[00:13] welcome to Google Cloud. Next,
[00:16] we are going to be talking about
[00:19] something that we are often asked which
[00:22] is impossible to do.
[00:25] Ship product faster.
[00:28] By the way, it has to be a
[00:30] differentiator that no one has done
[00:32] before.
[00:34] Ship it globally quickly.
[00:38] And you ask about headcount.
[00:41] That's that's a good question. We'll
[00:42] come back to that. So, how do we achieve
[00:44] that? Um, and I think the answer lies
[00:48] somewhere with the two-letter acronym AI
[00:51] and agents, which is what we are going
[00:53] to be exploring today.
[00:55] My name is Patik. I lead a cloud phops
[00:58] practice here at Google. Joining me here
[01:00] is Andre. He leads uh financial services
[01:03] here at Google. And the ultimate proof
[01:05] point we have brought is James who's the
[01:08] CIO of One United Bank and have uh been
[01:12] sharing experiences around how AI has
[01:14] been helping the bank internally,
[01:16] externally and for greater good.
[01:19] The way we will be spending our time
[01:20] today is looking at value realization
[01:23] framework. Right. So, how do we move
[01:25] from AI hype, right, just go at it to
[01:30] thinking about hard ROI numbers? This is
[01:34] something that the executives, the board
[01:36] is specifically asking many of our
[01:38] companies to do. Next up, we have James
[01:42] who will be sharing his experiences in
[01:44] terms of how they have been using AI for
[01:47] the past several years going back to
[01:50] 2020. the lessons learned, the ROI that
[01:53] the team have received not only to
[01:55] improve internal teams productivity but
[01:58] also making a differentiated customer
[02:00] experience externally and they are going
[02:03] to be also building something of a
[02:05] greater good which I will let James
[02:08] speak about. Um and then finally we will
[02:11] move on to a demo. For demo piece we
[02:14] actually picked a use case that we think
[02:16] would be resonating well with all of
[02:18] you. Uh this is about an army of AI
[02:22] agents helping you manage cloud spend at
[02:26] scale. So think about reporting,
[02:29] optimization,
[02:31] >> forecasting, allocation and so and so
[02:33] forth.
[02:34] >> All right. So with that said, Andre, how
[02:38] should we be thinking about value
[02:40] overall?
[02:41] >> Yeah. Uh and we'll start with uh the
[02:45] most expensive thing that customers are
[02:47] faced with right now. And it really
[02:49] boils down to looking through uh this
[02:52] chart that we have here that shows um
[02:56] you know what's been happening over the
[02:58] past few years. And as you can see in
[03:02] this chart, there's two races that are
[03:04] that's happening. First is the
[03:06] innovation race which is really driven
[03:08] by the boardroom pressures to uh invest
[03:12] billions of dollars into AI capabilities
[03:16] only to six months later to a year find
[03:18] out that the hard ROI and the savings
[03:22] that they were looking to receive and
[03:25] the business value created no longer
[03:27] exist. So and it wasn't that it was the
[03:30] business uh technology that failed. It
[03:34] was more so that no one asked the
[03:36] business value question before the first
[03:39] line of code was written.
[03:41] The second race here is that you can see
[03:43] is in the dark blue line. It's slow,
[03:46] steady. Uh and that's because every
[03:49] dollar that is spent is tied to business
[03:52] outcome.
[03:54] Uh and you know the harsh reality here
[03:57] especially for lean teams is you can't
[04:00] afford to run these science projects
[04:03] every dollar that is spent you have to
[04:06] answer the question where's the ROI so
[04:09] we wanted to enable you and create a
[04:11] framework uh that we can leave with you
[04:14] that really helps you guide you to
[04:17] participate in this value realization
[04:19] race. Next slide. And the way we think
[04:22] about it is across three phases. The
[04:24] first phase is identifying where the
[04:27] pain is the loudest. Right? So this is
[04:29] fixing the friction. Identifying all of
[04:32] the repetitive manual task that can be
[04:36] automated. Um and from a financial
[04:38] services perspective that is being able
[04:41] to open a bank account. There's loan
[04:44] processing. That's password resets.
[04:47] Right? all of that work that's high
[04:49] volume that's tedious uh but low
[04:51] complexity uh and that can be automated.
[04:55] The second is automation and that is
[04:58] where we're actually deploying agents in
[05:01] your across your enterprise ac across
[05:03] your environment and we use that as a
[05:06] way to buy back capacity. So how can we
[05:08] get time back to the teams and we do
[05:10] that through automation and the last
[05:13] phase here uh really to me I think
[05:17] allows us to really drive value and
[05:19] that's the transformation because now we
[05:22] freed up the time and the capacity from
[05:25] uh your resources and teams to now so
[05:28] they can invest in building products uh
[05:31] that's capable of uh competing against
[05:35] their competitors. So James let me bring
[05:37] you in uh tell me about one united bank.
[05:41] >> Sure.
[05:42] >> And tell me how this framework is
[05:44] applicable to your
[05:45] >> Awesome. Well, hey, thanks for having
[05:47] me, Andre. Uh, really appreciate the
[05:49] opportunity to tell everyone a little
[05:50] bit about who we are and what we've been
[05:51] able to accomplish. Uh, so, uh, One
[05:54] United Bank, uh, is America's leading
[05:56] community development financial
[05:58] institution. Uh, what does that mean? It
[06:00] really means that we're a missiondriven
[06:02] organization. Uh, we've been championing
[06:04] financial services and financial
[06:06] literacy for economic advancement for
[06:08] years. Um, you know, we're also a FDIC
[06:12] insured bank. Um, but we're a digital
[06:15] bank. So we operate a little bit more
[06:17] like a fintech and find ourselves
[06:18] competing against uh the likes of the
[06:20] neo banks that are out there. Uh we find
[06:23] that what we are doing differently is
[06:25] we're leveraging our technology to help
[06:27] us deliver unique products and services
[06:29] to our customers on that national basis.
[06:32] Um that said we're doing it with a
[06:34] pretty lean team. We got about 150
[06:36] employees internally uh that are helping
[06:39] us support that mission and objective.
[06:42] >> Awesome. So James, if I were to go back
[06:45] to what Andrew just mentioned, right,
[06:46] the playbook, which is the friction, the
[06:49] automation, and then ultimately the
[06:51] transformation. Yeah.
[06:54] >> Tell me more about the friction.
[06:56] >> Yeah. Like what let's let's go back
[06:59] 2020. What happened there?
[07:01] >> Yeah, sure. like going back to 2020 um
[07:04] you know I think if you uh think about
[07:06] where we were uh in time when United was
[07:09] well established as a digital bank and
[07:11] we were originating you know new
[07:13] customers on a monthly basis at a pretty
[07:15] steady clip and it was slowly improving
[07:17] and in increasing especially as we hit
[07:19] those periods of time where PPP was a
[07:21] thing and everyone was looking for you
[07:23] know digital services from their couch.
[07:26] Uh and then I think we all remember over
[07:28] the summer that summer uh there was the
[07:29] George Floyd incident and there was a uh
[07:32] outreach that uh was looking for
[07:34] organizations that were valueminded like
[07:36] One United.
[07:37] >> Uh what we saw was we saw that our
[07:39] customer base doubled in about a 60-day
[07:42] period.
[07:42] >> Um you know, and if we think about what
[07:44] that was in COVID, I mean, we were faced
[07:47] with the challenge of that didn't mean
[07:48] we were going to get more headcount. We
[07:50] were really honestly struggling with how
[07:52] are we socially distancing people in our
[07:54] call center? How can you bring in more
[07:56] people, you know, into, you know, an
[07:58] office and train them up and get them up
[08:00] to speed to be able to support financial
[08:02] services? I mean, that's a really
[08:03] difficult task on a regular day.
[08:06] >> Yeah. And, you know, you think about it,
[08:08] right? You had this one challenge on one
[08:10] hand, but then you had a great
[08:12] opportunity.
[08:13] >> Yeah.
[08:13] >> Right. So, talk to us about what that
[08:16] looks like and why you can't just throw
[08:17] people at it, right?
[08:18] >> Yeah. Sure. So I think that you know if
[08:21] you go back to that moment in time what
[08:23] what happened was we had this great
[08:25] phenomenon of a of a whole new customer
[08:27] base coming in and you know that meant
[08:30] that while our origination systems had
[08:32] done a great job of getting everyone in
[08:34] we were now faced with the inevitable
[08:36] questions and struggles new customers
[08:38] have and they didn't have the
[08:39] opportunity to go into branches to
[08:41] satisfy them. So what we were faced with
[08:43] was a flood of uh new customer inquiries
[08:46] uh for support coming through multiple
[08:48] channels. So when you talk about the
[08:49] friction, you know, people people will
[08:52] certainly come and reach out to you
[08:53] through email and they'll reach out to
[08:55] you through your website, but people
[08:57] when it comes to their money, they end
[08:58] up calling, right? They end up on the
[09:01] phones. So when we looked around at
[09:02] where our friction was, we were seeing
[09:04] that our biggest problem was in being
[09:05] able to service those customers who are
[09:07] coming in through the voice channel. We
[09:09] had a legacy phone system that had that
[09:11] crazy phone tree. You know, it was one
[09:13] of the places that, you know, we had
[09:14] done a lot of work in the phone tree and
[09:15] we thought, "Oh my god, we've built this
[09:17] great phone tree." Well, people coming
[09:19] in have complicated questions and those
[09:22] complicated questions are hard to find
[09:23] answers to in a seven layer phone tree.
[09:25] Um, so what we did was we made our first
[09:28] kind of pivot uh to looking to agents to
[09:31] help us. Right? This was early though uh
[09:33] in agents. We were looking at dialogue
[09:35] flow at the time.
[09:36] >> But what we moved to was we moved to a
[09:38] natural language processing approach. We
[09:41] as we simply asked the question of our
[09:42] customers, how can we help you? And
[09:44] instead what that did was it allowed
[09:46] them to say, "Hey, I want to call center
[09:48] agent." Of course they're going to say
[09:49] agent. That's the first thing they say.
[09:50] So we were able to prompt them to give
[09:52] us a little bit more description. And
[09:54] what it resulted in was intentbased
[09:56] routing. It took away the complexity of
[09:59] how do I find the right place in the
[10:00] bank to get the help that I need. Right?
[10:03] So that agent really translated that um
[10:06] that that voice u question into an
[10:09] actual actionable piece of information.
[10:11] We took that information of the intent
[10:13] and the routing. We certainly got the
[10:15] customer to the right place and and
[10:17] alleviated the friction for the
[10:18] customer, but we were also building up
[10:20] the knowledge of what they were calling
[10:22] for and in a lot more detail than, you
[10:24] know, we were getting from the old phone
[10:26] based, you know, one, two, three pushing
[10:28] of buttons.
[10:29] >> Gotcha. So, we talked about value
[10:32] realization, right? So, talk to us about
[10:35] how this deployment helped your
[10:37] business. what were some of the KPIs
[10:39] that came out of it and um how did you
[10:43] ultimately receive value on the end of
[10:44] this?
[10:45] >> Yeah, sure. So, I think that you know
[10:47] when we think about a call center agent
[10:50] um you know their their their time is
[10:52] measured, right? I mean every phone call
[10:53] that comes into the call center agent
[10:55] you know every call needs to be handled
[10:57] and solved. So what we were doing was we
[11:00] kind of took the next step. So we
[11:01] identified our friction, we found a good
[11:03] solution and we started getting the
[11:04] data. Then what we moved on to is being
[11:06] able to say, "All right, how can we help
[11:08] automate some of the things before it
[11:09] gets to the agent, we took away the need
[11:12] for the agent to be doing things like
[11:14] authenticating the caller as they came
[11:15] in, we used dialogue flow to ask the
[11:18] basic authentication questions. So we
[11:19] were able to do things like screen pop
[11:21] the customer information. They had the
[11:23] full 360 degree view of that customer,
[11:25] an explanation of why they were calling,
[11:27] and we had done the work of building up
[11:29] our knowledge base so that we were able
[11:30] to surface the knowledgebased answers to
[11:33] our employees as that call was coming
[11:35] in, really arming them with the
[11:36] information that they needed.
[11:38] Additionally, because we were doing the
[11:40] intentbased routing, we were able to
[11:42] move to a place where we could bring in
[11:45] less skilled staff to handle the more
[11:47] repetitive problems. So that meant that
[11:50] we turn we translated um you know the
[11:53] two or three-w week period of having to
[11:55] train a call center agent who has to
[11:56] flip to five different screens to do
[11:58] something in a bank down to hey we need
[12:00] to train them on one topic one intent
[12:02] one purpose. Uh that allowed us to take
[12:04] training from multiple weeks down to
[12:06] days. Um, so those were some of the, you
[12:09] know, I'd say key metrics that, you
[12:11] know, really allowed us to pull back
[12:13] some of the value from our call center
[12:15] agents and be able to scale, uh, and and
[12:18] really, I think, drove some of the ROI
[12:19] of the initial, uh, build out of the
[12:21] application.
[12:22] >> So, I I love the story. Um, you
[12:26] mentioned that 2020 opportunity is
[12:29] knocking on your door. You have a
[12:30] challenge that's a make or break
[12:33] situation and a decision needs to be
[12:35] made. It's a problem that you cannot
[12:37] throw money or people at which no longer
[12:40] solves the problem but may have
[12:42] exacerbated the problem. So you turned
[12:44] to AI not because it is cool but it's
[12:48] because you it's a necessity at that
[12:50] point and then the team actually did
[12:54] deploy it successfully and created a
[12:57] differentiated experience for the
[12:59] customers as well as relieving that
[13:00] pressure internally and absorbing that
[13:03] opportunity that that that was knocking
[13:05] on the doorstep. So, so now I feel like
[13:09] if I were in your position, I there's
[13:11] two opportunities for me, right? There's
[13:12] two two choices. One is to to cash in
[13:17] the success from the story and then, you
[13:20] know, retire early, go to beach, die
[13:23] equities on the hand uh or um double
[13:27] down on finding another use case where
[13:30] where this could be applicable. I feel
[13:32] like unfortunately you may have picked
[13:34] them later.
[13:35] So there's no beach involved yet?
[13:37] >> There's no beach involved yet. All
[13:38] right. So, so tell us what what hap what
[13:41] did you do with this success for the
[13:42] first use case?
[13:43] >> Yeah, sure. So, um the first use case,
[13:46] you know, what we did was we took the
[13:48] data that we were now gathering and we
[13:50] were able to pivot and certainly do more
[13:52] case deflection. But one of the key
[13:55] things that was also happening was with
[13:57] success and attention, I think we all
[13:59] know brings not necessarily always the
[14:02] best actors. So we were also faced with
[14:04] a bit of a challenge in the fraud arena
[14:06] where we were receiving a lot of dispute
[14:08] cases, friendly fraud type of activity.
[14:10] So what we did was we built upon what we
[14:13] had already put in place with Google
[14:14] dialogue flow to start handling the
[14:17] dispute cases that were coming in
[14:19] differently. We reused the
[14:21] authentication mechanism from our agent
[14:23] to authenticate callers before we handed
[14:26] them off to the dispute intake team to
[14:28] make sure we had a good understanding of
[14:29] who the customer was and that we were
[14:31] able to associate them with a case
[14:33] inside Salesforce. We were able to then
[14:36] use the 360deree view of our customer,
[14:38] the real-time disputes API from our
[14:41] partners over at Viserve, and through
[14:43] application integration, one of Google's
[14:45] frameworks uh for connecting our
[14:47] enterprise systems, we were able to feed
[14:49] that into Gemini to allow us to perform
[14:52] a quick risk assessment that we were
[14:54] able to write back to the Salesforce
[14:56] case, arming our loss mitigation team
[14:59] with the critical information that they
[15:01] needed. You know, just like I said
[15:03] before, you know, you've got call center
[15:04] agents that are flipping between five
[15:06] screens. Your loss mitigation team, you
[15:08] know, that they're digging into at least
[15:10] three or four or five different systems
[15:12] to have a good picture of who the person
[15:15] is when they're facing a dispute case.
[15:16] That takes a lot of time. Uh so for our
[15:19] loss mitigation team, the concept of
[15:21] having Gemini help them identify the
[15:24] most obvious risks, you know, that were
[15:27] coming in through these dispute cases
[15:29] has saved them a ton of time. Um, and
[15:32] it's allowed us to prioritize based upon
[15:34] the risk level that Gemini was able to
[15:36] put together by looking across the
[15:38] transactions and the dispute data and
[15:40] the customer history. Uh, whereas in the
[15:43] past, all they could look at really was
[15:45] the most highest dollar value disputes
[15:48] that were coming in. They weren't
[15:49] catching the repetitive, you know, oh,
[15:51] this person filed a $50 dispute this
[15:53] week. Oh, they filed a $50 dispute last
[15:55] week. You know, it really did allow them
[15:57] to connect the dots. So it was injecting
[16:00] AI and Gemini into the workflow that
[16:03] really saved us all this time on the uh
[16:05] fraud side.
[16:06] >> Oh, very cool. So, so you are
[16:07] incrementally now building and using AI
[16:10] to drive one of the internal process
[16:13] which is fraud. And I love that not only
[16:17] it helps you semi-automate and maybe
[16:19] automate uh the information gathering
[16:22] synthesis from multiple different data
[16:24] sources but that's actually also happens
[16:26] to be a great use case for AI to help
[16:29] the agents internally uh drive this very
[16:32] fast and and I love the discovery which
[16:34] is like hey
[16:37] you set out to solve the fraud problem
[16:40] at scale but now you're also handling
[16:43] cases that would not have been possible
[16:45] before because there's capacity to even
[16:49] manage that.
[16:51] >> Now going back to the ROI which is I
[16:53] think the heart of the conversation is
[16:54] okay you you did all this what does it
[16:57] mean from from a from a hard ROI
[16:59] perspective?
[17:00] >> Yeah. So I mean I think that you know uh
[17:02] first of all like the the fact that
[17:04] we're able to automate all this and pull
[17:06] it together allows our loss mitigation
[17:08] team to get to more of those dispute
[17:09] cases. I also think that introducing the
[17:13] authentication mechanisms allowed us to
[17:15] also reduce the number of inbound fraud
[17:17] cases. I think it like kind of pushed
[17:19] the fraudsters in a different direction.
[17:21] Fraudsters tend to like the easiest
[17:23] target, right? You introduce a little
[17:25] bit more friction for them, they tend to
[17:26] go a different direction. Um so that
[17:29] said you know I think that for us you
[17:30] know we definitely saw a huge decrease
[17:32] in the overall uh losses that we were
[17:34] experiencing uh in the reggie category
[17:37] the year after we introduced uh this
[17:39] agentic fraud uh approach.
[17:41] >> Very cool.
[17:43] >> Yeah. And you know going back to this
[17:46] framework that we put together right
[17:48] seems like you're following every step
[17:50] here right we've identified the friction
[17:53] points which is the detection fraud
[17:55] detection and then we actually deployed
[17:57] agents here where we are now having a
[18:00] aentic workflow to help mitigate that
[18:03] fraud uh detection process. The last
[18:07] thing that we have not talked about is
[18:09] how do we transform the investment
[18:12] capital from it. So tell me a little bit
[18:14] more about how one United Bank was able
[18:17] to use that investment to pursue other
[18:20] opportunities and capabilities.
[18:22] >> Yeah. So um I think that's that's a
[18:24] great question and I think honestly you
[18:26] know you build on success right that's
[18:28] what you do uh and one United being a
[18:30] missiondriven organization you know we
[18:32] really uh have been putting ourselves
[18:34] out there for years now focusing on how
[18:37] we can uh improve economic advancement
[18:40] uh across the country. Um up on the
[18:42] screen you'll see that there's a number
[18:43] of different statistics here that really
[18:45] just highlight what is out there um
[18:47] across the country. A true struggle um
[18:50] in in wide areas of um you know the our
[18:54] customer base and and broad-based
[18:57] America. You know, I think the thing
[18:58] that always strikes me, the number that
[19:00] I of all of these, it always makes me
[19:02] the most, you know, um, realize the need
[19:05] is the fact that there's a real problem
[19:07] where people can't handle an unexpected
[19:10] $400 expense. Um, you know, that's
[19:13] that's a real struggle and a real
[19:14] problem that tells you how tightly
[19:16] individuals are managing their finances.
[19:18] We've got people who are watching every
[19:20] dollar that comes in and every dollar
[19:21] that goes out. Uh, and they really don't
[19:24] have anyone helping them, right? there's
[19:26] no one there to be have their backs on a
[19:28] normal day in and dayout basis. So what
[19:30] we did was in the in the face of the
[19:32] success we had deploying agents against
[19:34] those couple of ROI use cases that we
[19:37] saw was that we actually built out um a
[19:39] new a new service a differentiator our
[19:42] transformation item
[19:43] >> right we we built out um a service that
[19:46] we're calling wise one uh wise one is a
[19:49] personal financial coach um in your
[19:52] pocket we we built it with Google Gemini
[19:54] uh we built upon our success in the
[19:56] Google dialogue flow arena using
[19:59] conversational agents now and generative
[20:00] AI uh to be able to certainly service
[20:04] customers but also help them with their
[20:05] finances. Uh and how do we do that,
[20:08] right? How is it possible that we're
[20:09] going to be able to help them understand
[20:10] when there's an opportunity to save or
[20:13] when there's a need to, you know, move
[20:14] money to cover a bill? Uh what we're
[20:16] able to do is we um use an agentic
[20:20] framework, right? Um to accomplish this.
[20:22] Uh we have a number of enterprise data
[20:25] streams that we're uh leveraging to uh
[20:28] bring into the wise one agent. Uh you
[20:30] know things like our customers budgets,
[20:32] their goals. Um certainly you know we're
[20:35] using our FAQs and our knowledge base to
[20:37] help guide them as well as the decades
[20:40] of financial literacy content that we've
[20:42] been building up to be able to answer
[20:44] questions. So, we use the Wise One
[20:47] orchestration Engine really to be the
[20:49] financial coach that's basically got all
[20:52] these subcategories of classified agents
[20:55] that it can come back and answer
[20:57] questions on, be instructive,
[20:58] supportive, um, and really be kind of a,
[21:02] you know,
[21:03] your your your safety net in your pocket
[21:06] when it comes to your finances. That
[21:07] independent third party, you know, that
[21:09] can come in there. You know, I think the
[21:11] thing is is that, you know, this is
[21:12] something that others, you know, may
[21:14] strive to do, but the the thing that One
[21:16] United is uniquely positioned to do as a
[21:18] bank and a financial institution is be
[21:20] stewards of your financial data and give
[21:23] you the ability to make great um great
[21:25] financial uh decisions and achieve good
[21:28] outcomes. So, we're really excited about
[21:30] being able to move forward our mission
[21:32] uh through a service like this um and
[21:34] leveraging the technology to really do
[21:36] something different.
[21:38] I am definitely going to be the user of
[21:40] vice one. Um even though I do follow a
[21:43] certain financial discipline, I think
[21:45] there is always an opportunity to learn
[21:47] more and I love the fact that bank's
[21:51] mission is to drive the literacy from a
[21:54] from a from a objective standpoint and I
[21:57] think using AI and I guess in this case
[22:00] army of AI agents to drive that mission
[22:03] forward um it's it's truly unique. I
[22:06] haven't seen this across many other
[22:07] financial institution as well. And kudos
[22:10] to you and your team for driving this.
[22:12] >> Yeah. Well, I think you know again it's
[22:13] the early successes. It's building on
[22:15] successes that so you can go into the
[22:17] boardroom and have the conversation
[22:18] about being able to build a service like
[22:20] this, right? You have a proven track
[22:22] record of success by reducing those
[22:23] frictions and delivering the
[22:25] automations. That's what gives you the
[22:27] opportunity to transform the business.
[22:28] So
[22:29] >> I I want to go back to one thing that
[22:31] you mentioned in your introduction which
[22:33] is 150 employees.
[22:35] >> That's right.
[22:37] Um I think I think it's incredible right
[22:39] like you have a living proof point of
[22:43] lean teams using AI as a capacity a
[22:48] workforce multiplier in terms of driving
[22:52] um you know maybe differentiated
[22:55] customer experience externally or
[22:57] driving internal uh productivity or even
[23:00] you know driving a mission forward for a
[23:02] greater good. So love love the story.
[23:05] Thank you James for sharing that. Um
[23:08] what we have next is a demo um for one
[23:12] of the use cases I think going along the
[23:14] similar trajectory that Andre was
[23:16] speaking about is identifying the
[23:19] friction uh moving to automation. Um and
[23:22] then finally thinking about
[23:23] transformation. So one of the use case
[23:25] that we have is uh cloud phops. We have
[23:28] found that several uh organizations may
[23:32] have PHOPS team which are lean. Some may
[23:35] not even have a PHOps team and there are
[23:37] uh one person wearing multiple hats
[23:39] managing cloud spend at scale. Right. So
[23:42] I think the idea here is how do we have
[23:46] AI and AI agents help you manage cloud
[23:48] spend because I think we know that you
[23:51] just have to be uh uh responsible in
[23:54] terms of spending your dollars in cloud
[23:56] and thinking about ROI throughout. So
[23:59] let's switch to uh the demo piece.
[24:07] Are you all able to see the screen?
[24:10] All right. Uh, quick poll.
[24:14] Dark mode or light mode?
[24:18] >> Dark mode.
[24:22] Okay.
[24:24] Light mode.
[24:27] Oh.
[24:28] >> Oh, boy. Pretty even.
[24:30] >> I feel like light mode missed it by
[24:31] three people. Just quickly counted that.
[24:35] All right. So, we will switch it back to
[24:37] dark.
[24:38] All right. So this is my uh interface
[24:42] into speaking to army of AI agents. I'm
[24:45] talking to a platform supervisor. So I'm
[24:48] just going to ask hi
[24:51] what can you do?
[24:55] um just to understand the capabilities
[24:57] that the uh platform has and there is
[25:01] number of things it can do from
[25:04] inventory to cost optimization analysis
[25:08] pricing ROI pull request governance
[25:12] policy architecture okay that's a lot so
[25:14] I'm going to ask a simple question and
[25:16] say
[25:18] show me top five services spent in the
[25:23] last 30 base.
[25:27] So I'm interested in just generally
[25:29] understanding where my teams are uh
[25:32] spending money on cloud um and what goes
[25:36] on behind the scenes is orchestrator
[25:39] based on the intent of my question
[25:41] reroutes it to a specific specialist. So
[25:45] in this case that specialist happens to
[25:47] be billing. Billing uh is connected with
[25:50] Bitquery through MCP server. It creates
[25:53] a SQL on the fly, looks at the billing
[25:57] data and come up with the answer. So
[26:00] that's what this information is.
[26:04] Um, now just to show you, we have
[26:07] several agents working behind the scenes
[26:09] depending upon the intent of the
[26:11] question I have. Um, and it's the
[26:15] platform supervisor that decides and
[26:17] defines whether I need access to one
[26:20] agent or multiple agents to accomplish
[26:23] users tasks or goal.
[26:27] So I'm going to ask a little bit more
[26:28] deeper question. Um, show me cost
[26:32] breakdown
[26:33] by namespace within GKE.
[26:42] So um for anyone who may not have worked
[26:46] directly with Kubernetes engine,
[26:47] Kubernetes engine tends to be a little
[26:49] bit more complicated because of its
[26:50] multi-tenant architectures. You could
[26:53] have multiple app teams using same
[26:56] cluster for application development and
[26:59] I think what that does is it creates a
[27:02] challenge of allocating shared services
[27:04] cost. So typically what teams do is they
[27:08] use different namespaces for running
[27:11] their services within their nodes. So
[27:14] you know typically when you want to look
[27:15] at hey what is my service A cost versus
[27:18] service B cost um just looking at a
[27:21] cluster information it may not be
[27:23] enough. So you may need to go a level
[27:25] down there as well. So this is again
[27:28] based on my question it's still billing
[27:30] but now it is able to find out nested
[27:33] information um that may not be readily
[27:35] visible uh if you if you look at the
[27:38] billing file uh but it is able to
[27:40] identify it and capture it and then show
[27:41] it to the team. Um u it actually is also
[27:45] pretty interesting to see that what part
[27:48] of my cluster is actually being
[27:50] unallocated meaning uh it is not doing
[27:53] any useful work for my business
[27:55] objective. So in this case I can see
[27:57] unallocated spend is roughly around $350
[28:00] uh for a for a given month. So that was
[28:02] interesting. Um and uh Gemini here also
[28:07] gives me some important tips about hey
[28:09] this is unallocated maybe that's the
[28:11] avenue that you need to look into in
[28:12] terms of optimizing your spend.
[28:16] So I will go ahead and say that okay my
[28:18] computer engine cost is 315 for last
[28:21] month create a budget alert for me
[28:28] as soon as the cost
[28:31] goes over let's say 500.
[28:35] Um so again the idea here is um if it
[28:38] goes above 500 just let me know so I can
[28:40] do something about it. Um, in this case,
[28:43] based on the intent of my question, it's
[28:45] the budget specialist that is being
[28:46] called under the hood. Um, and it goes
[28:49] in and it actually creates a budget for
[28:51] me. So, if I go to my cloud console, if
[28:55] I look at my budgets and alerts
[28:58] um, and you see that the budget has been
[29:02] created u by the AI agents behind the
[29:05] scenes.
[29:07] Um, this actually does few things. One,
[29:10] it gives your finance team, your PHOPS
[29:13] team, your app teams ability to create
[29:16] and manage their own budgets and kind of
[29:18] driving that accountability to the edge
[29:20] without having them either work through
[29:23] themselves through AM permissions,
[29:25] authentication, authorization or depend
[29:28] on some other team to be able to do so
[29:31] for them.
[29:33] All right, so I think we showcased that.
[29:35] Let me go ahead and ask a little bit
[29:36] more detailed question about um now we
[29:40] are turning towards you know just show
[29:42] me information about uh clusters that I
[29:45] might be running in my uh systems. Um so
[29:48] in this case uh it identifies based on
[29:50] the intent of my question that this is
[29:52] an inventory related information. It's
[29:54] asking for metadata. So again it uses
[29:57] MCP but this time it uses MCP for
[29:59] Kubernetes to get cluster information um
[30:04] for me. So it says I have one cluster
[30:06] it's running in central one um it's
[30:09] using E2 standard 4 um and then some of
[30:12] the information that is important for me
[30:14] in terms of phops profile any labels
[30:16] that are attached to it for cost
[30:18] allocation.
[30:21] Um it also provides me a useful question
[30:23] at the end. Would you like to scan it
[30:25] for cost optimization opportunity?
[30:28] Oh yes, please let me know if there are
[30:33] any optimizations.
[30:38] So this one is a slightly involved agent
[30:42] um because what it is doing now is it is
[30:46] looking at your observability data. So
[30:49] your CPU utilization, memory
[30:52] utilization, it looks at for last 30
[30:54] days. Uh it looks at uh the requests,
[30:58] the usage, the limits. Uh basically it
[31:01] uses Goldilocks behind the scenes to
[31:04] identify whether a workload can be right
[31:07] sized in a safe manner because yes, we
[31:10] do want to optimize cost but we don't
[31:12] want there to be a performance issue. Um
[31:15] and this is a balance which is extremely
[31:18] hard to maintain and this requires quite
[31:21] a bit of analysis behind the scenes. So
[31:24] what what what I have feed into the
[31:26] agent behind the scenes is you look at
[31:28] CPU for P99 which is which means 99% of
[31:32] the time if the CPU usage is low we can
[31:36] be safe that it can be right sized. Now
[31:39] we also don't want to right size it to
[31:41] it to its max. So we might be adding a
[31:43] buffer of 20% or 30% on top of it.
[31:46] Right? So we are taking into account all
[31:48] of the guardrails, all of the
[31:49] performance factors behind the scenes
[31:52] before a recommendation is created. This
[31:54] recommendation is truly custom based on
[31:59] your needs. Right? So it's not coming
[32:02] from PHOPSHub. It's not coming from
[32:03] active assist. It's something that a
[32:07] human would technically do in order for
[32:09] this to come up. AI agent is doing
[32:10] exactly the same behind the scenes. So I
[32:13] have a report. Um 95% of my cluster is
[32:18] idle. Um it gives me information about
[32:21] utilization and it also tells me there
[32:23] is a primary target of this app that can
[32:26] be right sized.
[32:29] So I will say if I were to implement
[32:33] this,
[32:36] how much would I save?
[32:40] Which is a normal question about like
[32:43] hey are we even thinking about this in
[32:44] the right direction or not? And what it
[32:46] does is it again calls a different sub
[32:49] agent pricing specialist that looks at
[32:51] my contractual negotiation with Google
[32:54] Cloud. It identifies the node that I'm
[32:58] running in the location US Central 1 for
[33:02] E2 standard and then it identifies vCPU
[33:06] memory savings and then club that
[33:08] information in one page. So essentially
[33:11] I can save 16 bucks which is good enough
[33:13] for two Starbucks. So I would go ahead
[33:16] and
[33:18] don't go to Starbucks reserve then you
[33:20] will get only one um but regular
[33:22] Starbucks is fine. So, I'll go ahead and
[33:25] I'll say, "Okay, I want to save $16."
[33:28] Um, so I'll go ahead and create
[33:31] implementation.
[33:33] Yes, please go ahead and
[33:36] implement this.
[33:40] Now, now if I were to do this in
[33:42] production, I think several people will
[33:44] get a heart attack. Um, so what we
[33:48] really want AI agents to be doing is not
[33:51] implementing it, but do something smart.
[33:54] So it actually looks at my code repo. It
[33:57] creates what we call a pull request,
[34:00] which means here is my suggested changes
[34:02] in your code and then it reroutes the
[34:06] request to an application team. Do you
[34:09] agree with this?
[34:11] So I'm going to click on this pull
[34:14] request. I'm going to look at the code
[34:16] change that was done by my EI agent and
[34:20] I'm going to pretend that I know what
[34:22] it's doing and I'll say yes this is
[34:25] amazing
[34:27] please go ahead and merge this.
[34:31] What happens behind the scenes is this
[34:33] is what we call infrastructure as a
[34:37] code. So as soon as my
[34:40] code has been merged, which means it has
[34:42] been reviewed and marked safe for
[34:45] production, um it actually creates a
[34:48] CI/CD job behind the scenes. I'm using
[34:50] GitHub action runner to do this. It
[34:52] connects to my cloud, connects to my
[34:54] Kubernetes, deploys my code, rightize my
[34:58] workloads. Everything happens in one
[35:00] swoop.
[35:07] So this will take
[35:12] two seconds to be done. I waited for it.
[35:15] Okay. So it's done now. Um which means
[35:18] my workload in Kubernetes should have
[35:22] been right sized.
[35:25] So I go to Kubernetes engine,
[35:32] go to workload
[35:34] and then I think this is the app and we
[35:38] will wait for a few more seconds and you
[35:40] will see this red line drop which is the
[35:43] vCPU usage and the memory usage you see
[35:46] on both the charts. We will come back to
[35:48] this. Um let me go ahead and show you
[35:51] one more thing.
[35:57] back.
[36:01] Okay. The other thing is an often time
[36:05] question comes back is like allocation
[36:07] like I want to allocate my cost to
[36:08] different teams and for that labels and
[36:11] tag should be applied appropriately and
[36:14] that itself can go into its own
[36:17] different barrel of conversations with
[36:19] the app teams. Um so in this case you
[36:21] can actually ask AI agent to do uh to
[36:24] help you with labeling management as
[36:26] well. So show me cloud resources that
[36:30] doesn't
[36:32] have labels environment and then it
[36:35] scans across my cloud um and then will
[36:40] provide me information about uh
[36:41] resources that is missing the label
[36:43] environment. So in this case the cluster
[36:45] is actually missing it.
[36:48] So I'll say go ahead and apply a label
[36:56] uh environment equal to finance.
[37:00] Um and the idea here is not only agent
[37:02] is able to identify it but it is also
[37:04] able to rectify information on your
[37:06] behalf. Now, in this case, I did give
[37:09] agent access to directly go ahead and
[37:11] make the change because I know it's a
[37:12] safe change to be made regardless if
[37:15] it's a production environment or a
[37:16] non-production environment. So, that has
[37:19] been done. Okay, I'm coming back here.
[37:22] You see the spike drop here, which means
[37:23] the workload has been right sized. I go
[37:25] to cluster. I look at the labels and you
[37:28] will see that the environment label is
[37:30] now
[37:32] being applied.
[37:37] And I'm going to ask a final question,
[37:40] which is, what's the weather
[37:45] in Jersey City?
[37:49] I'm from New Jersey, so I just care
[37:51] about what's going on there. And the
[37:53] agent will come back and say, "No,
[37:55] that's outside of my domain knowledge."
[37:57] So the idea here is that you can
[38:00] certainly guardrail your agents to to
[38:02] ground it into the data source that you
[38:04] have. Um you can however ask any
[38:07] questions about GKE like if there are
[38:11] different modes and what should I be
[38:12] using um and it will certainly give you
[38:15] answer to that but if you ask anything
[38:16] outside of this domain it will restrict
[38:19] you from doing that. So I hope you all
[38:22] enjoyed the demo and