Latest / CXOTalk / What is Google's Agentic AI Strategy? (Explained by Google Cloud's CTO) | CXOTalk #897
Transcript
- 0:00Google is betting billions on AI agents and multimodal models.
- 0:06But what's really behind the strategy?
- 0:09Today on CXO talk number 897, Will Grannis, chief technology
- 0:16officer of Google Cloud, takes us inside what you need to know
- 0:21now, I'm your host, Michael Krigsman.
- 0:24Let's get into it. My team and I work with our top
- 0:28roughly 150 customers around the world world to make sure that as
- 0:31this technology factory that we're spinning up all the time
- 0:34and releasing new products, that they know how best to leverage
- 0:37that technology for their own specific business use cases and
- 0:41needs. And then on the other side, we
- 0:43also work on kind of new and emergent problems in technology
- 0:47to try to solve them. An example of this could be
- 0:50right now as you build 10s, hundreds, even thousands of
- 0:54agents, agent alignment against tasks is a really difficult
- 0:58problem. So that's one of the things that
- 1:00more of an R&D mode we work to make easier for our customers.
- 1:04What's happening in the agent world?
- 1:06And then we can talk about what you're doing specifically with
- 1:09agents. Well, agents have quickly become
- 1:11kind of the surface through which many consumers and
- 1:15businesses are starting to experience the power of AI.
- 1:19And if you, if you think about in historical context, we've
- 1:22always been looking to automate tasks that human beings, it
- 1:27isn't necessarily the highest expression of our value.
- 1:29So for example, you know, we in spreadsheet software, we have
- 1:34been running formulas for many years to try to automate the
- 1:37calculation of certain numbers and to automate a task of
- 1:41calculation. That's something computers do
- 1:42really well and humans don't need to necessarily spend their
- 1:45time on. If you've ever worked on a
- 1:47laptop and you've ever tried to put a macro or like an
- 1:50automation on your laptop, that was, you know, an early form of
- 1:53task automation. Because if you're continuously
- 1:56clicking or doing the same thing over and over again, you want to
- 1:58automate that so it gets done a little bit more quickly and more
- 2:02efficiently. So agents are kind of like
- 2:05almost like this third wave of automation and intelligence and
- 2:09that they can take intent, they can take it in a variety of
- 2:12ways. It can be typed, it could be
- 2:14spoken and it can execute tasks on your behalf.
- 2:18And this obviously has profound potential impact in today real
- 2:22world impact across number of consumer use cases as well as
- 2:25business to business use cases. And it's not unusual in my, you
- 2:29know, I mentioned earlier, I work with, you know, 150 ish
- 2:32customers around the world, top customers and across all
- 2:36industries. It's not unusual for me to run
- 2:37into an organization these days like Highmark Health, which has
- 2:42an agent that they've deployed, an agent platform they've
- 2:45deployed where their employees who have a routine questions
- 2:48about benefits, processes, procedures within the within the
- 2:54organization. Before they'd have to go to like
- 2:56websites and scroll through it. They'd have to ask somebody if
- 2:58they know you know how to how to book travel or how to expense
- 3:03something or how to provide or how to set up a conference room.
- 3:07All of those kind of like basic everyday tasks the organization
- 3:10has to do. All that knowledge now has been
- 3:14indexed and is now available through a very simple interface
- 3:17that today I think roughly 60,000 users at Highmark Health
- 3:22are interacting with agents to get the knowledge that they need
- 3:25to execute the work that they need to execute almost
- 3:28immediately. Now let's take a moment to learn
- 3:31about Emeritus, which is making CXO talk possible.
- 3:35If you're a business leader navigating change or driving
- 3:39growth, explore executive education programs with
- 3:44Emeritus, a global leader. They offer programs developed in
- 3:48collaboration with top universities, designed for
- 3:51decision makers like you. There's a program tailored to
- 3:55your goals, whether that's AI and digital transformation or
- 3:59strategy and leadership. Find your program at
- 4:03www.emeritus dot org. What are the the complexities of
- 4:10creating these agents and what are you doing?
- 4:15What Google doing to make it easier?
- 4:18We are consolidating our own approach to agent development
- 4:23and AI development. And just yesterday we launched
- 4:26something called Gemini Enterprise, which brings
- 4:28together all the different pieces of AI and agent creation
- 4:33and deployment and management into one singular stack and
- 4:39package. And what this has done is this
- 4:41created what we like to refer to as the new front door for AI in
- 4:45the workplace. And it consists of 6 core
- 4:48components. So first, the chat interface or
- 4:52the AI interface. This is something that people
- 4:54are now used to using. Four years ago was a brand new
- 4:56concept. Now it's, you know, it's very
- 4:58commonplace. So this is the first piece of
- 5:01this Gemini Enterprises, the chat interface.
- 5:03Now underneath that are all the models that feed the
- 5:06intelligence. So this could be Gemini 2.5 Pro,
- 5:09it can be Gemini Flash, it could even be multimodal models such
- 5:12as VO and Imagine. And those are the models that
- 5:15help people execute their tasks. And that's the brains of the
- 5:18execution engine behind agents. Underneath that is the agent
- 5:22platform. So individuals within an
- 5:24organization can build this kind of task automation level of
- 5:27agents, but also multiple organizations within one company
- 5:31or firm can come together and build multi agent workflows
- 5:36using this third tier, this agent platform.
- 5:40Underneath that you have out-of-the-box agents.
- 5:42So we have data science agents, we have deep research agents, we
- 5:47even have coding agents that are available to get value from the
- 5:51platform out-of-the-box and then create scaffolding and enrich
- 5:54them over time. There's anybody who works on
- 5:58agents knows you need access to data and context.
- 6:00So there's also a set of connectors to 3rd party data
- 6:03sources like your ServiceNow implementations, workflows,
- 6:08Oracle, Salesforce, as well as even JIRA, Confluence and, and
- 6:15databases like Google Bigquery. So those connectors are a key
- 6:18part of this Gemini enterprise stack.
- 6:20And then finally, and maybe most importantly, as agents
- 6:24proliferate governance and security, so making sure that
- 6:28the agents are created by the right people with the right
- 6:31policies, have the ability to execute their tasks safely and
- 6:34securely within an enterprise context.
- 6:36And so that's job number one for us in making agents more useful
- 6:41and more available was bringing together all of the ways that we
- 6:44help organizations with AI agents.
- 6:47There are a couple of, there are there are also a couple of
- 6:49really important things I think translate beyond even Google's
- 6:53approach to this for any organization in general when
- 6:56they're building and deploying agents.
- 6:59So number one, I mentioned this earlier is context.
- 7:04Think of agents as a very explicit task execution machines
- 7:10and within an organization, first step is to make sure that
- 7:13you actually have written down or available to, you know this
- 7:18task automation, your business process or your business
- 7:21workflow. One interesting work right now,
- 7:23Michael, is that in regulated industries such as financial
- 7:28services, healthcare, public sector, those organizations for
- 7:32many years have been highly documented in terms of their
- 7:35business process and workload, their standard operating
- 7:37procedures. And so they actually have a jump
- 7:39start into the agent world, where as we tend to think of
- 7:42technology as you're coming to regulated industries second or a
- 7:46little bit later in the world of agents, they're seeing time to
- 7:50value a little faster because they already have all of those
- 7:53key business processes and data documented that could then feed
- 7:56the agent execution and the agent rules through a workflow.
- 8:00So the context and the data is critical.
- 8:02I just want to remind everybody that right now you can ask your
- 8:07questions. If you're on Twitter X, use the
- 8:11hashtag CXO talk. If you're watching on LinkedIn,
- 8:17just pop your questions into the LinkedIn chat.
- 8:21And this is from Anthony Scrifignano, who is a very
- 8:24prominent data scientist. And he says one of the biggest
- 8:28challenges with RPA robotic process automation was
- 8:33automating previous tasks that were not designed for
- 8:38automation. And so how do you advise clients
- 8:42to be agentic ready before they rush to deploy agents in the
- 8:49enterprise? Every customer that I've worked
- 8:52with that has been successful in these early agent
- 8:54implementations, they've all picked a problem or a business
- 8:57process for which they had data and context.
- 9:00They had standard operating procedures.
- 9:03It was a, it was a known problem in a workflow.
- 9:06And so to your point around trying to automate the wrong
- 9:10thing, they take the time and the diligence and they, and it's
- 9:13kind of in the old world, we would call this application
- 9:15rationalization. I guess in the new world we
- 9:17might call this, you know, agentic workflow
- 9:19rationalization. And that's taking inventory of
- 9:21all the things that you might choose to do and being very
- 9:24methodical about picking one that has the data, the context,
- 9:27the documentation. That's absolutely critical.
- 9:31Then once you, you know, you've identified that very specific
- 9:34workflow that you're trying to automate and execute, it's also
- 9:39extremely important to have a sense of how you're going to
- 9:42evaluate completion and success And evals actually, and
- 9:48evaluations tend to be one of the stickiest parts of the kind
- 9:52of agentic development workflow today because think about it
- 9:55like a manager at a desk. And in the past, the work that
- 10:00that would come to that manager would be gated by how fast
- 10:03humans could bring decisions or bring execution tasks to that
- 10:07manager. And so, you know, it's a steady
- 10:09stream of them. Today, agents and automation can
- 10:12bring many, many tasks for approval, and it overwhelms a
- 10:17human's ability to make decisions fast enough.
- 10:19So the only way out of that trap is to design evaluations and AI
- 10:23as a judge or AI as a critic, or AI as the evaluator, so that the
- 10:28AI itself and the agents themselves can evaluate task
- 10:31completion, whether it's good enough or not, and either send
- 10:34agents back to do a better job or move on to the next step in
- 10:38the workflow. And that is a unique problem
- 10:41that kind of transcends our PA, which is RPA being kind of like
- 10:44single task, single automation. These are multi task, multi step
- 10:48automation concerns now. And you know, looking ahead a
- 10:52little bit, one of the things that we see is that most of the
- 10:55value from agents will be derived from multi step, multi
- 10:59task, multi agent work flows. And I, you know, an example of
- 11:03this could be there's a really interesting problem in telco and
- 11:10in customer service in telco. And that is that when your Wi-Fi
- 11:15goes out, you know, that's a very, it's, it's in our hyper
- 11:19connected world, Wi-Fi going out is a very emotional moment for
- 11:22human beings, myself included. And so when the Wi-Fi goes out,
- 11:27the first thing you want to do is you want to get help.
- 11:28And the typical workflow for getting help is you go to the,
- 11:31the website and you try to scroll through, you know,
- 11:33whoever your provider is or the mobile app, you try to find how
- 11:36to debug the problem. It's a very, it's a very chat
- 11:40kind of 2010 interaction model. Well, in the world of agents.
- 11:45And there's a telco in the UK that's doing this right now that
- 11:51we're working with. What they do is they have an app
- 11:54experience. And if someone is experiencing
- 11:56an outage of the Wi-Fi and they still have access to cellular,
- 12:00they pop up a live multimodal experience where the customer
- 12:05support agent will come up. They will have the individual
- 12:08point their camera at the equipment that they have in the
- 12:11house. And they're able to figure out
- 12:13how to do troubleshooting based on the equipment that's in the
- 12:15location, you know, of the apartment or the house, that
- 12:18person said through a completely native multimodal experience.
- 12:22And so that is an example of, you know, you have an
- 12:24orchestrator agent, you have sub agents underneath the hood that
- 12:28are characterizing, you know, a video and, and orchestrating the
- 12:32chat and orchestrating the text. So all these different agents
- 12:35are coming together to create this very seamless support
- 12:38experience. And that is a very that's a very
- 12:42modern multi, multi step, multi work, multi agent workflow.
- 12:47Now let's quickly hear from Emeritus, which is making CXO
- 12:51talk possible. If you're a business leader
- 12:54navigating change or driving growth, explore executive
- 12:58education programs with Emeritus, a global leader.
- 13:03They offer programs developed in collaboration with top
- 13:06universities, designed for decision makers like you.
- 13:10There's a program tailored to your goals, whether that's AI
- 13:14and digital transformation or strategy and leadership.
- 13:18Find your program at www.emeritus.org.
- 13:24You made a very interesting point several times, which is
- 13:31this idea of intelligence or judgement.
- 13:34I don't think you use those terms as being a key
- 13:39distinguishing factor. It seems to me that's the point
- 13:43that actually changes all of this.
- 13:46It's absolutely true, and we call it evals or AI as judge or
- 13:51AI as a critic, but it is the step at which companies
- 13:56themselves and organizations themselves also have to, in some
- 14:00cases for the first time, actually document what their
- 14:03decision rules are. You know, I come from the
- 14:06enterprise. I've been working in enterprise
- 14:07technology for a few decades now, longer than I would like to
- 14:11specify very specifically. But one of the things that
- 14:16happens in a lot of companies and most companies that I work
- 14:19with and I've worked in is there's the process, there's the
- 14:22documentation, and then there's the human judgement, unspoken
- 14:27norms and things that sit underneath the surface.
- 14:30And so in order to give, you know, the software is very
- 14:33explicit, if you say, you know, I want, we're working with a
- 14:38company right now and part of what we're working on is people
- 14:44want to buy HomeGoods, but they want to be able to see it in
- 14:48their, in their kind of living space before they choose.
- 14:51And so if you're going to create an AI as kind of a designer to
- 14:55automate these processes and to create this kind of like before
- 14:57you buy it, like let's put it in your space, that's a very
- 15:01complex problem because you're dealing with physics, you're
- 15:03dealing with inventory, real time inventory.
- 15:06There's a lot of very complicated pieces to making
- 15:08this design work really well. And So what we've had to do is
- 15:13we've had to architect evaluation at multiple steps in
- 15:16the process. The first step is, OK, so
- 15:19someone, the AI has come back and said, you know, I want to
- 15:24create this type of set up in this room.
- 15:26Well, First off, does it defy the laws of physics?
- 15:29A funny little quirk is one thing that we found when we were
- 15:31first starting to do this is that, you know, manipulating
- 15:33objects in three-dimensional space.
- 15:35Sometimes we'd have like a couch sitting on its side.
- 15:38And so we had to actually instruct, you know, give
- 15:41explicit instructions to the evaluating agent to say, you
- 15:44know, it has to sit with, you know, the bottom oriented
- 15:46towards the bottom of the space. We actually had to Orient the
- 15:49physics. So that's step one.
- 15:51Step 2 was, is this item actually, you know, an item
- 15:54that's in our inventory, you know, a very specific item.
- 15:57And so that was a second evaluation and critique 1/3
- 16:00evaluation and critique was there are certain clusters.
- 16:04If you, if you ever go to like a home design showroom, you'll see
- 16:07that they purposefully put like an end table and a chair and a
- 16:09lamp and a, and a, and a client. They put them all together in
- 16:13clusters. One of the things we learned is
- 16:14that these clusters should meet certain brand criteria or brand
- 16:18sensibilities. And so depending on which
- 16:20company are working with, they may have different clusters.
- 16:23So the clusters had to match the design sensibilities and the
- 16:25brand sensibilities of that. So we're already in like 4
- 16:28different evaluation steps and that it's that methodical
- 16:34definition and methodical evaluation and instrumenting all
- 16:38of that, that to your point, that really unlocks the scale.
- 16:42Once you get that right, then you can send lots of different
- 16:45requests and lots of different users through that same
- 16:47pipeline. And it meets the brand
- 16:48guidelines that matches the laws of physics.
- 16:51And it actually matches inventory can create some pretty
- 16:53magical experiences. It's really interesting the way
- 16:56you describe these eval points, putting them through the entire
- 17:03process frequently enough. Would it be correct to say that
- 17:08you are making many, many, many, many small course corrections?
- 17:13Is that an accurate way to say it?
- 17:15Yeah, that's absolutely true. And, and one of the most
- 17:17interesting things to me about agents and I've been around AI
- 17:21both, you know, kind of pre generative AI and, and now you
- 17:24know, generative AI and agents for about 20 years.
- 17:27And one of the things that's really interesting to me is once
- 17:30you get into agents and agentic workflows, you're actually
- 17:33creating a trail of behaviors, documents, log files, telemetry
- 17:40that is at a scale that is that far surpasses kind of the
- 17:44business intelligence. Like the operations logs of the
- 17:48Business Today are all gated on manual work flows.
- 17:51There's a human in every in every loop.
- 17:54Well, the more that you get to agents and automation and these
- 17:57automated work flows, you're actually creating more data and
- 18:00exhaust to analyze the behaviors of the agents themselves.
- 18:04So it's kind of a, you know, it's kind of a, a recursion,
- 18:07which is like the more agents you have, the more interactions
- 18:09you have, the more telemetry you have, which then you can analyse
- 18:12your behaviors and you can actually refine them.
- 18:14So to your point, Michael, it's an iterative cycle.
- 18:17Nothing works the first time. Nothing works the first time
- 18:20when we you know, from the most simple single flow agent to
- 18:23handling, you know, like an inbound issue in customer
- 18:26service to the most complex work flows of like trying to do
- 18:29financial reconciliation at the end of the quarter.
- 18:32You know, that has to pull off treasury and has to pull off of
- 18:34systems of record and has to be, you know, match SEC rules for
- 18:38disclosures. You know, that is you know, this
- 18:43volume, this this analysis and being very methodical about
- 18:47capturing what the agents are doing is both one of the
- 18:51greatest challenges and one of the greatest opportunities.
- 18:53Like debugging multi agent work flows right now is very very
- 18:56complicated. We have a very interesting
- 18:59question on LinkedIn from Stephanie Satsots who is an
- 19:04account executive with Work Day and she and she says our agents
- 19:10role or skills based, which is to say one agent equals one job
- 19:16to be done, which I think is so interesting.
- 19:20We'll give in the lengthy discussion you were just
- 19:24describing around multi agent technologies working together.
- 19:30What we see today are kind of 2 ends of the spectrum.
- 19:31On one end are kind of single purpose, single task agents that
- 19:36individuals themselves might build within their company to
- 19:41let's say you've got meetings coming up this week and you're
- 19:43like, you know what, I want to prep for these meetings.
- 19:45Well, today it's, you know, it's kind of a singular agent to
- 19:48multi data source and back through a singular agent where
- 19:51you say, let's get me prepped for these meetings.
- 19:53Could you tell me who I'm meeting with, what I ought to
- 19:55know about these folks? And so that single agent
- 19:58workflow is pretty easy to compose.
- 20:00As long as you have the connectors to, you know, or the
- 20:02data access and context access for like your calendar for, you
- 20:06know, a form of CRM for even e-mail, then you can gather the
- 20:11context and agents are and they are pretty good at returning.
- 20:14Hey, here are the people you're meeting with.
- 20:15Here's when you're meeting with them.
- 20:17Based on what interactions you had in the past, what we know
- 20:20about them, here's some things you might want to prep.
- 20:21So on one end spectrum, they're these kind of task specific,
- 20:25they're not really role specific.
- 20:26It's more task specific execution, you know, very
- 20:29quickly against known data sources.
- 20:32On the other end of the spectrum are these more what you're
- 20:36describing, which are these role based agents, which have many,
- 20:41many functions that occur within a role.
- 20:43It's just like if you're a financial analyst at a bank, you
- 20:46have multiple tasks and multiple jobs that you execute every
- 20:50single day. And once you start getting into
- 20:53that end of the spectrum, you're actually composing multiple
- 20:55agents to execute multiple tasks in concurrency or sometimes in
- 21:00the serial form. So one example of this is that
- 21:04for example, in at Harvey AI, it's a legal AI company right
- 21:10now. They have a is there's kind of
- 21:13they, they have like a role based AI.
- 21:14You can think about this as you're kind of like a paralegal
- 21:18where this paralegal AI goes out and it'll do, it'll analyze
- 21:22contracts, it'll do due diligence.
- 21:24It'll, you know, it'll do all of this work that before, like a, a
- 21:28paralegal, you know, might have to go and search through and
- 21:32index all this information and bring it back.
- 21:34The AI now can go and index that information, synthesize it,
- 21:37bring it back. And now a paralegal, you know,
- 21:40the human paralegal is more like steering and the AI paralegal is
- 21:44more like consolidating and analyzing.
- 21:46That's much more of a role based agent.
- 21:50Arsalan Khan on Twitter X, who's a regular listener, asks this.
- 21:55He says when it comes to AI agents, how many low hanging
- 22:02fruits should folks go after until they address, quote what
- 22:07he says the entire forest? And also who decides what is
- 22:12holistic AI agent deployment? It depends what your objectives
- 22:16are. The best advice I can give you
- 22:18is if you personally and or your organization aren't already in
- 22:23the building and construction of early agentic work flows, you
- 22:27should get started right away. There is.
- 22:31So that's capacity building. I've been working in advanced
- 22:35technology for a really long time and I can tell you it
- 22:37always takes longer than you think to get used to using a new
- 22:42technology. And so just to even start the
- 22:44learning curve just to get comfortable with the tools.
- 22:46You know, it's different in the past, maybe you work with AP is
- 22:49well, MCP servers and the flow of agents are a little bit
- 22:53different. You know, we even now have
- 22:54computer use which can bypass some of the, you know, need to
- 22:59construct APIs in front of data. And now you can just interact
- 23:01with, you know, the screen in front of you without having to,
- 23:05you know, drop a bunch of semantic ties between, you know,
- 23:08data sources and intent. So there's all these, there's
- 23:11all these emerging technologies that if, if you haven't like
- 23:14rolled up your literally rolled up your sleeves like I have, and
- 23:17if you're not, you know, experimenting, then you're kind
- 23:20of falling behind a little bit. So on that case, I would say,
- 23:24you know, I talked about Gemini Enterprise earlier.
- 23:25One of the great things about Gemini Enterprise Oregon, you
- 23:28know, these platforms is that you can go right to the user
- 23:31interface, the main page and you can start building an agent
- 23:34right away. You can and and like a Gemini
- 23:37Enterprise, I know because I was just building 1 this morning.
- 23:40You know, you can go in, you can specify what type of agent it
- 23:43is. You can specify the data you
- 23:44want it to connect to that you have access to based on your
- 23:47organizational policies and you can get up and started right
- 23:51away with building a So do that right away.
- 23:54Now the other question which is go tackle the, you know, the big
- 23:58nasty multi agent, multi workflow kind of those role
- 24:02based agents that Stephanie had referred to earlier.
- 24:06That's something that requires multiple stakeholders usually
- 24:09within an organization and usually comes from a top down
- 24:14business imperative or urgent business imperative.
- 24:17This could be, you know, we need efficiency in our software
- 24:21engineering or we need lift in revenue.
- 24:25And so we need more engagement, you know, at our on kind of like
- 24:29our storefront from a digital consumer journey in retail, for
- 24:33example. And so I would say depending on
- 24:36what your role is and the nature of your company and where you're
- 24:38at in the life cycle, you know, it might be a good idea to just
- 24:42start building and or you might be in the place where it's time
- 24:46to bring the finance department together, central IT together,
- 24:49the line of business and actually start to construct and
- 24:53build these multi agent workflows.
- 24:56But it's all going to be kind of like where you sit, where your
- 24:58organization sits in a period of time.
- 24:59Let's talk for a moment about the organizational challenges
- 25:05associated with multi agent orchestration.
- 25:11In a way it's it's kind of easier to talk about the
- 25:14technology. The technology is hard, but but
- 25:18people are harder. So tell us your thoughts about
- 25:22that. The most complex problems that
- 25:25we deal with when it comes to implementation of technology at
- 25:27scale are human and organizational issues.
- 25:30So number one, I mentioned this in this in the previous
- 25:33question, willingness to use. If you think about every great
- 25:38technology, wave, mobile, you know, big data, AI, cloud AI,
- 25:47they were all started both from like an organization's
- 25:51understanding that they need to do something differently.
- 25:52But usually it's by people within their own organization
- 25:55who are experimenting with new technology.
- 25:57You know, we, we call the shadow IT or, you know, kind of
- 26:00consumerization of IT. People discover these
- 26:02technologies, they see how it could be leveraged for their
- 26:04business and they're bringing it in from the bottom up.
- 26:07And so one of the things that, you know, I have been really
- 26:11appreciative of at least in like Google's approach to technology,
- 26:14is that we really encourage folks to experiment with new
- 26:17technology and we're always watching for where the hotspots
- 26:20show up. So for example, we figured out
- 26:23that AI and, you know, this kind of new generative AI wave, you
- 26:27could access it through the command line interface if
- 26:29there's a really, you know, a smart integration done.
- 26:32So Gemini CLI, yeah, it was this thing that we just released.
- 26:35It was a small team that worked on it and, you know, it exploded
- 26:39immediately. But it was our organization's
- 26:41willingness to allow a small number of engineers to try
- 26:44something and to, you know, allow it the space and time.
- 26:47Not a lot of resources, by the way, that's a misnomer.
- 26:50It's actually very scrappy, very bootstrappy.
- 26:52But just allowing and, and fostering and being willing to
- 26:57encourage experimentation is super, super important.
- 27:00Second organizational thing that I highly recommend is when
- 27:04you're getting into agents, again, they're going to take
- 27:06instructions and they need rules of engagement and they need
- 27:10explicit guardrails. And if you don't have standard
- 27:13operating procedures or decision frameworks within your
- 27:15organization for how you want to, you know, how you want to
- 27:18guide your own operations, you can't guide agents because they
- 27:22can't interpret intent very well.
- 27:25They're very explicit right now. And probably the last one is,
- 27:31you know, we'd mentioned this at the beginning in, in a kind of a
- 27:35different lens, but there are in every organization, there are
- 27:38hidden rules, agendas, norms. You know, it's important that to
- 27:43understand that if you ask software to do something, it
- 27:47will do it. If you don't have software to do
- 27:49something, it won't do it. And as you're constructing, you
- 27:52know, agents within a large organization, a lot of the, a
- 27:55lot of the, a lot of the outcomes that people are
- 27:59disappointed with, you can trace back to.
- 28:03They just assumed that would be able to make a logic leap that
- 28:06wasn't explicitly given to it. And that's a different way.
- 28:10You know, today, a lot of what the way that like managers and
- 28:13leaders and organizations work is they trust the humans that
- 28:16they hire and they put in these positions to bridge the gap
- 28:19between what is stated and what is unstated and what's needed
- 28:21for the business. And so being really aware of
- 28:25that and, and getting in that culture of experimentation will
- 28:27offer will often surface hidden rules and norms that the
- 28:31organization didn't even know that they had, because the
- 28:33software will break. You're the CTO of Google Cloud,
- 28:37and we want the AI to intuitively know and understand
- 28:46our implicit rules of engagement and our implicit culture, so we
- 28:53can just let the software do its thing and help us.
- 28:57Think about it this way. If you receive Gemini, for
- 29:01example, Gemini has been trained on, you know, an enormous amount
- 29:07of information that is accessible to everyone.
- 29:11So if you're looking for outcomes that are very specific
- 29:13to your industry and or your company, then it also requires
- 29:17pairing this AI with grounding or other data sources that will
- 29:21allow it to understand the unique context of your business,
- 29:24your industry, or your use case. There are no shortcuts.
- 29:27So you know, these frontier models need, and I think the
- 29:31biggest opportunity for organizations is to bring their
- 29:35data, their contacts, and their understanding partner with these
- 29:39frontier models. And it's amazing AI that's being
- 29:41provided and create something that is specific to their
- 29:45organization and their industry. And that's where differentiation
- 29:47and that's where competitive advantage lies.
- 29:49It doesn't lie in accessing the same model everybody has access
- 29:52to. It lies in combining the power
- 29:55of the incredible power of these, you know, AI models that
- 29:59have a general understanding of intent based on the data they've
- 30:02been trained on. But you might have very domain
- 30:04specific language within your industry.
- 30:06So for example, I spent a a bunch of time in manufacturing
- 30:10and industrials. You want to talk about a place
- 30:12that has its own language. You know, there are terms, very
- 30:15specific terms that mean very specific things.
- 30:17And AI isn't necessarily going to be trained on what those
- 30:21terms mean. And so it won't have the ability
- 30:22to bridge, you know, like that semantic layer.
- 30:27You're going from general interpretation of how of what a
- 30:31term might mean to a very domain specific term.
- 30:33This also shows up in code. A lot of organizations have
- 30:36built their proprietary software languages.
- 30:38You see this a lot in financial services and healthcare
- 30:43companies. And the AI wasn't trained on
- 30:48that data out-of-the-box. So as an organization, it's that
- 30:52combination of bringing your domain specific data and
- 30:56language and understanding intellectual property and
- 30:58combining that with this AI in a privacy Safeway that creates the
- 31:02real, the really big outcome. So, Michael, I hate to
- 31:04disappoint you, but you know, you it, it, it requires these
- 31:08organizations to do both to leverage that AI, but also, you
- 31:12know, using a platform like Gemini Enterprise, bring in and
- 31:16ground it in the realities and the specific knowledge of their
- 31:19industry. This is a question again from
- 31:21Anthony Scrifignano. And then we're going to move to
- 31:23some new folks. But he's asking about the role
- 31:28that active inference can play in all of this, to make agentic
- 31:35AI better, more useful, and so forth.
- 31:39The more inference and the more this kind of like this loop that
- 31:43needs to be created is going to be an opportunity for
- 31:46organizations to continuously improve the quality of their
- 31:49agents. And you know, hill climb on
- 31:53performance. And you know, one thing having
- 31:56worked in AI for a while too, is it's, it's a never ending
- 32:00journey. Like training AI isn't the end,
- 32:03it's the beginning. And having the sensibility about
- 32:06continuous improvement and continuous loops of of training
- 32:11and feedback are really important to making agents
- 32:15improve their outcomes over time.
- 32:17This is from Justin Kavanaugh and he's asking what are the
- 32:23most practical ways that small businesses can start integrating
- 32:27AI at the infrastructure or data level, not just to save time,
- 32:33but to future proof how they attract customers and compete
- 32:37with enterprise organizations so that they are not left behind.
- 32:42It's a really important question, actually.
- 32:44The most practical method is to just leverage the native
- 32:47capabilities of the cloud. So for example today in Big
- 32:50Query in Google Cloud, we have native AI integrations and
- 32:54native AI inference in the data system itself.
- 32:57And I think trying to bolt it on is a is a costly and and time
- 33:05intensive proposition. And so for the smaller to medium
- 33:08sized organizations, I would say leverage the native capabilities
- 33:11of the vendors and the cloud partners that you have.
- 33:15And in Google Cloud's case, you know, if you need intelligence
- 33:20baked in to the data systems, you know that already comes
- 33:23standard and what we provide. I just have to amplify Will a
- 33:28comment that that you made earlier, which is just get
- 33:34started. If you're a small business, the
- 33:38more you can gain familiarity with the kind of services, for
- 33:41example, that Will was just describing, the better off
- 33:45you're going to be. And you'll then you'll learn and
- 33:47you'll know how to how to take those tools and apply them to
- 33:52your specific business. And so much now like in our
- 33:56platform, for example, AI is everywhere and embedded into the
- 34:00services themselves at every core component.
- 34:03So it doesn't matter whether you know you're talking about a
- 34:05storage system, a database tier. If you're talking about even
- 34:09like compute and how we optimize compute for specific jobs, you
- 34:13should really just leverage the cloud providers integrations
- 34:17natively of AI into these into every tier that might be
- 34:21supporting business applications, your website and
- 34:23what have you. We have a question on LinkedIn
- 34:26from Kenroy Benedict, who says, do you see an increase in public
- 34:32services using AI outside of simple chat bots?
- 34:36And I'm not sure whether he means public cloud services or
- 34:41in the public. Sector so as yeah, the founder
- 34:44of Google public sector and a board member still, which I'm
- 34:48very proud of the explosion of AI in in public services is
- 34:52starting to happen. And I'll give you you know, one
- 34:55example that is I think really, really cool and that is one of
- 35:01the one of the most important functions that public sector
- 35:03organizations fulfill is that they provide help when people
- 35:07need it. And so, for example,
- 35:09unemployment benefits at the state level can often be or at
- 35:15the end or at the federal level, but unemployment benefits are a
- 35:18really big deal because this is a person's most vulnerable time.
- 35:21And one of the things that AI has enabled states to do with
- 35:26the state of Wisconsin and others is instead of having
- 35:30someone submit an application for benefits and going back and
- 35:32forth with a bunch of paper and that taking, you know,
- 35:34potentially months, meanwhile, this person is suffering.
- 35:37They don't have access to the resources they need.
- 35:39It's a very, it can be a very, very significantly, you know,
- 35:44painful time for an individual and their families.
- 35:47AI has enabled interactions and the processing of claims, for
- 35:54example, to go from, you know, weeks to hours or days through,
- 36:00you know, the initial submission of information to AI, triaging
- 36:04the severity inherently of the cases, pushing, you know, top
- 36:08cases to the top, providing initial recommendations.
- 36:12And, you know, all of those kind of automation speed up steps are
- 36:17the difference between, you know, thriving and, you know, at
- 36:22least getting back on your feet or not.
- 36:25And we're going to, you're going to see a lot more in public
- 36:28sector services at the federal, state and even internationally.
- 36:32Greg Walters on LinkedIn says he sees a world where all
- 36:37applications and functions are contained within the LLM and the
- 36:41AI. What say you?
- 36:43And I'm going to ask you to answer that pretty quickly.
- 36:46Wow, that's my answer. I, I love the vision.
- 36:50I do think that one of the more exciting capabilities of these
- 36:54models is their ability to provide tailored user interfaces
- 37:00like we call them like ephemeral apps or ephemeral UI.
- 37:04And that means that the AI becomes the UI and that the
- 37:07application UI is less important.
- 37:09And so in that way, I do agree with the vision that so for
- 37:13example, if you have access, if the LLM has access to the data,
- 37:15has access to the user intent, we've seen behaviors where AI
- 37:20now can spin up the user interface ad hoc and create what
- 37:23is essentially an ephemeral AI app immediately.
- 37:26We have crazy question, a good question from Arsalan Khan on
- 37:30Twitter X he says, and very quickly, please.
- 37:34He says these AI agents are like soldiers who will follow orders.
- 37:39Who will be the generals, colonels or even military police
- 37:44to provide leadership and guidance to these soldiers?
- 37:48Should that be high level AI agents too?
- 37:51Well, it will depend on the stakes of the workflow that
- 37:54you're dealing with. So for example, if you're
- 37:55dealing with Transportation Safety, those are always going
- 37:57to be human in the loop, human monitored workflows.
- 38:02But if you're dealing with, I'm trying to create 3000 short form
- 38:05videos for certain brand outcomes to go create an
- 38:09advertising campaign, you don't necessarily need to have a human
- 38:12in the loop because of the stakes of getting something
- 38:15wrong. So think about it in terms of
- 38:17stakes and what happens if you get something wrong.
- 38:19And you can probably back into which work flows are going to be
- 38:21more human in the loop and which are going to be more agent
- 38:23driven. Let's talk about deployment of
- 38:28agents in the in the enterprise. Can you describe the common
- 38:34patterns of successful agent deployments and AI in general,
- 38:39and again relatively quickly please?
- 38:41#1 success criteria is picking a problem that is actually likely
- 38:47to be somewhat solved by agents and higher forms of automation
- 38:53and faster task task execution. So a good example of this would
- 38:56be in financial services. You know, we're working with the
- 39:01bank right now that has seen the that has speed up their research
- 39:06function. So now they can go cascade
- 39:08across all the documents and information that they have and
- 39:11bring back to analysts and customer relationship managers
- 39:14synthesize pieces of research in a matter of hours or minutes
- 39:19where it used to take multiple days or a week that research to
- 39:22come back. And so that was a problem.
- 39:24They knew that they had the data, it was very meaningful to
- 39:28their customers, which is number 2.
- 39:30And so specific problem had the data meaningful to their
- 39:33customers and it was something that would get better over time
- 39:37iteratively. So those are kind of key
- 39:39components that we see over and over again.
- 39:41What are the most significant challenges that you see as
- 39:46organizations are trying to deploy agents?
- 39:49Where do they run into trouble? Lack of data, lack of context to
- 39:53the agents in the models is the number one trapdoor.
- 39:58A second one is realistic expectations in the early going.
- 40:05These are iterative loops and many organizations like to like
- 40:10to set up projects so that they're, you know, if they don't
- 40:14get it, they're used to being really like experts in their
- 40:17field. You know, you can think of like
- 40:18aerospace, very precision oriented manufacturing.
- 40:23It's important to have a culture where you understand that an
- 40:26agentic software development, the first few iterations are
- 40:30probably not going to have a quality or an efficiency that
- 40:33you really love, but it's in the iteration and the fast iteration
- 40:36that you get the results that you want.
- 40:37And so that's probably the second trap door is the way that
- 40:41we construct projects for success out-of-the-box in many
- 40:44cases is exactly the opposite of the way that these projects work
- 40:48and will make agent development projects successful.
- 40:50I will also say the third really important factor Michael, and
- 40:54this goes across whether it's agents, AI or G or even
- 40:58technology projects in general is leadership modelling the
- 41:03importance and get getting involved in this.
- 41:06The leaders that and I'm talking about CEOC suite senior leaders
- 41:10rolling their sleeves up. It was a, it was a very
- 41:13important moment to all of us at Google and I think, you know,
- 41:15signal to the market when, you know, our CEO Sundar was asked,
- 41:20you know, hey, you know, what are you doing with AI?
- 41:22He's like, I'm vibe coding and he went into a long exposition
- 41:25of what he's actually doing with AI.
- 41:27What that does is it says, I'm doing this, we're all in this
- 41:30together. I, I'm committed to this.
- 41:32We're committed to this. And I can't, I can't emphasize
- 41:35enough how important the leader modeling the support of this
- 41:39exploration, this kind of new wave of exploration, how
- 41:41important that is in every in every way to agent development.
- 41:45But you're really forcing organization, business leaders
- 41:49therefore to become technologists and the CEO or
- 41:56CFO. Obviously Google's a special
- 41:58case because you're developing these products, but the average,
- 42:01you know, manufacturing CEOCFO, is it really realistic to expect
- 42:07them to get into the guts of? Absolutely, yeah, absolutely.
- 42:12And they all want to and see, that's the great part is my job
- 42:16is to make it so the technology is mostly invisible.
- 42:19So instead what they see is a platform, you know, like again,
- 42:23Gemini Enterprise where they can go and they can just in plain
- 42:26language scope an agent, give IT system instructions, give it
- 42:30initial prompts, click button on data sources and build an agent.
- 42:36Now, it doesn't mean that CE OS necessarily all CE OS should be
- 42:39out there building multi agent complex work flows for, you
- 42:43know, highly specific tasks. But it's my firm belief that
- 42:47they're all capable of participating materially in this
- 42:51technology wave. And that's one of the things
- 42:53that makes it so powerful and so ubiquitous, Michael, is that AI
- 42:56has gone from a department or a very like weird and mystical
- 43:00technology to part of everyday life.
- 43:03If you type on a keyboard, a virtual keyboard on your phone,
- 43:07you've got AI under the hood that's anticipating the words
- 43:09that you're trying to type and trying to serve you up a word.
- 43:12Any search that you run every single day, it's all around you.
- 43:16And for the first time in my career, the platforms to
- 43:21actually deploy advanced AI are accessible to everyone.
- 43:24Can you talk about nano banana? Explain what nano banana is?
- 43:30And I have to say, I just think it's incredible.
- 43:34So please just just briefly tell us about Nano Banana.
- 43:38So Nano Banana is our image generation model, latest image
- 43:41generation model. And this is a big theme,
- 43:44Michael, and that is that multimodal AI is the future for
- 43:48a couple reasons. Number one, if one picture is
- 43:51worth 1000 words, a video is worth a million pictures.
- 43:54And as humans, how we experience the world is through our senses,
- 43:58sight, speech, sound and AI. Now these multimodal AI models
- 44:06like Nano Banana can take images as inputs along with, you know,
- 44:12prompts and it can create new things.
- 44:15You know, if you if you take Nano Banana and the ability to,
- 44:18you know, create images and you extend it, we're able to create
- 44:23videos as well. VO is a video creation model
- 44:25that we have under the hood of multimodal model extended.
- 44:28Even more, we have a Gemini Live, the API, and I mentioned
- 44:33that that use case earlier of being able to take your phone,
- 44:36flip it around, show it something, show the AI
- 44:38something, and it knows inherently that that's a cable
- 44:41modem. This is the model and you can
- 44:43start your debugging right away. That's nano Banana is a glimpse
- 44:49into the multimodal future of AI and that's how humans want to
- 44:54interact with AI. They don't want to have to type
- 44:56things all the time. Sometimes they just want to show
- 44:57you something. So you know, hey, here's what
- 45:00I'm looking at, you know, give me some background on, you know
- 45:05this and it can explain mathematical equations, it can
- 45:08explain geolocations. That is, that's the future of a
- 45:12is is multimodal immersive experience and nano bananas one
- 45:15form. Elizabeth Shaw says how can you
- 45:18tell what's real from the hype with respect to AI?
- 45:24Then how can you turn that into something that's practical?
- 45:29Measure, measure, measure and be very transparent.
- 45:32I even have to coach my own team on this.
- 45:34When you're building something, the natural human tendency is to
- 45:38show all the best parts of it and to talk as though you know
- 45:42everything is great in our team. And what I would encourage for
- 45:46all of you that are working with AI and especially building
- 45:48agents right now is it's possible, but it's not always
- 45:53easy. And it's important to show the
- 45:55initial is, is to gain the credibility and be very
- 45:58realistic about how it's going and measure incrementally.
- 46:02We break, we break long agentic development workflows into very
- 46:06incremental steps. And we measure and we talk about
- 46:09what we need to do to hill climb on performance and or efficiency
- 46:12at every single step rather than saying, oh, it's going to be
- 46:15able to do all this great stuff. We start with very atomic, very
- 46:19specific and very measurable outcomes we're looking for, and
- 46:22we make a decision at each one of those gates about whether to
- 46:24proceed, whether to change or whether to stop.
- 46:27And that type of transparency in every step of development and
- 46:30that very sober look at what it can and can't do will help bust
- 46:34the hype cycle. Also, for those of you that are
- 46:35leading projects, highly encourage you not to make big
- 46:39statements about what this about what AI is going to do and
- 46:43instead think very soberly about one of the biggest problems that
- 46:47your organization is facing. And how might automation speed
- 46:52multi modality? How might those things combine
- 46:55to chip away at solving? You know what are always going
- 46:58to be some really complex problems that aren't you know
- 47:00you're not going to solve through magic AI wand.
- 47:03Very, very good advice. You're, you're saying
- 47:06essentially approach it from a very practical standpoint about
- 47:10what problems you need to solve rather than AI is great and
- 47:15we'll change our lives and solve everything in the world.
- 47:18Yeah. And I will also say that in the
- 47:20history of technology, we always underestimate how profound the
- 47:23changes are going to be and we overestimate how fast they're
- 47:26going to happen. And I think that also creates
- 47:28some of this hype cycle that you're referring to, which is
- 47:32now we can see the potential, but that potential will take
- 47:36years to realize. To the point earlier, Michael,
- 47:40if you don't get started now, you don't climb the learning
- 47:43curve and you don't create the incremental successes that
- 47:45eventually lead to the breakthroughs.
- 47:47And you're always watching somebody else do it.
- 47:48Where is this all going in the next 6 to 18 months?
- 47:53What can we expect from your vantage point, from what you
- 47:55know right now? First off, the models will
- 47:58continue to improve in capability, in quality and
- 48:03you'll see more multi native multimodal capabilities.
- 48:06Just like today, you know where you can go to Gemini and you can
- 48:12interact through voice, you can interact through video, you can
- 48:17interact through images. That multi modality is going to
- 48:20get more and more seamless and it's and it's going to feel more
- 48:22natural to interact with AI. Second, a continuous improvement
- 48:27around the out-of-the-box that are available to help people get
- 48:32started and Gemini Enterprise more and more connectors to data
- 48:35sources. So you know, for those of you
- 48:37that have, you know, SharePoint or JIRA or Confluent or you
- 48:41know, all these different, all of those connectors and many,
- 48:43many more are coming. More improvement around
- 48:47governance and security and especially in where value is
- 48:51transacted. So today, many of the agent
- 48:53workflows that I see, they're not commerce transactions, but
- 48:57we just released the agent payment protocol AP2 as part of
- 49:01this kind of standardization stack.
- 49:03So MCP A2A, which is the agent to agent protocol and now the
- 49:08agent payment protocol, because we're going to want agents to
- 49:12fulfill tasks that include commerce.
- 49:15And that is going to be a significant area of focus is
- 49:18making sure that that's done well, it's done safely, it's
- 49:21done securely. And AP2 Agent payment protocol
- 49:26is the first step in that journey.
- 49:27And with that, I'm afraid we're out of time.
- 49:32Will Grannis, Chief Technology Officer of Google Cloud.
- 49:34Thank you so much for being with us, and I hope you'll come back
- 49:37another time. Yeah, you got it, Michael.
- 49:39It was a real pleasure. Thanks for the questions
- 49:41everybody. And everybody, thank you for
- 49:44those great questions. Before you go, I want you to
- 49:47write this second. Go to cxotalk.com, subscribe to
- 49:51our newsletter. We have truly great shows that
- 49:54are coming up and we want you to join us.
- 49:57Everybody, I hope you have a great day.
- 49:59Thank you guys. Are you guys ask the most
- 50:02amazing questions. You guys are awesome.
- 50:03Thank you so much. Take care, everyone.
- 50:06We'll see you soon.