[00:00:04] So joining me today is Paul Lewis. He’s
[00:00:07] the CTO of Python. He spent almost 30
[00:00:09] years in IT and also served as a global
[00:00:12] CTO at Hitachi Ventara. He’s helped
[00:00:14] enterprises through every major tech
[00:00:16] shift and now he advises boards on
[00:00:18] technology investment and works at the
[00:00:20] CIO and CTO analyst at Gigaw. Our topic
[00:00:22] of discussion today is AI adoption
[00:00:25] versus AI advantage. And Paul, welcome
[00:00:27] to Techen. It’s really good to have you
[00:00:29] on the show.
[00:00:30] >> Glad to be here. Thanks for the invite.
[00:00:32] And it’s Python versus
[00:00:34] >> But there’s an interesting story, right?
[00:00:35] They
[00:00:36] >> It’s easy to read and say that’s it
[00:00:38] feels like Python, so therefore you
[00:00:40] pronounce it that way. But Python’s
[00:00:42] historyology is based on Pytha who is
[00:00:45] the high priestess to the Oracle of
[00:00:47] Deli. And we were an oracle services
[00:00:49] company. So it’s derived from Pytha and
[00:00:52] therefore we’re all Pythons.
[00:00:54] >> Ptheians. That’s a very wellthoughtout
[00:00:56] name.
[00:00:58] and somebody really spent some time
[00:01:00] there, right? So,
[00:01:02] >> exactly.
[00:01:02] >> Creating all the rigs and everything.
[00:01:04] So, this is good. That’s an interesting
[00:01:06] story. Starting off with Paul, you’ve
[00:01:08] you lived through all the
[00:01:10] transformations, technology, you know,
[00:01:11] the big technology changes like cloud,
[00:01:13] big data, social, everything, and now
[00:01:16] AI, right? So, what are your thoughts on
[00:01:18] that? Is it actually a different wave or
[00:01:20] it’s just that it feels that you know
[00:01:22] it’s just moving faster because
[00:01:23] previously technology used to change in
[00:01:25] decades versus now it’s like months. So
[00:01:28] there is a dichotomy right. I recently
[00:01:31] published an article called the enduring
[00:01:34] rules of IT and those enduring rules
[00:01:36] were true 35 years ago when I started
[00:01:39] running it in the banking environment
[00:01:41] and true today when AI is innovating.
[00:01:44] Realistically the rules of it like I
[00:01:47] have to build business cases to create
[00:01:49] some capital purchases. I have
[00:01:51] enterprise friction like who can go to
[00:01:53] production and why. I have real
[00:01:55] governance compliance and risk concerns
[00:01:57] that have to be true. Those haven’t gone
[00:01:59] away with APIs, with cloud, with
[00:02:01] blockchain, with AI. It all is in the
[00:02:04] same category of of what has changed
[00:02:06] with AI is our risk tolerance for
[00:02:10] implementing things that are less than
[00:02:13] 100%. Imagine a world 5 years ago where
[00:02:16] I would implement technology in
[00:02:18] production that had a 83% accuracy rate
[00:02:21] versus now you know out of the box
[00:02:23] models are somewhere in the 65 to 67%
[00:02:26] accurate and with prompting grounding
[00:02:28] you know access to internal data sets I
[00:02:31] can get that up to 85 to 90% but that’s
[00:02:33] still 10% off perfect. I’ve had
[00:02:35] conversations with CIOS of banks that
[00:02:38] say, I’m still good implementing these
[00:02:40] things in production as long as it
[00:02:43] doesn’t affect things like online
[00:02:45] banking, the ATM. People still need
[00:02:47] access to money. But if I want to
[00:02:49] interact with a customer and, you know,
[00:02:51] understand digitally the language and
[00:02:53] provide a digital answer to that world,
[00:02:55] then I’m perfectly willing to accept
[00:02:57] some potential problem as long as I’m
[00:02:59] not from a compliance perspective sort
[00:03:01] of focused on doing that. What’s
[00:03:03] interesting is you know a good chunk of
[00:03:04] my career we’ll call the first 15 to 17
[00:03:07] years has been very IT ccentric and
[00:03:09] highly regulated world the next decade
[00:03:11] itachi is a very human ccentric OT
[00:03:14] operational technology related work
[00:03:16] where per perfection is required right
[00:03:19] you don’t want an LLM managing the
[00:03:21] landing of a plane stopping of a train
[00:03:23] right 100% is the requirement very
[00:03:25] deterministic world I want this plane to
[00:03:27] land every time I ask this plane
[00:03:30] >> there’s no there’s a 90% landing it.
[00:03:32] Right. So,
[00:03:33] >> exactly. Exactly. And then when you look
[00:03:36] at, let’s say, autonomous vehicles,
[00:03:37] right, the Whimos of the world, that
[00:03:39] it’s a mix of both, right? You’re not
[00:03:41] going to have an LLM breaking the card,
[00:03:43] but you are going to have a whole bunch
[00:03:45] of LM determining appropriate routes and
[00:03:47] patterns on how to get, right? It’s an
[00:03:49] interesting mix of both. So, I think
[00:03:50] what you’re seeing is, especially AI
[00:03:53] adoption in the enterprise space, less
[00:03:55] risk intolerance and more mixing of
[00:03:58] determinative and non-determined sort of
[00:04:00] solution sets. And because technology in
[00:04:02] the AI were just changing so quickly,
[00:04:04] the one non-functional requirement that
[00:04:06] wasn’t the highest priority item in the
[00:04:08] past, which was replaceability as
[00:04:10] compared to availability, right? That
[00:04:12] becomes the number one thing is that if
[00:04:14] I put an AI use case in production, I
[00:04:16] need to make sure I can replace the LLM,
[00:04:18] I can replace the agentic workflow, I
[00:04:20] can replace the prompting, I can replace
[00:04:22] the grounding at any one point because
[00:04:24] the technologies change every single
[00:04:26] month. I got to deal with that change,
[00:04:28] right? So AI ops as a function different
[00:04:31] than IT operations becomes a skill set.
[00:04:33] That was a long answer but that
[00:04:34] >> no that absolutely I think uh the
[00:04:37] question deserved answer like this. So
[00:04:39] absolutely most companies that you talk
[00:04:41] to today or meet today do you think
[00:04:43] they’re adopting AI they or they just
[00:04:46] buy it and it’s just another line item.
[00:04:48] Then
[00:04:48] >> the one consistent truism of engagements
[00:04:51] with let’s say 100 CIOS every single
[00:04:53] quarter it’s inconsistent. It is very
[00:04:56] rare. I I can’t come to you now and say
[00:04:58] most customers are at maturity to five.
[00:05:00] I can come to you now and say the vast
[00:05:02] majority of customers are all over the
[00:05:04] spectrum. I see many many who you know
[00:05:07] have implemented AI in production but
[00:05:10] mostly for SDLC right they’re they have
[00:05:12] dark factory implementations because
[00:05:14] they’re a product company great but they
[00:05:15] have very limited deployments across the
[00:05:18] rest of the organization for personal
[00:05:20] task. Then I’ll go to other customers
[00:05:22] heavily invested in the Gleans, the
[00:05:24] Geminis, the co-pilots and have
[00:05:27] empowered 10,000 people with that tool
[00:05:29] but have built nothing in terms of the
[00:05:31] high ROI AI. Then I see lots of
[00:05:34] customers that simply have a
[00:05:35] prerequisite. They can’t do AI because
[00:05:38] their IT ORC chart looks like 1992. They
[00:05:41] don’t have they have a data state that
[00:05:43] crosses 15 different platforms or they
[00:05:46] don’t have a data platform to
[00:05:48] consolidate data that already exists,
[00:05:49] right? they just have a bunch of
[00:05:51] prerequisites that are preventing them
[00:05:53] from doing any AI innovation.
[00:05:55] >> And a lot of times I think it’s really
[00:05:57] what I feel it’s it’s unlearning what
[00:05:59] you know and then learning it all over
[00:06:02] again. I think that’s the biggest push
[00:06:04] that that this technology change has
[00:06:05] really brought right there’s a certain
[00:06:07] mindset with which you have been working
[00:06:09] for the last 15 20 10 odd years even I
[00:06:11] face that right so and then you have to
[00:06:14] kind of unlearn that and then start
[00:06:16] using it. So buying I guess is the
[00:06:18] easiest of things that you can do. Just
[00:06:20] say hey get 10,000 cloud account set up
[00:06:22] there and let’s give it to everybody.
[00:06:23] Then the harder step that is is a
[00:06:26] equipping or training the organization
[00:06:29] or the workforce to think AI first. So
[00:06:31] it’s like more when it was you know
[00:06:33] cloud you had this cloud native
[00:06:35] development and every developer every
[00:06:37] engineer really had to change and wire
[00:06:39] re you know kind of rewire themselves to
[00:06:41] thinking cloud first and then I go to
[00:06:43] other solutions. Now it’s like AI first
[00:06:45] and then I figure out what I have to do
[00:06:47] and how to leverage it. So
[00:06:48] >> it’s a bit of human incentive problem
[00:06:50] which is why AI for the most part sort
[00:06:53] of the roadblock the speed bump has been
[00:06:55] change management right most part
[00:06:57] especially users but if you if you don’t
[00:06:59] think about AI as displacing a job but
[00:07:01] enhancing in fact we tend to talk about
[00:07:03] use cases in terms of three big
[00:07:05] categories what would be tasks so things
[00:07:07] I need to AI enable myself my day-to-day
[00:07:09] work AI workflows which are think of
[00:07:12] those as departmentals things that
[00:07:14] happen within the department that I need
[00:07:15] to automate in multiple ways and then
[00:07:17] finally AI journeys, things that are
[00:07:19] interdep department, things like order
[00:07:21] to cash, hire to retire, like the large
[00:07:24] complex things which one could easily
[00:07:26] argue is just a collection of workflows.
[00:07:28] So we can provide a tool and make it
[00:07:30] available for tasks, right? So you know
[00:07:32] it’s 10,000 people. Let me give you the
[00:07:34] Geminis, the co-pilots, the gleans of
[00:07:36] the world and make your job personally
[00:07:39] better and I can provide them a tool set
[00:07:41] to say, let me make sure that what’s
[00:07:43] automatable within the department is
[00:07:45] automatable via aic model. And then
[00:07:48] finally, let me help you document what
[00:07:50] those endto-end journeys look like and
[00:07:51] then figure out what workflows I should
[00:07:53] automate in that world to make that a
[00:07:56] higher ROI as distinct projects versus
[00:07:59] individual projects. And that way the
[00:08:00] individual doesn’t see it as a displace.
[00:08:02] They see it as an empowerment to say
[00:08:04] you’re not going to teach me to replace
[00:08:06] the procurement but you are going to
[00:08:08] teach the procurement administrator to
[00:08:09] do things
[00:08:10] >> and do them faster. So you know exactly
[00:08:13] >> deploy it in a sales team. You don’t
[00:08:14] need to update the CRM. You don’t need
[00:08:16] to do this. You don’t need to go back to
[00:08:18] a recording take notes that all of that
[00:08:20] gets you know kind of automated. You you
[00:08:22] focus on doing sales. So you know you
[00:08:24] get get an hour more to do more calls
[00:08:27] and you know just potentially talk to
[00:08:30] potential customers. So that’s that’s
[00:08:31] the idea right. So save time on those
[00:08:33] things. I think it also varies
[00:08:34] department to department. So sales could
[00:08:36] be a very easy use case to identify how
[00:08:38] it can you know a 10x effect can create
[00:08:40] it but maybe accounting and finance
[00:08:42] isn’t right. though it’s really a time
[00:08:45] >> and the lesson learned especially over
[00:08:47] the last year maybe 18 months was there
[00:08:50] was this assumption expectation
[00:08:51] especially from the technology providers
[00:08:53] that one can DIY AI in that the
[00:08:57] procurement administrator the sales
[00:08:59] operations person will build their own
[00:09:01] no code agents and automate the
[00:09:03] department when in fact all the other
[00:09:05] low code no code technologies that
[00:09:07] existed for the last 20 years this isn’t
[00:09:09] any more effective than that rightly
[00:09:12] all of those People have jobs to do and
[00:09:14] they don’t want to learn technology to
[00:09:16] displace their task. They want to manage
[00:09:18] payroll, right? They they want to on
[00:09:20] board and recruit people, right? Like
[00:09:22] they have actual disciplines and those
[00:09:23] actual disciplines require human skill.
[00:09:25] That’s interesting news. So what we have
[00:09:27] learned is instead of diying within the
[00:09:30] obvious non-technology teams within an
[00:09:32] organization is why don’t you create a
[00:09:35] mechanism a centralized means to build
[00:09:38] those noode agents on their behalf and
[00:09:40] then train them on how to use those
[00:09:42] agents. That’s the win. Don’t ask them
[00:09:44] to build them. Let’s build them on their
[00:09:45] behalf so we know it works. Great. And
[00:09:47] there’s reusability also great. And then
[00:09:49] enable them to actually start performing
[00:09:51] and using because if you don’t do that
[00:09:53] and you expect people to build their
[00:09:54] own, they will try it’ll get frustrated
[00:09:57] and it’ll have a negative effect on the
[00:09:58] actual usability. But even worse is that
[00:10:01] you’ll have a whole bunch of smart
[00:10:03] individuals who do in fact build a whole
[00:10:05] bunch of stuff and now they’re not
[00:10:07] displaceable. Imagine a world in 5 years
[00:10:09] from now where a couple of really smart
[00:10:11] people, the most important people in
[00:10:12] finance, have built 120 personal agents
[00:10:16] for delivery of their T and then they
[00:10:18] win the lottery and then you have to
[00:10:19] hire a new person. Imagine the
[00:10:21] onboarding complexity of that new
[00:10:23] person. It’s almost impossible. So now
[00:10:25] you have a succession planning problem.
[00:10:27] You’ve made personal automation
[00:10:28] important, but that personal automation
[00:10:30] has created more uniqueness in that
[00:10:33] person that was already unique problem.
[00:10:36] Right? That that is a practical problem
[00:10:38] outside of AI.
[00:10:39] >> So I was talking to another to somebody
[00:10:41] else and and he suggested that we treat
[00:10:44] agents as employees. You know when we
[00:10:46] create agents we onboard them in a phys
[00:10:49] in a similar fashion that you would
[00:10:51] onboard any employee. So kind of treat
[00:10:53] them in that way as well. So maybe that
[00:10:56] >> I don’t buy into that philosophically
[00:10:57] mostly because agent as employee has no
[00:11:01] responsibility on accountability right
[00:11:02] there. You can’t fire that person.
[00:11:04] >> You can turn them off though. I hope we
[00:11:06] can Sure. But if nobody’s accountable or
[00:11:09] responsible, then
[00:11:11] >> that’s a challenge. Yes. So somebody has
[00:11:13] to own but somebody then has to own that
[00:11:15] agent, right? So Right. So that part is
[00:11:17] there and I hope we can still turn them
[00:11:19] off in 2 years. Let’s see. So walk me
[00:11:22] through your your normal workday. How
[00:11:23] how are you using AI in your workflow?
[00:11:26] >> So we have several tools as you can
[00:11:27] imagine. We’re we’re kind of Gemini
[00:11:29] Enterprise Central because we’re we’re
[00:11:32] Google partner in many ways. So I
[00:11:33] definitely use Gemini Enterprise. Uh,
[00:11:34] but I am an AI first blank page. It is
[00:11:38] the least likely thing to happen is for
[00:11:40] me to open up a slide, a document, a
[00:11:41] sheet, blank page, and then start
[00:11:43] typing. What’s much more likely to
[00:11:45] happen is I’m opening up Gemini
[00:11:47] Enterprise Cloud, drafting my thoughts
[00:11:50] on a topic that I know I have to create
[00:11:52] content for, have the draft get created,
[00:11:55] have me force conversation with this
[00:11:57] with this uh UI to build a better
[00:12:00] version of the thing I was thinking
[00:12:01] about producing. double-checking,
[00:12:03] tripleeing, expanding the research,
[00:12:05] building out the content that I think is
[00:12:07] reflective of my needs, getting it to a
[00:12:09] point where I think it’s finalized, and
[00:12:11] then taking that output, putting it into
[00:12:13] a second AI tool and having that verify
[00:12:17] the exact same that is 100% of the time
[00:12:20] of the content that what I don’t use it
[00:12:21] for is the much smaller, right? Draft an
[00:12:23] email to do respond to this chat. These
[00:12:25] are things that I still think requires
[00:12:26] human experience. I think it requires
[00:12:28] sarcastic tone. I think it requires
[00:12:31] aphability, right? there is a human
[00:12:33] communication interface that while I
[00:12:35] could have Gemini enterprise say you
[00:12:37] know build a poll bot it’s still not
[00:12:39] going to look like and respond like Paul
[00:12:41] does in a highly sarcastic affable way
[00:12:43] right so so so that exist but my role
[00:12:46] CTO tends to be a lot of content
[00:12:48] creation either external content
[00:12:50] creation or internal and it’s AI
[00:12:52] exclusive until such I have some
[00:12:54] secondary AI value in content so that’s
[00:12:57] the good news is over time it learns my
[00:12:59] voice it learns my content it learns my
[00:13:01] point of do. In fact, I can ask it very
[00:13:03] frequently, what’s my point of view on
[00:13:04] X? And it because it sees all of the
[00:13:07] content we’ve ever co-produced together
[00:13:08] and and I use that purposely co-produced
[00:13:10] that it’s pretty 95% accurate in what I
[00:13:14] think. At times you feel it probably
[00:13:15] knows you more than what you know
[00:13:16] yourself because
[00:13:19] I mean we are limited by memory and you
[00:13:22] know we’re limited to we can’t store
[00:13:23] everything and it has somehow crazy how
[00:13:26] it works at times and how accurate it
[00:13:28] can be in one sentence and how off it
[00:13:30] can be in one sentence. So you know it’s
[00:13:32] it’s both at the same times and you’ll
[00:13:33] be I’m I’m usually surprised at how
[00:13:35] right it is and yet how wrong it can be
[00:13:37] at at certain situations. And that’s
[00:13:39] where I guess the human in the loop, the
[00:13:41] human in the nose, the human in the you
[00:13:43] know the process and everything and that
[00:13:45] comes in. It cannot have that context
[00:13:48] that you have and you have built it over
[00:13:52] you know 18 20 whatever your working
[00:13:54] experience and everything else. So I
[00:13:55] think yeah you’re 35 m 20 I mean I’m the
[00:13:58] same person’s context time and you know
[00:14:01] that’s that’s what has been developed.
[00:14:03] I think there’s there’s a little bit of
[00:14:05] value in that though, right? Like what I
[00:14:06] wouldn’t wanted to do is make
[00:14:08] assumptions outside of my perspective. I
[00:14:10] think what humans are still pretty bad
[00:14:12] at is describing their own person.
[00:14:13] Right? So So if I’m starting a content
[00:14:15] creation, I’m doing a a live dump of
[00:14:19] what I believe is true in my head,
[00:14:21] poorly documented, poorly articulated,
[00:14:23] poorly organized because that’s not how
[00:14:25] people think and work, right? They
[00:14:26] create random notes of stuff. Goal
[00:14:28] number one is organize my random notes
[00:14:31] of stuff. Goal number two is ask me
[00:14:33] about the gaps in the random list of
[00:14:36] things that I probably should also be
[00:14:37] thinking about on this topic. Ask me
[00:14:39] about that. I will give you my opinion
[00:14:41] on those gaps. So now at least you have
[00:14:43] an organized opinion set of this topic.
[00:14:46] Then draft document. You have to go
[00:14:48] through the pre-work. Right? Humans are
[00:14:49] notoriously bad at having an opinion out
[00:14:51] of the box. They’re very very good at
[00:14:52] revision and very very good at sort of
[00:14:55] repeating and applying their own point
[00:14:56] of view. They’re very very bad as a
[00:14:58] general rule is starting from
[00:15:00] >> I mean yeah they’re right actually I can
[00:15:02] so oftent times if I ask it to draft an
[00:15:05] email response I have written my own
[00:15:07] response and I would just pick up a few
[00:15:09] lines that I wasn’t really able to
[00:15:11] articulate that well and I would pick up
[00:15:13] those lines from his long email and I’ll
[00:15:15] just plug them in my email. Right. So
[00:15:17] now I using your word I’ve co-produced
[00:15:20] an email that is better than what I had
[00:15:22] started off with.
[00:15:23] >> Correct.
[00:15:23] >> Yeah. Right. and and you’ve you’ve
[00:15:26] discovered the exact value which is the
[00:15:28] gap in your thought process, right? To
[00:15:30] say
[00:15:31] >> correct,
[00:15:31] >> I know I want to say 10 things. I
[00:15:33] probably should say 12, but I’m not
[00:15:35] remembering the other two or it’s just
[00:15:37] not top of mind or
[00:15:38] >> human can only remember 10 things. So
[00:15:40] the 12 things was just hard to even
[00:15:42] articulate anyway. Right. So you’re
[00:15:44] using it for the two, you’re not using
[00:15:46] it for the 10.
[00:15:46] >> Yes. And at the end of the day, email
[00:15:48] still looks like as if it was written by
[00:15:50] a human. Right. So that’s another
[00:15:51] problem that
[00:15:52] >> by the way the same thing but by the way
[00:15:53] the same thing applies to code. We’re
[00:15:55] using a relatively simple sort of email
[00:15:57] taskbased approach. But if I look at
[00:15:59] building complex agentic workflows via
[00:16:01] cloud code or just full applications
[00:16:03] comparatively, it’s the same thing,
[00:16:05] right? A developer has it has their own
[00:16:08] experience and they’ve built their own
[00:16:09] software, but they absolutely have
[00:16:11] conceptbased gaps in application
[00:16:13] creation. Right? Realistically, anyone
[00:16:15] individual developer didn’t build in its
[00:16:18] entirety and entire SAS. They
[00:16:20] contributed to, right? of all the great
[00:16:22] contributed twos that’s 20% of anything
[00:16:24] that’s ever been produced and used and
[00:16:26] therefore that 80% gap of things I’ve
[00:16:28] never done before that’s where you know
[00:16:30] AI automations and code development are
[00:16:32] helping you fulfill on that gap to say
[00:16:35] if I look at other applications you
[00:16:37] haven’t thought about a bunch of other
[00:16:38] things here’s those bunch of other
[00:16:40] things and let me help you build that in
[00:16:42] the background while you’re building the
[00:16:44] things you already know how to build
[00:16:45] >> a few questions around data foundation
[00:16:47] and data in probably in 2013 you wrote a
[00:16:51] piece called you don’t have a big data
[00:16:52] problem is that you know you don’t have
[00:16:54] an AI problem the 2026 version of that
[00:16:57] >> pretty much right the the reality is
[00:16:58] both pieces effectively to say it’s not
[00:17:01] a tool problem it’s a organizational
[00:17:03] model problem right you like you you’re
[00:17:05] missing what we would generally refer to
[00:17:07] as an AI operating model so you’re
[00:17:08] missing the head of AI somebody who’s
[00:17:10] both accountable and responsible for
[00:17:12] making financial andor strategic
[00:17:14] tactical decisions of AI you’re missing
[00:17:16] the concept of so COE’s to both build no
[00:17:20] code agents enable and train and to sort
[00:17:23] of build agentic workflows outside of
[00:17:25] the application development team, you’re
[00:17:26] missing some some technology disciplines
[00:17:29] that you’re not used to having. And
[00:17:30] while you might have bought a tool,
[00:17:32] realistically your organization requires
[00:17:34] many tools. And the big impact as you
[00:17:36] talked about in terms of in the
[00:17:37] prerequisites is it’s quite possible all
[00:17:40] of your awesome ideas can’t be
[00:17:42] implemented because you have actual
[00:17:44] roadblocks and prerequisit overcome to
[00:17:46] support. The good news is there has been
[00:17:48] an evolution in how I would generally
[00:17:51] think about AI use case. In the past,
[00:17:53] one would think of a AI use cases as a
[00:17:55] super a single agent to answer all
[00:17:57] questions, which is kind of what the
[00:17:58] evolution of enterprise search used to
[00:18:00] look like, right? Just go in enterprise
[00:18:02] search, you wouldn’t see any agents, you
[00:18:04] just ask any question. Enterprise search
[00:18:06] tool would go to the sources to discover
[00:18:08] it. What we figured out was that
[00:18:10] enterprise AI from a enterprise search
[00:18:12] perspective couldn’t overcome a
[00:18:14] knowledge management, right? Every
[00:18:15] organization has multiple files of the
[00:18:18] same content existing in multiple
[00:18:19] places. So if I was a grocery store
[00:18:21] chain and I bought a bunch of other
[00:18:22] grocery stores, I might have a store
[00:18:24] list in 10 different places, only one of
[00:18:25] them’s right. So if I asked Enterprise
[00:18:27] Search, who’s the manager of store 1 2
[00:18:29] 3, nine times out of 10, it was wrong.
[00:18:30] Right? So it’s exacerbating a knowledge
[00:18:33] management problem. And if I haven’t
[00:18:34] fixed the knowledge management problem,
[00:18:36] this isn’t going to help. I a super
[00:18:39] agent is going to be the least effective
[00:18:40] approach, which is why the new design
[00:18:42] process, especially within enterprise
[00:18:44] search, but holistically also true, much
[00:18:46] more discreet agents that do much more
[00:18:47] discreet function. Right? So I’m going
[00:18:49] to call upon the HR onboarding agent to
[00:18:52] do onboarding employee. I’m not going to
[00:18:55] call the HR agent to that that can do
[00:18:57] all HR functions. Right? So the more
[00:18:59] discrete function, the more singular
[00:19:01] data it has access to, the more discrete
[00:19:04] actions it can take. That’s the win.
[00:19:06] Which means I have more agents rather
[00:19:09] than less agents. But those more agents
[00:19:11] do something very specific and therefore
[00:19:13] much more likely to be true honest. So
[00:19:15] that has that has changed the design
[00:19:17] philosophy just just of agents
[00:19:19] architecturally.
[00:19:19] >> Yes, absolutely. So how do you get a you
[00:19:21] know 30-year-old data architecture ready
[00:19:24] for agents without a multiple year
[00:19:27] overhaul first? What could be
[00:19:28] >> that’s the kind of news the good news
[00:19:30] what I just described is it changes the
[00:19:32] philosophy and how important it is to
[00:19:34] deal with your 30-year-old data state in
[00:19:36] a world where I don’t have a data lake a
[00:19:38] data platform or data warehouse you know
[00:19:40] a lakehouse and I don’t have all of that
[00:19:42] data consolidated so I can make either
[00:19:44] analytical or action based decisions I
[00:19:46] would need a super agent to deal with
[00:19:47] that but if I don’t have super agents
[00:19:49] and I build agentic workflows the
[00:19:51] closest to the source that I’m only
[00:19:53] worried about the currency of the source
[00:19:55] itself so I’m not looking at a data
[00:19:56] state of a thousand databases of which
[00:19:59] I’m running on average n minus 5
[00:20:01] version. I’m only worried about the
[00:20:03] actual source that I want to run this
[00:20:05] agentic workflow and whether it’s
[00:20:07] correct. So you would do the same with
[00:20:09] the with let’s say if the data
[00:20:10] architecture is modern you have lakes
[00:20:12] warehouse and everything you would still
[00:20:13] prefer you know dedicated agents for
[00:20:16] dedicated task versus
[00:20:17] >> correct the closest to the source
[00:20:19] >> the closest to the source you would yeah
[00:20:21] you would do that in either case I if
[00:20:23] even if it’s an old architecture setup
[00:20:24] or
[00:20:25] >> correct but people out of the box were
[00:20:27] saying you know why don’t I build all of
[00:20:29] my agentic workflows from the lakehouse
[00:20:31] data warehouse as a sort the irony of
[00:20:33] that is that in order to take action one
[00:20:35] needs context realistically a data
[00:20:37] warehouse lost a good chunk of its
[00:20:39] context. If I have multiple addresses
[00:20:41] for multiple sources and there are
[00:20:43] three, four, five field addresses in the
[00:20:45] data warehouse, it’s a it’s a three-
[00:20:46] field address. So therefore, I’ve
[00:20:48] already lost the context of the address
[00:20:50] through the sources. So if I had an
[00:20:52] agentic function to update the address,
[00:20:54] almost impossible as an example. So it’s
[00:20:56] actually the least valuable place to
[00:20:58] draw data from and and in action is a
[00:21:01] data warehouse. best possible place is
[00:21:03] the actual sources themselves. Which
[00:21:04] means the best possible place to deploy
[00:21:06] or build an agent isn’t outside in a
[00:21:10] technology framework. It’s actually
[00:21:11] inside the SAS platform that’s using the
[00:21:14] tool. Right? So in in fact the vast
[00:21:16] majority of agents one would use produce
[00:21:18] consume is inside the CRM inside the ERP
[00:21:21] inside the HLS inside the MEES because
[00:21:24] it’s the closest to the data. It’s the
[00:21:26] closest to the user. It deals with all
[00:21:28] of the access control and identity
[00:21:31] supported to do that and it’s of course
[00:21:33] context rate. All of those things are
[00:21:35] true and it’s for the most part included
[00:21:37] in the price in the SAS software you
[00:21:38] have done. The irony of course is that
[00:21:40] the SAS prices have increased
[00:21:42] dramatically 20 to 25%. Right? So so you
[00:21:44] know it’s included but you’re just
[00:21:46] paying more for the SAS.
[00:21:48] >> But that being said it’s clearly the
[00:21:49] best place to put it.
[00:21:50] >> I was reading up you also kind of work
[00:21:52] with uh Shoulick with their AI master
[00:21:56] programs. I hope I got the name right.
[00:21:58] >> School of business, which is the school
[00:21:59] of business of York University.
[00:22:01] >> Awesome. So, what’s the one skill most
[00:22:03] AI graduates are still missing? What do
[00:22:05] you think is or they should be something
[00:22:07] that’s something that they should be
[00:22:08] focusing on
[00:22:09] >> entirely? Most students are thinking
[00:22:11] about it as sort of a tool-based problem
[00:22:13] set and it’s an operating model based
[00:22:15] problem set. The vast majority of sort
[00:22:17] of the academic world don’t have a lot
[00:22:19] of practical experience. Uh so the
[00:22:20] practical experience being in IT and
[00:22:22] understanding enterprise friction is
[00:22:24] always the thing you have to it’s very
[00:22:25] easy to think of and I’m sure you’ve
[00:22:27] heard this term many times for deployed
[00:22:29] engineer. So each one of the technology
[00:22:30] companies will deploy a forward deployed
[00:22:32] engineer into your environment to help
[00:22:34] you build out those use cases. But
[00:22:35] what’s the thing they can’t do? Deploy
[00:22:37] to production. They can’t right they
[00:22:38] don’t have the credentials. They don’t
[00:22:39] have the skill set. They don’t have
[00:22:40] access to the systems to do that. And
[00:22:42] you probably shouldn’t let an external
[00:22:44] vendor deploy to your production
[00:22:46] environment. There’s there’s a whole
[00:22:47] bunch of enterprise frictions why that’s
[00:22:49] true. learning the enterprise friction
[00:22:51] is the biggest gap in the academic world
[00:22:53] before they apply
[00:22:54] >> that’s always been a industry academia
[00:22:56] gap across the world anywhere right so
[00:22:58] because the industry
[00:23:01] operates on a different level the
[00:23:02] academia never really gets that exposure
[00:23:04] of what really goes on you know at 1:00
[00:23:06] a.m. in the night production DB is down
[00:23:08] what do I do things like
[00:23:09] >> but I think it’s even more true with AI
[00:23:11] because AI started as a consumer tool
[00:23:14] set right I went into chat and I did
[00:23:17] cool things and I could apply that in my
[00:23:19] real world so why can’t I this just
[00:23:21] apply that literally in my job
[00:23:23] >> right and the reality is I can’t
[00:23:26] limitations
[00:23:26] >> I also believe FDE I think they need to
[00:23:30] really really have very strong product
[00:23:34] thinking product management or the kill
[00:23:36] set kind of approach as well because now
[00:23:39] you’re actually albeit it might be a
[00:23:41] small workflow it might be a small
[00:23:43] internal but at the end of the day it’s
[00:23:45] really determining you know
[00:23:46] understanding what the requirement is
[00:23:48] and then creating a product out of it so
[00:23:51] more than coding I because let’s say you
[00:23:53] you can push off coding to the agents to
[00:23:55] a certain extent but it’s more that you
[00:23:58] have to define the problem first so you
[00:24:00] know you have to state down the problem
[00:24:02] it’s not the technology that matters
[00:24:03] it’s not that but at a very basic level
[00:24:06] before the bigger systems kick in. You
[00:24:07] really have to understand the problem
[00:24:09] and being able to define it. So I guess
[00:24:10] that one skill is something that all
[00:24:13] engineers or developers would be
[00:24:15] required to do in you know the next
[00:24:17] >> I think rapid business understanding is
[00:24:20] incredibly helpful for an FD correct
[00:24:22] >> and even rapid I don’t want to say a
[00:24:24] word prototyping but if I were to say if
[00:24:26] I could demonstrate an applicability of
[00:24:28] this business problem this task or
[00:24:30] workflow or journey in a technology even
[00:24:33] if it’s not live that’s still helpful in
[00:24:35] the business problem definition process.
[00:24:37] So it’s not demo for the purpose of
[00:24:39] deploying to production, demo for the
[00:24:41] purpose of understanding the business.
[00:24:42] >> Correct. I mean I I often you know just
[00:24:44] tell my team at times as well that maybe
[00:24:46] waterfall is coming back in this sense
[00:24:49] because requirements has always been the
[00:24:52] major challenge right. So all of these
[00:24:54] models, waterfall, iterative
[00:24:56] development, agile, canban, all of these
[00:24:59] kind of stem from the fact that you know
[00:25:01] requirements, what do we need to build?
[00:25:03] How do we build? And and now an agent if
[00:25:06] it can code 36 hours consecutively, it
[00:25:08] can build a fairly decent app. You need
[00:25:10] to sell it to the requirement. So that
[00:25:12] entire waterfall methodology might just
[00:25:15] come back now because you have to% but
[00:25:17] in fairness having led multiple app dev
[00:25:20] teams in my career sometimes they were
[00:25:22] agile in name only they were sprints but
[00:25:24] they were small they were just shorter
[00:25:27] waterfall projects
[00:25:29] >> exactly I agree 100%.
[00:25:32] That’s yeah I mean we have been so much
[00:25:35] the client wants a fixed budget I don’t
[00:25:37] have more than spend so obviously the
[00:25:39] scope becomes fixed. Okay, great. So,
[00:25:41] we’re locking in the scope. Okay, great.
[00:25:42] Just break it down into sprints and call
[00:25:44] it a day. So, yeah, that’s uh and FDS
[00:25:46] would then be, you know, moving forward,
[00:25:48] you’ll be more like functional
[00:25:50] consultants that you have in ERPs,
[00:25:52] right? So, uh you’re deploying SAP,
[00:25:54] Oracle, Udu, Dynamics 365, whatever. And
[00:25:58] then there’s a functional consultant who
[00:25:59] understands how the uh you know supply
[00:26:02] chain module gets installed, how the
[00:26:03] finance system gets installed, how the
[00:26:05] those guys I think because then they
[00:26:07] have the domain knowledge knowing how
[00:26:09] the finance team would work in a
[00:26:11] manufacturing or or what is procurement
[00:26:13] looking at in a mold manufacturing
[00:26:15] organization or a textile manufacturing
[00:26:17] organization or whatever and then being
[00:26:19] able to understand those processes
[00:26:20] better. Oh,
[00:26:22] >> and just as importantly, being able to
[00:26:24] identify what’s not AI. Like, so very
[00:26:26] frequently, I’ll go a customer workshop
[00:26:28] and we’ll we’ll have a list of 400
[00:26:30] amazing. 350 of them are analytics.
[00:26:33] >> Correct.
[00:26:33] >> Right. It’s I need to list that dot dot
[00:26:35] dot. It’s like, okay, well, let let me
[00:26:37] fire this up for you. Of the remaining
[00:26:38] 50, 30 are kind of out of the box
[00:26:41] machine learning, right? Segmentation,
[00:26:42] classification, BI dashboards, right? Of
[00:26:44] of the last 20, uh, 15 of them are very
[00:26:48] complex, very expensive. They may or may
[00:26:51] not be able to afford this multi-quarter
[00:26:52] project, but let’s identify them. But
[00:26:54] then five are kind of the sweet spot
[00:26:55] ones, right? They hey, we could do this
[00:26:57] within a quarter. You already have the
[00:26:59] data available. We don’t have a lot of
[00:27:01] prerequisite speed bumps. I do foresee
[00:27:03] an RO like let’s do those five, right?
[00:27:06] And sometimes you need FTE to really
[00:27:08] narrow down that awesome automation
[00:27:10] request into a determinative or
[00:27:12] non-determinative answer
[00:27:15] >> or BI or dashboard. They’ve got to make
[00:27:18] that determination because end user. And
[00:27:20] even in in in workflow automation
[00:27:22] systems, you don’t need an agent every
[00:27:24] single time. At times, it’s just right.
[00:27:26] So you you’re getting documents from
[00:27:28] three different sources. You have to go
[00:27:30] through them, extract them, and then
[00:27:31] let’s put them somewhere. You don’t need
[00:27:32] an agent to do all of this. All you need
[00:27:34] you need a script to extract those
[00:27:36] documents, put it here. Then maybe an
[00:27:38] agent extracts information from them,
[00:27:40] you know, OCR or whatever, and then
[00:27:42] writing back is again an API call and
[00:27:44] just write it back. Right? So that
[00:27:45] >> there’s a lot of deterministic workflow
[00:27:47] in workflow. Yeah, that’s I mean I’m
[00:27:50] glad I I can share this with somebody
[00:27:52] else as well because these are really
[00:27:54] problems that we’re running into. We
[00:27:56] just need AI just apply AI and yeah is
[00:27:58] not even needed over there. It’s good
[00:28:00] old, you know, RPA or standard
[00:28:02] deterministic workflows.
[00:28:04] >> And I think that’s kind of the good
[00:28:06] news, right? I could walk you through
[00:28:07] hundreds of let’s say AI use cases we’ve
[00:28:09] built and I could tell you the the
[00:28:10] highest ROI ones use the most advanced
[00:28:14] AI tool set which is computer vision.
[00:28:17] Computer vision is clearly the one
[00:28:18] that’s escaped the gardener hype hypocha
[00:28:21] right it is in the it is an insightful
[00:28:23] tool that is mostly good if you have a
[00:28:26] bunch of digitized documents and you’re
[00:28:28] just pulling trying to pull out fields
[00:28:29] from it using computer vision that is
[00:28:31] 97% accurate 97% accurate is probably
[00:28:35] better than the people you have rekeying
[00:28:37] that document manually yes do that but
[00:28:40] even that agentic workflow is mostly
[00:28:43] deterministic with the exception of the
[00:28:45] computer vision part
[00:28:46] >> we’ve actually done that for several
[00:28:48] providers. They’re keying in thousands
[00:28:50] of documents and they have handwritten
[00:28:52] documents. They have very fine PDFs.
[00:28:54] They have very fine written you know the
[00:28:57] use case where it’s the most effective
[00:28:59] is going through those handwritten
[00:29:00] documents and extracting information
[00:29:01] from it. LM is good at it even LM or you
[00:29:04] know right so so yeah absolutely I agree
[00:29:07] rest of everything they’re just you know
[00:29:09] normal deterministic workflows that you
[00:29:11] can integrate and set up. Let’s talk
[00:29:12] about a little bit about governance and
[00:29:15] you know sharing data with AI security
[00:29:17] perspective. How would you kind of
[00:29:19] implement that in organization? What are
[00:29:21] the things to watch out for? So if you
[00:29:23] look at governance a little broader it
[00:29:25] is notorious for having a trailing
[00:29:28] function of governance versus a leading
[00:29:30] function of governance. Now AI has to
[00:29:32] have a leading function. You have to
[00:29:33] think about it in terms of how much
[00:29:35] money am I willing to spend? Who gets to
[00:29:37] make decisions of those decisions? who
[00:29:40] gets to determine what good looks like.
[00:29:42] Like what level of accuracy is valuable
[00:29:43] to me? Who has to deal with the bump in
[00:29:46] the night? But the bump in the night
[00:29:47] isn’t when the agent failed. The bump in
[00:29:49] the night was when the accuracy goes
[00:29:50] from 87%. How are you instrumenting
[00:29:53] that? And how are you observing that?
[00:29:55] And then if you’re observing it, how
[00:29:57] you’re actioning that change. It’s quite
[00:29:58] possible that 87 to 83 is based on the
[00:30:01] source. It’s based on the prompt. It’s
[00:30:02] based on the model. It could be based on
[00:30:04] so many things. But that 4% gap is a
[00:30:06] million dollar impact, right? because
[00:30:08] you’re making 4% worse responses. That’s
[00:30:10] a big number, right? Correct. So that
[00:30:11] governance needs to be in place both
[00:30:13] people and technology to instrument
[00:30:16] observe and action well before my first
[00:30:18] AI is deployed into production for
[00:30:20] workflows in journey for tasks pretty
[00:30:22] much out of the box of the tool sets
[00:30:23] that are at their disposal, right? They
[00:30:24] get to do that. Now I think it’s fair to
[00:30:26] say that the in production operating
[00:30:29] function set the most nent part of any
[00:30:31] of the tools available for you to right
[00:30:33] in fact when I look at all the tools
[00:30:35] sort of out out of the box right now
[00:30:36] only Gemini enterprise agentic framework
[00:30:39] has all of the things it needs like an
[00:30:41] agent registry agent observation agent
[00:30:44] logs um agent ids you know all the
[00:30:47] things you need to both instrument
[00:30:48] observe and action in production doesn’t
[00:30:51] even have a deployment environment right
[00:30:53] there there’s no means to deploy an
[00:30:55] anthropically built agent in an
[00:30:57] anthropic environment. So it’s already a
[00:30:59] big gap. So I now have to deploy that
[00:31:00] somewhere else and you have to hope that
[00:31:01] somewhere else has what you can’t do is
[00:31:03] treat it like an application, treat it
[00:31:05] like compute, treat it like everything
[00:31:07] else I do in even though it’s an IT
[00:31:09] asset and LM and aentic workflow um are
[00:31:12] all still IT assets I need to deal with.
[00:31:14] The IT team are unfamiliar with these IT
[00:31:18] which is why you have AI ops sort of
[00:31:19] function. So I guess the big answer here
[00:31:21] is you have to think about it before you
[00:31:23] deploy to production instead of
[00:31:25] notoriously AI thinks about it several
[00:31:27] years after I deploy to production
[00:31:29] because of the potential net impact your
[00:31:31] business and I’m not even talking about
[00:31:32] financial it just might reputation 4% of
[00:31:34] the time um it’s giving you the wrong
[00:31:37] answer that might have a detrimental
[00:31:39] thing to your to to your business as a
[00:31:41] whole right so I have to think of that
[00:31:42] as as a privacy concern and it has to
[00:31:44] coexist with your data governance like
[00:31:46] if you have AI products then that they
[00:31:49] have to use the data products that
[00:31:50] exist. Just like I have data steward, I
[00:31:52] better have AI agents. Uh just like I
[00:31:54] have data security categorization,
[00:31:58] classification, I need AI security
[00:32:00] categorization and classifying. There’s
[00:32:02] a parallel universe where I have to
[00:32:05] think about those as living beings as I
[00:32:08] do data products as I do.
[00:32:10] >> So closing question, you’ve walked
[00:32:11] companies adopt every major technology
[00:32:14] for over 30 years now. What’s the
[00:32:16] hardest lesson that has taught you you
[00:32:18] know about technology leadership or life
[00:32:20] in general?
[00:32:21] >> The hardest lesson has been the people
[00:32:24] centric impact of implementing it is
[00:32:26] very clear especially IT shops have been
[00:32:28] more concerned with the non-functional
[00:32:31] requirements of the application than
[00:32:32] they are the functional requirements. So
[00:32:34] they worried about uptime, availability,
[00:32:37] performance and not worried about impact
[00:32:40] to the individual if they choose to use
[00:32:42] the AI flips that AI is now much more
[00:32:46] impactful to the person to the human
[00:32:48] whatever tool set you. So now I have to
[00:32:50] rethink how it enablement works and it
[00:32:53] has to be a people centric enablement
[00:32:55] versus everything else in IT is a tool
[00:32:57] ccentric enable doesn’t work it won’t
[00:32:59] work definitely won’t work in IT and
[00:33:01] >> that’s true. So Paul, it’s it was a
[00:33:03] pleasure having you on Techan. I hope
[00:33:06] you enjoyed the conversation as much as
[00:33:08] I I did it as well. And I really
[00:33:10] appreciate your time and your cander and
[00:33:12] you know the open conversation we had.
[00:33:13] >> Thank you very much. Appreciate it.