[00:13] What a morning. We've had an incredible
[00:16] set of announcements for Gemini
[00:18] Enterprise today, including the launch
[00:21] of the Gemini Enterprise Agent platform
[00:24] and more on how Gemini Enterprise works
[00:26] with Workspace in the spotlight session
[00:28] immediately following Thomas's keynote.
[00:31] Thank you so much for joining late in
[00:33] the day and after a packed first day at
[00:36] Next. Uh in this session we'll share
[00:38] more details on what's new for Gemini
[00:41] Enterprise, in particular the
[00:43] application level solution that allows
[00:46] employees to leverage AI and agents
[00:48] throughout the enterprise.
[00:50] I'll grab a clicker since I didn't grab
[00:51] that.
[00:53] I'll go to the next slide. Uh my name's
[00:54] Jamie Deggeer. Uh I'm a product leader
[00:57] on Google Cloud AI team working on
[01:00] Gemini Enterprise, our agent platform,
[01:02] developer tools.
[01:04] And I'm joined by an esteemed panel of
[01:07] uh colleagues today, including Dave
[01:09] Teggart
[01:11] who's our engineering leader for Gemini
[01:13] Enterprise
[01:14] and two of our most innovative customers
[01:17] that are transforming how their
[01:18] organizations work with Gemini
[01:21] Enterprise.
[01:22] Aaron Purcell is the managing director
[01:24] of digital innovation at KPMG
[01:28] and Lisa Nabb is the head of agentic use
[01:31] case management at one of Germany's
[01:33] largest insurance companies, Signal
[01:35] Iduna.
[01:38] So, let's dive in starting with Gemini
[01:40] Enterprise.
[01:42] We launched Gemini Enterprise in October
[01:44] and have seen tremendous adoption across
[01:47] industries.
[01:49] Gemini Enterprise is that front door
[01:51] to AI in the workplace.
[01:54] It gives every employee the ability to
[01:56] leverage AI and agents connected to all
[01:59] of your enterprise data
[02:01] and with the ability to take actions in
[02:03] all of your enterprise systems.
[02:07] Importantly, from day one Gemini
[02:09] Enterprise has been built to be cloud
[02:11] first and enterprise grade. And today we
[02:14] announced the underlying platform for
[02:16] that with the Gemini Enterprise Agent
[02:18] platform and a bunch of expansions to
[02:21] how we do that and govern that across
[02:23] your organization.
[02:26] Gemini Enterprise allows
[02:28] your employees [clears throat] to
[02:29] leverage a whole team of agents
[02:31] to collaborate with other employees and
[02:34] with agents in a persistent team memory
[02:37] space
[02:38] and it enables every employee to build
[02:41] their own agents through no code.
[02:46] And the impact of this is real.
[02:49] We have customers like Anaplan that have
[02:52] rolled this out across the organization
[02:54] with 80% of employees using it
[02:57] building thousands of no code agents.
[03:00] Macquarie Bank in Australia who is
[03:03] transforming to become an AI driven bank
[03:06] um
[03:08] literally envisioning a future where
[03:11] everything can the whole bank can
[03:12] operate through AI
[03:14] and they've been able to drive a 40%
[03:16] increase and help center self-service.
[03:19] They're using agents to help with fraud
[03:21] prevention, personalized recommendations
[03:23] and more.
[03:25] And both Aaron and Lisa will talk more
[03:27] about how they're using it at their
[03:29] organizations in a few moments.
[03:33] Now, importantly, as I said, this is all
[03:35] built to be cloud
[03:37] first and enterprise grade. And the
[03:40] enabler for that is the Gemini
[03:42] Enterprise Agent platform.
[03:45] The way you can think about this is that
[03:47] the agent platform brings a few things.
[03:50] It brings all of Vertex AI with the
[03:52] ability to leverage
[03:54] all the Google models, models like
[03:56] Anthropic Claude, open source models and
[03:59] more for your agents across your
[04:01] organization.
[04:03] You can connect those with all the
[04:04] tooling you need to ground them with
[04:06] data, connect them to systems, etc.
[04:09] And then it adds a set of new
[04:11] capabilities just launched today
[04:14] for governing those at scale, including
[04:17] an agent gateway, a new agent identity
[04:20] service that's generally available
[04:23] agent registry and tools and skills
[04:25] registries
[04:27] and more that come from other parts of
[04:29] Google Cloud and are managed and made
[04:31] available through this one agent
[04:33] platform for you.
[04:35] And it combines some new innovations
[04:36] coming from our DeepMind team helping
[04:39] with how you achieve quality at scale
[04:41] with those agents and improve quality
[04:43] over time.
[04:46] But as the app on top of that, you can
[04:49] think of it taking those building blocks
[04:51] and putting them to work.
[04:53] So, registry's great if it says that
[04:55] Jamie has access to 100 agents.
[04:58] But how as an employee do I see those
[05:00] agents?
[05:01] How do I share them with my team
[05:03] members?
[05:05] And how do we connect this to all of
[05:06] your enterprise systems?
[05:08] And this is the what Gemini Enterprise
[05:10] app brings on top of this with built-in
[05:12] connectors for over 50 first-party
[05:14] integrations the ability to add a
[05:16] connector from any MCP server
[05:20] um personalized ranking for every
[05:22] employee of how we bring the best data
[05:24] to every interaction with AI or an agent
[05:28] um and front-end surfaces to use this
[05:31] through the Gemini Enterprise web
[05:33] interface
[05:34] mobile apps and third-party surfaces
[05:37] like Slack, Workspace and more.
[05:43] So, some of the new capabilities
[05:45] coming to Gemini Enterprise at the
[05:47] application layer this week include a
[05:49] number of updates across how you work,
[05:51] how we connect to data and then manage
[05:54] and operate that.
[05:56] Um I'm going to go through a few
[05:57] highlights of those next. Um but we have
[06:01] a lot more and
[06:03] um happy to kind of cover those in in
[06:05] follow-ups and additional materials for
[06:07] you following the event.
[06:09] The first is our no code agent designer.
[06:13] This is really important because what we
[06:15] hear from organizations is they want two
[06:17] things. They're kind of almost the yin
[06:19] and the yang.
[06:20] They want the ability to have agentic
[06:22] reasoning. They want to leverage AI to
[06:24] do long-running, complicated tasks.
[06:28] To do that
[06:29] you need a full agent that can leverage
[06:32] a sandbox environment, write their own
[06:34] code, create their own tools on the fly,
[06:36] connect to your systems, evaluate their
[06:38] own work and iterate in the sandbox
[06:40] environment to achieve a high-quality
[06:43] outcome.
[06:45] So, that's one. That's sort of
[06:46] long-running agents. But two, they also
[06:49] want the ability to put the outcomes of
[06:51] that on guardrails in specific
[06:54] operational policies
[06:57] that are different for different
[06:58] employees, different teams and different
[07:00] scenarios.
[07:02] And so with Gemini Enterprise and the
[07:03] agent designer we allow you to do just
[07:05] that. You can combine
[07:07] long-running agents with a no code
[07:10] editor that allows you to create those
[07:12] guardrails and policies that the agent
[07:14] operates within to have deterministic
[07:16] outcomes for what you're doing with
[07:18] those agents.
[07:20] And I'll show an example of a
[07:21] long-running agent
[07:23] in my demo in a moment but these are
[07:25] agents that can um
[07:27] as I mentioned have a long-running
[07:29] [clears throat] file system, carry on
[07:31] for days of work or um maybe even run
[07:34] forever across sessions with shared
[07:37] context
[07:38] execute code in a secure sandbox
[07:42] leverage a computer use like a browser
[07:44] or other application
[07:46] but do so in secure sandboxes that are
[07:49] in the cloud
[07:50] and are governed by your secure policy
[07:52] security policies and try instead of
[07:54] trying to manage a world in which agents
[07:57] can do that compilation of code and
[07:59] creation of tools and access all your
[08:01] systems on each of your employees'
[08:04] desktops or laptops um which is much
[08:07] harder to secure.
[08:12] With both of those features, I think
[08:14] what's really exciting is that the way
[08:16] that every employee works with AI
[08:18] shifts.
[08:19] It shifts from a world in which
[08:22] employees go to an agent or to the AI
[08:24] assistant and chat with it in a
[08:26] synchronous fashion to complete a task
[08:29] to a world in which you have a team of
[08:31] agents operating in the background for
[08:33] you
[08:34] and they only need to come to you when
[08:36] they need input from you or when they
[08:38] have results and updates for you.
[08:41] And to give one central place to manage
[08:43] all of those, we're introducing inbox
[08:46] which gives you a view of
[08:48] all of your agents operating in the
[08:49] background and the ability to manage
[08:51] notifications and respond with updates
[08:53] to each of them.
[08:55] Importantly, each employee can set up
[08:57] their own preferences
[08:59] for how they push those notifications to
[09:01] the mobile apps
[09:03] to chat or Slack
[09:06] or email
[09:07] or other clients and continue to
[09:10] interact with those agents from those
[09:12] clients as well and and manage through
[09:14] multiple surfaces.
[09:17] We're also adding support for skills.
[09:18] So, every employee can build their own
[09:21] skills to operationalize procedures that
[09:23] they want to use regularly and leverage
[09:26] those with their agents.
[09:28] And Canvas, so you can create documents,
[09:31] whether it be presentations, documents
[09:34] or other things that you want to then
[09:36] share and open inside of Workspace or
[09:39] Microsoft PowerPoint, Word and Microsoft
[09:42] 365.
[09:45] And we're announcing projects.
[09:49] Projects allow employees to work in one
[09:51] shared space with a fixed set of context
[09:54] and data and a shared conversation
[09:56] history
[09:57] that's long-running and kept over time.
[10:00] And they can not only collaborate with
[10:02] their human employee co-workers, but
[10:05] also with their digital agent co-workers
[10:07] in this space.
[10:10] So, that's a quick overview of a few of
[10:11] the highlights coming. Next, let me give
[10:14] you a demo of what some of this looks
[10:16] like
[10:17] in a real scenario. So, if we switch to
[10:20] the demo,
[10:22] um
[10:23] let's imagine that
[10:27] we're a bank,
[10:28] um and this bank wants to automate loan
[10:31] processing. I'm told by some of our
[10:33] customers loan processing is like the
[10:35] most expensive part of the bank. There's
[10:37] like thousands or tens of thousands of
[10:38] loan officers in some cases. And
[10:41] processing a loan might take weeks or a
[10:43] month to go back and forth with a
[10:45] consumer to get the right information,
[10:48] go through the compliance checks, and
[10:50] other regulatory steps.
[10:53] So, to do this, the first thing that
[10:54] this bank has done is to create a
[10:57] long-running agent. They wrote this
[10:59] agent in full code
[11:01] with the ability to leverage a sandbox
[11:02] and more.
[11:04] And as an employee, I can use that agent
[11:06] by coming into Gemini Enterprise and
[11:08] asking, "Give me the latest 10
[11:10] unprocessed mortgage loan applications."
[11:13] Now, importantly, I didn't have to go
[11:14] and browse the agent gallery to find the
[11:17] right agent here. Gemini Enterprise
[11:19] understands which agents are available
[11:21] to me as an employee and understands the
[11:23] query and says the the best agent for
[11:26] this is the loan supervisor.
[11:28] And I can say, "Confirm. Yes, I'd like
[11:30] to ask that agent."
[11:32] Now, as this runs, it'll come back with
[11:34] a queue of um the loans for me, and I'll
[11:38] go ahead and ask to process the first
[11:39] one.
[11:42] Now, at this point, this is kind of a
[11:43] long-running task. This might take
[11:45] hours or or a day or other longer amount
[11:48] of time.
[11:49] So, next, let me show you what this
[11:51] looks like for the administrative team
[11:53] if we go under the hood.
[11:55] So, the administrative team inside of
[11:57] the GCP console
[11:59] can see and manage
[12:01] the agent the agent runtime. They can
[12:03] see the session for this. They see this
[12:05] request for the processing the latest
[12:06] loan application. They can see
[12:09] what the agent does under the hood. It
[12:11] calls document retriever, it validates
[12:14] documents, it checks the user's credit,
[12:16] has a risk assessment agent, compliance
[12:19] agent, etc.
[12:21] And when that comes back, if I go to the
[12:22] next tab where I ran this in advance, in
[12:24] this case, it comes back and says that
[12:26] this is flagged for human review. So,
[12:28] this is great. I'm using this
[12:30] long-running agent through Gemini
[12:31] Enterprise.
[12:33] But now, what I really want to do is
[12:34] automate the running of this inside of
[12:36] my procedures. Let's imagine that I'm
[12:39] the loan operations manager,
[12:42] I might want to put this into a set of
[12:44] policies that we run as an organization.
[12:47] So, to do that, I'm going to go and I'm
[12:49] going to create a no-code agent as that
[12:51] loan officer.
[12:53] And I've done this in advance, but just
[12:55] to see how this works,
[12:56] as a knowledge worker, I could come in
[12:58] and say, "Okay, I want this to run on a
[13:00] scheduled trigger daily,
[13:03] grab all the latest loan applications,
[13:06] and then call that full code agent with
[13:08] the list of applications."
[13:10] And from there, I want to put this into
[13:12] a set of conditionals. And I can
[13:14] actually refresh this.
[13:17] >> [clears throat]
[13:17] >> I can run and test this as an employee
[13:21] inside of this environment and see how
[13:23] it flows through each step, what
[13:25] actually happens in the logic.
[13:27] And with the conditionals I've set up,
[13:29] if it flags for human review,
[13:31] the loan officer would get a
[13:33] notification through Gemini Enterprise,
[13:35] and they could approve or deny that. And
[13:37] we'll actually surface if they're
[13:39] approving or denying that as a quality
[13:41] signal to you as the IT team, and in the
[13:44] future, over time, use that to help
[13:45] self-improve how these agents run at
[13:48] scale.
[13:49] Now, in this case, this flows through to
[13:52] a human review step, and uh it will
[13:55] create an incident inside of ServiceNow,
[13:58] and then email me a notification for
[13:59] that. For the sake of time, I ran that
[14:02] in advance in the next tab. You can see
[14:04] how it came through and created that. If
[14:06] I go to my email,
[14:08] I can check
[14:10] and open directly into ServiceNow the
[14:12] issue that was created. So, what did we
[14:14] see here?
[14:15] We created a full code agent, that's
[14:18] long-running agent, that works with the
[14:20] sandbox to create its own tools, write
[14:22] code, and evaluate its work and iterate
[14:24] until high-quality outcomes are
[14:26] achieved.
[14:27] But then we put that inside of a set of
[14:29] operational guardrails with a no-code
[14:31] agent that kind of makes this real for a
[14:34] real enterprise environment.
[14:36] Okay.
[14:37] Next, let's say that I would like to
[14:40] um report on this. And so, next, I'll
[14:43] ask,
[14:46] "Let's summarize all the loan
[14:47] applications from the latest batch."
[14:49] And this is actually functionality we
[14:51] built into the loan supervisor agent,
[14:54] where it will actually come back with
[14:55] some uh summary information for me, um
[14:59] which looks really great here inside of
[15:00] Gemini Enterprise. It even creates an
[15:02] HTML report, so I could open that HTML
[15:05] report, and this is something I could
[15:06] copy into an email and send to my
[15:08] leadership.
[15:10] But now, let's say that I also want to
[15:12] create uh a slide deck with this report.
[15:16] Um that takes a couple of minutes, but
[15:18] we can showcase the capability in the
[15:20] next tab to
[15:22] um [clears throat]
[15:24] open that slide deck that's created
[15:26] right inside of Gemini Enterprise
[15:28] through Canvas.
[15:31] So, here I have a fully interactive
[15:33] environment where I can edit my slides,
[15:35] and I can share them to my team through
[15:36] Workspace or open directly into the
[15:38] Workspace app.
[15:40] And going further, we also want to
[15:42] extend the ability to call these agents
[15:45] from across Workspace and other surfaces
[15:47] like Slack.
[15:49] And so, here, if my manager has asked me
[15:51] for the latest stats, I can go ahead
[15:54] I'll go into this window so you can see
[15:55] it better.
[15:56] Um I can go ahead and at-mention the
[15:58] loan supervisor agent. That's an agent
[16:01] from Gemini Enterprise.
[16:03] And I can ask it to
[16:06] uh
[16:07] summarize the latest loan applications
[16:09] processed in the latest batch, and it
[16:11] will respond to myself and my manager in
[16:13] this chat with all that information, so
[16:15] my manager can directly open that HTML
[16:17] report themselves.
[16:20] So,
[16:21] with that, I mentioned how all of that
[16:23] comes together to me in a new way of
[16:25] working with AI. Employees now have a
[16:27] team of agents that operate in the
[16:29] background. And so, the upcoming update
[16:31] to Gemini Enterprise is this. We're all
[16:34] seeing right inside of my main window
[16:36] updates that need my review. I can click
[16:38] to view those in my inbox
[16:41] and filter to say all of the
[16:44] agentic tasks that are completed, those
[16:47] that are currently running in progress,
[16:49] and more. And so, that's a quick look at
[16:51] some of what's coming to Gemini
[16:52] Enterprise over the coming announcements
[16:55] this week. Um next, I'm going to invite
[16:58] our head of engineering on stage
[17:00] on stage, Dave Tagare, who will walk us
[17:02] through the underlying platform and how
[17:04] we build this for quality. We can switch
[17:06] back to slides. Dave.
[17:16] All right, everyone.
[17:19] Firstly, thank you for coming, and we
[17:21] appreciate your support for Gemini
[17:22] Enterprise. Thank you, Jamie. Wonderful
[17:25] demo there.
[17:26] Uh and actually not a demo, right? Like
[17:27] we we're using some of these things or
[17:29] most of these things.
[17:31] Everyone, I'm Dave Tagare. I lead
[17:33] engineering for Gemini Enterprise. Um
[17:36] so, Jamie showed us an awesome demo.
[17:38] Let's uh you know, take a sneak peek
[17:40] under the hood and see what's actually
[17:42] happening, right? Uh so, here I have
[17:45] like what was the architecture as of
[17:47] this week, and what will be the
[17:49] architecture in in like the very near
[17:51] future. So, as of the past week, we
[17:54] actually switched to what I what we call
[17:56] a flattened planner plus skills
[17:58] architecture. Essentially, we moved away
[18:00] from the agent to sub-agent model into a
[18:02] more the planner is just a dispatcher,
[18:05] everything to it in the ecosystem is a
[18:07] tool,
[18:08] uh where you the planner gets a query
[18:11] and decomposes it into many, many
[18:12] fan-out queries, and sees the world as a
[18:14] tool. The tools execute uh
[18:17] Additionally, we have added skills
[18:19] support. It's already rolled out, uh
[18:21] wherein we are centralizing all the
[18:23] skills, so you every agent understands
[18:25] its own capabilities, uh which help
[18:27] prevent errors and hallucinations. Post
[18:30] roll-out, what we've noticed is that
[18:31] this architecture is actually giving
[18:33] about uh 40%ish drop in latencies at P95
[18:37] already. So, moving forward, how's the
[18:39] how's this changing, right? Uh moving
[18:41] forward, we're actually
[18:43] going to enable something called the
[18:45] agent mode. Uh this mode will support
[18:47] long-running agents with CLI plus
[18:49] dynamic skills, uh all the good stuff
[18:51] that enable you to run, you know, agents
[18:53] from like minutes or 20, 30-minute grain
[18:56] today to many weeks at the same time. Uh
[18:59] additionally, the architecture is also
[19:00] going to support uh universal context
[19:02] for longer-term memories, checkpointed
[19:05] state, because these long-running
[19:06] agents,
[19:07] let's say over a span of 2 weeks, 3
[19:09] weeks, 4 weeks, will have disaggregated
[19:11] sessions, so you need to know where to
[19:12] resume from, what your state was at that
[19:14] point, keep context with between
[19:17] different turns or sessions of this uh
[19:19] specific agent type, and then take it up
[19:21] from there.
[19:22] We're also going to roll out heartbeats
[19:24] and event and time-based triggers that
[19:26] enable, you know, pretty complicated
[19:28] functionality. The triggers can also be
[19:30] modeled along the lines of like your own
[19:32] custom policies and and feature sets. Um
[19:35] You can do this today with uh high-code
[19:37] agents on ADK 2.0, and these will soon
[19:40] be available out of the box on Gemini
[19:42] Enterprise in like in Q2 itself.
[19:46] Um
[19:46] >> [clears throat]
[19:47] >> Taking a look further into, you know,
[19:50] governance is a cornerstone of why we
[19:53] have Gemini Enterprise to begin with.
[19:54] So, to scale AI securely, we're bringing
[19:57] in enterprise-grade governance to your
[19:59] digital workforce. We're doing this
[20:01] through investments in five key areas
[20:03] that you see listed here.
[20:05] First is agent observability and
[20:06] traceability. We provide you with
[20:08] detailed logs and traces to see exactly
[20:11] what an which agents are being involved,
[20:13] what are they doing, what actions they
[20:15] are taking, and how your data is being
[20:17] processed, handled, all within a single
[20:19] console in GCP. Secondly, it's the agent
[20:22] registry. You know, in practice, you
[20:24] will have many, many agents
[20:26] proliferating through your workforce. We
[20:28] want to get them into a single pane of
[20:29] glass so you can view all of your agents
[20:32] within the Gemini Enterprise app. Admins
[20:34] can discover, manage, invoke any any and
[20:37] every agent across the workforce. Every
[20:40] employee can see the agents available to
[20:42] them and also initiate sharing following
[20:44] the governance principles our policies
[20:47] laid down by your organization.
[20:49] Now, in practice, agents need tools to
[20:51] do real work, right? We allow you to do
[20:53] that by bringing in your own MCP
[20:55] servers. B by your MCP is you know,
[20:58] public preview as of today. Connect them
[21:00] via our low-code, no-code agents into
[21:03] the designer that Jamie showed just now
[21:05] and instantly make them available
[21:06] through the Gemini Enterprise app.
[21:08] Fourth is agent identity. Just like
[21:10] humans, agents need clear identities and
[21:14] sort of like role-based access control
[21:16] associated with it.
[21:18] We give admins and users complete
[21:20] visibility and control over who an agent
[21:22] is, what it's supposed to do, and the
[21:25] associated permissions with it. And
[21:26] finally, you need a point of enforcement
[21:29] to tie it all together. This is where
[21:31] the agent gateway comes in,
[21:33] which lets admins allow and set policies
[21:36] within the GCP console and then enforce
[21:38] it for all ingress and egress traffic.
[21:41] So, if an agent goes rogue, it cannot
[21:43] exfiltrate any data outside of your
[21:45] organization.
[21:47] Digging in a little bit more into agent
[21:49] observability, uh
[21:51] agent observability includes audit
[21:53] logging, overview metrics, and agent
[21:55] tracing. Here's a quick glance into
[21:57] audit logging.
[21:58] Um
[21:59] and more importantly, what you get with
[22:01] this is full hotel compliance, which
[22:03] means you can export all of this data,
[22:05] these metrics, and logs into your
[22:07] favorite observability system and
[22:09] analyze them further.
[22:11] Uh a quick look here tells you like the
[22:14] the sessions you've had for a given
[22:15] agent, the number of turns per session,
[22:18] how many invocations, the traffic
[22:20] patterns, input-output token
[22:21] distribution,
[22:22] so on and so forth.
[22:24] Um
[22:25] on the agent management front, I think
[22:26] the big callouts here are
[22:29] effectively for agent security and
[22:31] registry. Let me focus on identity here
[22:34] a little bit more than the registry cuz
[22:36] Jamie sort of covered the registry
[22:38] pieces in his uh
[22:39] in his slides. Uh we're deploying GCP
[22:42] wide identity framework, which is based
[22:43] on SPIFFE as the standard. This is a
[22:46] huge differentiator against any
[22:48] competitive product in this space.
[22:50] Essentially, it gives you the security
[22:52] you need to trust running your agents
[22:54] across organizations with access to all
[22:56] of your data. It allows you to
[22:58] differentiate between Dave's agent, what
[23:01] Dave's agent did, meaning something the
[23:03] agent did on my behalf versus me doing
[23:05] something directly, right? Uh
[23:07] additionally, the ability to leverage my
[23:09] access or access on my behalf is
[23:11] ephemeral, which means after the session
[23:13] goes away, the access goes away and, you
[23:15] know, everything is closed and gated
[23:17] down after that point.
[23:18] Um
[23:20] Let me jump over to agentic security a
[23:21] little bit more. Here, we've released
[23:23] two sets of features and products or two
[23:26] sets of integrations into other product
[23:28] portfolios. First is Model Armor, which
[23:31] is already GA.
[23:32] Model Armor protects you and your
[23:34] application from prompt injections,
[23:36] jailbreaks, malicious content, bad URLs,
[23:39] stuff like that. It plugs directly into
[23:41] the gateway, which means it monitors all
[23:44] ingress and egress traffic in line with
[23:46] your company's policies. Similarly, SDP
[23:49] or sensitive data protection gives you
[23:50] PII
[23:52] uh detection and prevention policies
[23:54] right into your workflow so you can mask
[23:56] sensitive data uh before it crosses uh
[23:59] the client side.
[24:01] Jumping over a little bit more, how does
[24:03] the how does it all get together into a
[24:06] single point of enforcement? That's
[24:07] where the agent gateway comes into the
[24:09] picture.
[24:10] Uh agent gateway is already in preview
[24:12] and it serves as a single point of
[24:14] policy enforcement. Architecturally, the
[24:16] gateway sits between any client or your
[24:18] client, uh such as the Gemini Enterprise
[24:20] app, uh and your backend resources,
[24:22] which could include your models, your
[24:24] APIs, your tools, B by your MCP, so on
[24:27] and so forth.
[24:28] Its core function or functionality is to
[24:31] give you the ability to author policies,
[24:33] apply them in order to secure, govern,
[24:35] and monitor any agent to external system
[24:37] interactions.
[24:39] Whether an agent is communicating with
[24:40] another agent, executing a tool, or any
[24:43] kind of APIs outside of the GE
[24:45] environment, the gateway ensures that
[24:47] access authorization, security policies
[24:50] are consistently applied for all ingress
[24:52] and egress at a single request level.
[24:55] It also gives you the ability to capture
[24:56] logs, traces, and metrics at the gateway
[24:59] level itself, so you know what's
[25:00] happening and these are all pulled
[25:02] together into a single view in Pantheon.
[25:05] Um
[25:06] Now, let's connect all like I'm going to
[25:07] skip over to the demo now, which is
[25:09] going to be fun because I want to show
[25:10] you how all of these tie into your agent
[25:13] development life cycle.
[25:17] And thankfully, the session has not
[25:18] timed out. I see you didn't put it on
[25:20] the right tab, are you? All right. Thank
[25:21] you, sir.
[25:22] All right, perfect.
[25:25] As you can tell, this is Jamie's laptop,
[25:26] so I've no idea how to operate it. All
[25:28] right. So,
[25:30] when I
[25:32] All right. So, when I think about the
[25:33] agent development life cycle, right? I
[25:35] think about it in terms of like
[25:37] four categories. Category one is really
[25:40] you want to build an agent, you want to
[25:42] scale the deployments as in like scale
[25:44] the agent which is deployed in
[25:45] production. You want to have governance
[25:47] and control over what the agent is
[25:49] supposed to do, and you want to optimize
[25:51] over time to make sure there is no
[25:52] regressions, data skews change, traffic
[25:55] patterns change, your external
[25:57] integrations change. So, you want to
[25:58] make sure that what is in production
[26:00] keeps on running and humming along
[26:02] smoothly. So, that that's how like the
[26:04] agent platform is designed into these
[26:05] four phases: build, scale, govern, and
[26:07] operate. And we're continuing to add
[26:09] more and more features in each of these
[26:10] categories. Um
[26:12] in in the build phase, I think Agent
[26:14] Garden is is like the point where you
[26:16] start from. This gives you like
[26:18] pre-curated templates and recipes for
[26:20] spinning up an agent. All of these are
[26:22] ADK based, but you could bring in
[26:24] arbitrary recipes that your company may
[26:26] want to have into this and they don't
[26:28] need to be just ADK, they can be
[26:30] whichever framework of choice for
[26:31] authoring an agent. Um I personally like
[26:34] uh the LLM auditor and the travel agent.
[26:37] It's kind of nifty. I use it in personal
[26:38] capacity. Uh I also like the chatbot
[26:41] agent which gives you like a nice
[26:43] starting point to to build a chatbot.
[26:47] Skipping forward to what it means in
[26:49] production side, right? If I go, I can
[26:52] see all the deployments of the agents
[26:53] I've done. Seems like someone's been
[26:56] enthusiastic to spin up an agent 2
[26:57] minutes ago, which is awesome, perfect
[26:59] time for a demo.
[27:01] Uh but here you can see like a bunch of
[27:02] agents that are in production. Let me
[27:04] see what the travel orchestrator agent
[27:06] is doing. Uh you have a bunch of details
[27:09] when you click through the agent here.
[27:11] It gives you when it's created, it gives
[27:13] you the name, where it's deployed, so on
[27:15] and so forth. I can tell it's an A2A
[27:17] agent by the the card here in terms of
[27:19] the protocols and the minute. Uh and I
[27:21] can also tell that it has an external
[27:23] OAuth app associated with it. Uh it's
[27:27] leveraging three skills: Confluence,
[27:29] Trello, and Loom.
[27:30] Uh so, now you kind of get a picture of
[27:32] like what is all connected into it. You
[27:34] have a protocol, you have the skills
[27:35] associated, you have where it's
[27:37] deployed, so on and so forth.
[27:39] Um I personally like to
[27:41] you know, jump over to uh
[27:44] some of the
[27:45] observability elements of this where I
[27:47] can see all of the observability metrics
[27:49] in a single tab. Uh I can see my average
[27:51] latency, I can see agentic traffic.
[27:54] From what I can tell, this sawtooth wave
[27:57] represents a scheduled agent or a
[27:58] trigger-based agent. So, someone's
[28:00] running it ever so frequently, probably
[28:02] to figure out what the best pricing for
[28:04] a given thing is. Um and then, you can
[28:07] also see what the execution semantics of
[28:10] this uh agent are, what frequencies, so
[28:12] on and so forth. That's actually done
[28:14] via the session management tab. Uh so,
[28:16] session essentially represents an
[28:18] interaction between you and the agent
[28:20] itself. Like we basically keep track of
[28:23] the conversational history that you've
[28:24] had with the agent, any actions that
[28:26] your agent has taken
[28:28] in that course during the course of the
[28:29] execution, and we also allow uh we also
[28:33] retain memory for long-term agent
[28:35] execution, so we don't miss context in
[28:37] the subsequent turn or in the subsequent
[28:39] iteration of the session, right? Um
[28:42] Now, you could say like what is a
[28:43] memory? A memory is effectively a custom
[28:46] recipe that you might give in addition
[28:48] to us saving your prompts, histories,
[28:50] the agent outcomes, so on and so forth.
[28:52] I could say, you know what, I want to
[28:53] personalize my travel agent by only
[28:55] looking at memberships that I have with
[28:58] you know, my frequent flyer cards and
[29:00] tickets and so on and so forth. That can
[29:02] be modeled as a memory. It learns that,
[29:04] you know, Dave prefers to fly whatever
[29:06] Alaska Airlines. Um so on and so forth.
[29:10] Uh let me skip forward to
[29:13] the governance side of it where you get
[29:15] to see all of your agents in one view.
[29:17] Some are A2A agents, some are non-A2A
[29:20] agents. Gives you like the endpoint that
[29:22] it's going against. It tells you what
[29:24] the runtime is. For example, the
[29:26] internal ID assistant is on the agent
[29:28] platform runtime, and the marketing
[29:31] copywriter is a direct Gemini Enterprise
[29:33] agent as in it was created from within
[29:35] Gemini Enterprise. Uh something called
[29:37] the agent one seems directly on GKE. I'm
[29:40] going to assume it is still an A2A
[29:42] agent, but probably the underlying uh
[29:45] authoring framework could be like
[29:46] LangChain, LangGraph, what what have
[29:47] you. Um
[29:49] And that that sort of gives you like a
[29:50] perspective on,
[29:53] you know, what the registry is, how it
[29:54] ties back into the gateways, how it ties
[29:56] back into other associated policies, and
[29:59] how you build, deploy, govern and scale
[30:01] over time. All right?
[30:03] That's That's me. I am going to hand it
[30:05] over to our awesome partner from KPMG,
[30:08] Aaron, now.
[30:09] And they will walk you through their
[30:11] experiences with G.
[30:14] >> [applause]
[30:16] >> Thank you, sir.
[30:17] Awesome. Thank you.
[30:21] Hopefully y'all can hear me. Um thanks,
[30:23] Dave. Uh
[30:24] really excited to be up here. Appreciate
[30:26] Google inviting us. We're going to talk
[30:28] a little bit about our journey at KPMG
[30:31] uh with a I almost said Asia space. Do
[30:33] you want to enterprise?
[30:34] Uh it's been about 7 months. So, we're
[30:36] going to talk a little bit about our
[30:38] approach to Gemini Enterprise.
[30:41] First of all, just to give you a little
[30:42] bit of a high-level architecture
[30:43] thinking. Um we we look at Gemini
[30:46] Enterprise as a concierge layer.
[30:50] And we broke it down into really three
[30:51] components. Find and understand. So, the
[30:54] data fields or the data connectors.
[30:56] Being able to find information,
[30:58] understand information with Gemini.
[31:01] Then act. We're moving into the act
[31:03] phase.
[31:04] Being able to take action. When When I
[31:06] see something, when I find it, I
[31:07] understand it, I want to take action.
[31:09] And being able to do that, as as Jamie
[31:12] pointed out, across our enterprise
[31:13] systems. So, you can see here, this is
[31:16] how we're thinking about Gemini
[31:17] Enterprise and being able to use it to
[31:20] manage our enterprise space.
[31:23] The other thing we think about with
[31:25] Gemini Enterprise is that it's
[31:26] ultimately going to help us abstract a
[31:28] lot of things that are going on. A lot
[31:30] of us are going through digital
[31:31] transformations. A lot of us are going
[31:33] through journey to the cloud. A lot of
[31:35] us are dealing with the constant influx
[31:37] of agents. The idea of being, can we
[31:39] abstract that with Gemini Enterprise and
[31:42] be able to make it simple for the user?
[31:44] They don't need to know what agent it is
[31:45] they're talking to necessarily because
[31:46] Gemini Enterprise is invoking that
[31:48] agent. They don't need to know what
[31:50] system they're talking to. If you think
[31:52] about the previous slide, we've kind of
[31:53] abstracted that for them. So, ultimately
[31:56] they are just interacting with Gemini
[31:58] Enterprise to find, understand, and take
[32:00] action on a variety of different things
[32:02] that they're looking to do. Reducing the
[32:05] cognitive load off of folks, so they
[32:07] don't need to know, hey, I I just
[32:09] transitioned to a new SAP system, I've
[32:11] transitioned to a new Salesforce system.
[32:13] The idea of being that they're just
[32:14] interacting with Gemini Enterprise and
[32:16] at the end of the day, they're simply
[32:18] simplifying those interactions.
[32:24] So, give you a little sense of what our
[32:27] stats look like over the last 7 months.
[32:29] So,
[32:30] we rolled this out in September to about
[32:33] 38,000 users. As you can see, that
[32:36] second row, our adoption rate was about
[32:38] 90%. You heard a little bit about that
[32:40] on the keynote this morning.
[32:42] We've since extended it to some of our
[32:44] offshore resources.
[32:46] Roughly about 50,000 now total. Um we're
[32:48] increasing those numbers, but you can
[32:51] see we're rolling this out, the adoption
[32:52] is super high, it's super sticky, and
[32:55] we're really excited about the fact that
[32:57] that strategy is starting to take hold
[32:59] across the US.
[33:02] This gives you a sense of what we've
[33:03] seen in volumes since we rolled this
[33:05] out. Uh about 28,000 prompts a day.
[33:09] Uh we're These numbers are actually
[33:12] already been exceeded because it's a
[33:13] couple months old. Uh we're we're almost
[33:15] at 40 45,000 notebooks created.
[33:18] Um we've seen about 15,000 low-code
[33:22] agents get created. And we've also
[33:24] shared about 800 agents. This shows 600,
[33:27] but we've seen an increase to about 800
[33:30] agents being shared using the shared
[33:31] feature. So, we're really promoting that
[33:34] organic innovation. People building
[33:36] their own agents. You saw a lot about
[33:37] those agents, but being able to promote
[33:39] sharing of those agents, so we can
[33:41] capture a lot of that organic innovation
[33:44] activity coming from the ground up.
[33:47] This will show you a little bit about
[33:49] the prompt usage. We're trajecting up.
[33:51] Uh again,
[33:53] uh a little bit outdated, but you can
[33:54] see that that trend is moving to to the
[33:56] right place for us.
[33:58] Seemingly giving us a lot of that value
[34:00] and stickiness from from that Gemini
[34:02] Enterprise experience.
[34:05] Then we focus a little bit on some of
[34:07] the value. Uh for us, the value is seen
[34:10] in really two key areas. Number one,
[34:12] time saved in finding information. You
[34:14] think about that slide we started with,
[34:16] find and understand. This gives you a
[34:18] sense of we're able to really find that
[34:20] information and we're also ultimately
[34:22] able to uh decrease the amount of time
[34:25] people are out there looking for things.
[34:27] The other thing is is that the sentiment
[34:30] metrics is super high. Uh people are
[34:32] very excited about it. As you can see,
[34:34] they have confidence. They are
[34:36] definitely finding it useful, usable,
[34:39] and they have a high satisfaction rate.
[34:41] Uh so, this is really giving us a sense
[34:43] of what the the user sentiment is. Not
[34:46] only how much they're using it, but how
[34:47] much they enjoy using it and how much
[34:49] they have confidence in their use of it.
[34:54] Real-world application.
[34:56] >> [clears throat]
[34:56] >> We talked a little bit about organic
[34:58] innovation and some of the things that
[35:00] Jamie and and and Dave talked about. For
[35:02] us, this is a real-world example. We had
[35:04] someone go out and build an agent around
[35:07] IP risk. And in this particular case,
[35:09] the idea was we build marketing
[35:11] material, build a variety of different
[35:13] things that go out into the marketplace.
[35:15] Using this agent and the orchestrated
[35:16] functions in the agent designer to be
[35:19] able to create stages, steps, sub-agents
[35:23] to be able to manage those agents across
[35:25] the board. But ultimately being able to
[35:27] produce a document review agent that
[35:29] looks and finds any IP infringement,
[35:32] looks and finds any risk in content. So,
[35:34] ultimately the the people producing the
[35:37] the PDFs, producing the the marketing
[35:40] material, at the end of the day,
[35:42] validated whether or not there was any
[35:43] IP infringements or risk with releasing
[35:45] that data out into the market.
[35:49] So, in conclusion, some key takeaways.
[35:52] AI is a core team member. As you can
[35:54] see, for us, we really want to build
[35:57] from the bottom up, automate the work we
[35:59] do,
[36:00] uh be able to drive that organic
[36:01] innovation.
[36:03] That unified platform. Trying to create
[36:05] that abstraction. Being able to
[36:07] ultimately take it to a place where you
[36:09] no longer have to worry about what the
[36:11] systems are underneath. And then lastly,
[36:14] you know,
[36:15] find, understand, act. Create shorter
[36:17] timelines to be able to take action and
[36:20] be able to measure that that capability.
[36:24] All right. Now, I'd like to turn over to
[36:25] Lisa uh from Signal Iduna.
[36:30] >> [applause]
[36:36] >> I am very looking forward to
[36:38] use all the new features, I think, as
[36:41] you also do. But let's do a reality
[36:44] check how it's going with Gemini
[36:46] Enterprise in a large and old insurance
[36:50] company, which maybe is not so
[36:53] innovative sometimes.
[36:56] So,
[36:58] last year, exactly 1 year ago, we
[37:00] decided to give Gemini Enterprise a
[37:02] chance.
[37:03] And 6 months later, last October, we
[37:06] went live with Gemini Enterprise. We
[37:08] went live with 11,000 people, so all
[37:11] over employees.
[37:13] Why everyone? Because we believe in our
[37:16] strategy, first use, then case, and then
[37:20] value.
[37:21] Because we believe everyone in a company
[37:24] should use AI, should learn how to
[37:27] integrate AI in their daily working
[37:30] life.
[37:31] And then personal productivity leads
[37:35] automatically into use cases,
[37:38] for example, into agents, and into
[37:40] value.
[37:44] Last October, we had our big marketing
[37:46] campaign.
[37:47] You can see it there. You can see our
[37:49] office.
[37:50] And
[37:52] it was
[37:54] our biggest marketing campaign and a
[37:56] change management. So, to everyone to
[37:59] get everyone on board, we had a big day,
[38:03] but it's an ongoing journey. And this
[38:05] ongoing journey you can see in the
[38:07] numbers, and I think they speak for
[38:09] themselves. We have 80% monthly active
[38:13] user.
[38:14] We have 12 high-code agents that
[38:17] everyone in the company can use.
[38:21] And we have over 100 AI ambassadors. And
[38:24] these AI ambassadors are really really
[38:27] crucial for our AI adoption in our
[38:30] company.
[38:33] But a very good example for our success
[38:36] in Signal Iduna is our custom-built
[38:39] health care agent,
[38:41] which supports our employees with every
[38:44] health related inquiry from our
[38:47] customers. So, it's
[38:50] an custom-built agent. It's embedded in
[38:53] Gemini Enterprise, and it uses our very
[38:56] very large
[38:58] and sometimes old health care related
[39:02] database
[39:03] that every inquiry from customers can be
[39:07] answered. So, the database is used as a
[39:09] point uh
[39:12] uh ground truth to
[39:15] answer every inquiry from point of sale
[39:18] to claims management.
[39:21] And this example shows that the
[39:23] combination of AI agents and employees
[39:27] lead to very high value both on internal
[39:31] side and on the customer face.
[39:34] So,
[39:35] with this agent, our custom-built health
[39:37] care agent, we saw a 400% increase of
[39:41] usage within the first three months
[39:44] after launch, which showed that we
[39:46] really have hit a right spot
[39:49] for our employees.
[39:51] Also, the time answering our client
[39:54] injuries went significantly down. We're
[39:56] now
[39:57] 7 37% faster. And last but not least,
[40:02] the transfer rate went down to only 3%.
[40:06] So, an employee don't has to ask someone
[40:10] else.
[40:11] He can chat with our agent and get
[40:13] quickly an answer.
[40:15] So, if you want to take one point from
[40:18] my talk, please involve everyone you can
[40:22] possibly get in our AI journey that you
[40:25] can first have use and then case and
[40:28] value.
[40:29] Jamie, back to you.
[40:32] >> [applause]
[40:36] >> Amazing. Thank you. Thank you both.
[40:38] Thank you, Aaron. Thank you, Lisa. I
[40:39] think it's uh great to hear about all
[40:42] the new capabilities, but seeing it put
[40:44] to be practice across an organization
[40:47] with all the challenges that an
[40:49] enterprise has at scale for doing that
[40:51] data in a governed way, in a regulated
[40:54] environment, and with real people that
[40:57] have to change their way of working is
[40:59] is where this really kind of rubber
[41:01] meets the road. So, incredible examples.
[41:03] Thank you both for sharing those.
[41:06] Just to wrap up, there's a couple more
[41:08] capabilities that we announced this
[41:10] morning, and one of those is around our
[41:12] ecosystem and partnerships. We launched
[41:15] a major update to our agent marketplace
[41:19] with key partners like you see on screen
[41:21] here from ServiceNow to Workday to Box
[41:24] to Lovable, everything from startups to
[41:27] uh large systems integrators and ISVs.
[41:30] And employees get access to that
[41:33] marketplace right inside of the Gemini
[41:35] Enterprise uh dashboard
[41:38] where they can then browse those agents,
[41:41] request access to new agents from the
[41:43] marketplace, and your IT team can govern
[41:45] the policy to give them access.
[41:48] So, I'll welcome my team back on stage.
[41:50] You can you can get started now with
[41:52] this. Um
[41:53] there's a number of resources available
[41:56] to you
[41:57] um to leverage these through agents
[41:59] um to leverage context across your
[42:01] organization, to govern and manage it at
[42:04] scale,
[42:05] um and to connect it to to all of your
[42:07] knowledge.
[42:09] And um we have a number of sessions here
[42:12] at Cloud Next this week to continue to
[42:16] dive deeper into these capabilities with
[42:18] Gemini Enterprise, and we'd love your
[42:20] feedback on the session.
[42:24] >> [music]