways a fundamental
[00:42:30] rewrite and I think you’ll see that play
[00:42:32] out because if these systems don’t do it
[00:42:33] customers will do it for themselves
[00:42:35] because the cost of doing it is
[00:42:36] plummeted I don’t need some massive
[00:42:37] engineering team to go rebuild some
[00:42:39] silly CRM system that I use I just don’t
[00:42:42] there ain’t all that in that right so
[00:42:43] you’re going to have agents that are a
[00:42:45] part of these systems you’re going to
[00:42:46] have agents that employees use that span
[00:42:49] multiple systems just like a human
[00:42:51] they’ll be logging in into different
[00:42:53] systems interacting across them. So this
[00:42:55] layer is already being built by the by
[00:42:57] the anthropics of the world. You know,
[00:42:58] everybody is doing it. Many of these
[00:43:00] enterprise systems need to be rewritten
[00:43:02] and redesigned. I think agents as a
[00:43:05] layer and as an operating system within
[00:43:06] most enterprises will be doing a lot of
[00:43:08] the heavy lifting. So I think that
[00:43:10] that’s where things are headed.
[00:43:11] >> So it’s kind of a same rewrite that
[00:43:13] mobile once pushed, right? So 10 years
[00:43:15] back when mobile enterprise mobility
[00:43:16] came in, they had to really rethink on
[00:43:18] how mobile application function. But
[00:43:20] this is like 10x of that. Now you really
[00:43:22] have to think how agents would operate
[00:43:23] it. And there’s no human
[00:43:25] >> absolutely
[00:43:26] >> initiating things versus human is
[00:43:28] approving things or reviewing things.
[00:43:30] >> Absolutely.
[00:43:31] >> So yeah, I agree that that entire
[00:43:33] process has to and then think over and
[00:43:35] done. All right. So thank you sir taking
[00:43:38] out the time. Thank you. Thank you Ash.
[00:43:40] Great to chat with you man.
[00:43:41] >> Yes absolutely.