Grok Bot for Admins

0:02Okay. Um We're going to kick off. So, um my name is Rohan. I work on our Grokbot go-to-market team. Um essentially kind of helping companies figure out how to, you know, bring in Grokbot to their companies, drive adoption, um uh lots of fun stuff there. Pass it over to John. Um my name's John. I'm a product manager at uh SpaceX and I work on Grokbot for

0:26enterprise. Um I talk a lot with uh admins and security teams at uh large companies and uh yeah, looking forward to sharing more about what we've built in in collaboration with our customers today. Cool. So, just um some quick housekeeping. Um we are recording this. Um the the you can ask questions through the Q&A function. Um John's going to predominantly be driving. I will be um uh facilitating uh questions. So, please ask them. We'll hopefully have time at the end for um Q&A. And want this to be interactive in that way. So, please uh please jump in. Uh we'll share the recording and um

1:03um more afterward. Cool. So, uh this event here, uh this is mainly for admins of Grokbots and uh we're going to walk through what you need to think about as an admin as you roll out Grokbot to your uh organization. Um and the topics we'll cover today uh are how you can set guardrails for for what actions

1:26uh your bots are able to do. How you can make sure that your data uh stays secure. Um how we recommend you rolling Grokbot out to your uh team, uh including how you can create shared bots uh that can work across uh uh your team. And then lastly, how you can understand uh usage, uh how your team is using Grokbot, and also what the cost

1:48drivers are. Cool. Um but first, Rohan will give a quick introduction to what Grokbot is. Yeah. Next slide. Yeah, so Who's on the ground to help you bit in in the way we're thinking about Grok bots. So, you know, back in 2022, multiple chatbots were were launched. Well, ChatGPT was launched, soon after Claude was launched. This was really this idea of question and answer back and forth. Move over to agents. This is where majority of products are today, right? You can kind of um uh be chatting, but like kick off a task. It can maybe, you know, send an email, read your Notion docs, send a

2:24Slack message. A bit more of this kind of agentic behavior. Very, very common in engineering use cases, and and lots of others are starting to catch up. Where we see Grok bot is really a massive step change. And so, this idea of instead of you assigning a task and then it does it, it's really this idea of AI teammates, where it can run your work or your team's work end to end and loop you in as needed. And so, this is a big paradigm shift that I want to ground us in. Not just kind of like another version of like chatbot or even agentic behavior, but really thinking about Grok

2:57bot as an AI teammate that can complete work end to end. So, why don't you take us through a couple quick examples of this? So, you know, also like thinking about non-engineering use cases especially. One of one big one is sales inbounding that happens a lot. So, you know, SDRs oftentimes will research the company, check for an owner, looking at the left column here, qualify against an ICP, they'll write a reply maybe with a chatbot, book the meeting, route it to an account executive, and then AE piece together

3:27content before the call. And with Grok bot, what you're able to do is create a single bot. Let's call it a lead bot that does all of this work. So, it picks up each lead, researches it, it scores it against your ICP, it drafts a personal reply, and then it posts a brief and and it sends it proactively to the AE. So, it's doing all of this kind of you know, in an always-on mode, and and it's really kind of doing this end to end work, and uh pulling in the AE as needed. And so, you know, what this means is that today you might have an SDR that's using tools

3:58uh with days per lead, and with Grokbot you actually have it's an AI teammate that schedules leads in minutes, and then the AE can just approve. And if we take this out, this is sort of like a very focused um task, but if we jump to next slide, um what's what we've done internally is actually really, really abstracted this high level. So, we created our own um internal um bot called support bot. What this does is it uh you know, pulls in from plain linear data doc, all our different kind of um sources, and it does many, many different jobs. So, it investigates, it watches the queue, it

4:31actually looks at X signals, um it'll also do some quality review, and it'll take the first stab at actually um solving the issue. And in it'll escalate to a human as needed, it'll pull people in, but it's really doing a lot of this end-to-end work. And what we saw internally is that we were actually able to scale our team 175% without hiring by doing these three jobs, um 2 million salary savings, and the biggest kind of thing here is that there's really no engineering required. It was all sort of creating bots within the platform. Uh so, really want to ground us in thinking about these end-to-end workflows. If we

5:04go to the next slide, um n- not just end-to-end workflows, but end-to-end workflows in a multiplayer uh capacity. So, the vision with Grokbot is, you know, before on the left here, this is now we're jumping to marketing, um you had uh a PM write the brief, then passing it to marketer who drafts the plan, manager, designer, lots and lots of handoffs, and lots of different, you know, um people to to work alongside with. On the right, what you can see here is um a Slack thread. And so, you have Maya who's a marketing manager, kicks off a thread saying iOS app is pushed to Thursday uh launch. Priya,

5:37who's a PM, calls Launchbot, says, "Can you run this end-to-end?" Launchbot drafts three artifacts. Jordan, who's a designer, says, "Make it enterprise ready." And then Launchbot does that, and kind of keep going back and forth. Maya approves, Jordan approves, LaunchBot schedules, and LaunchBot proactively comes back two days later saying, you know, oh, hey, there are

5:5542,000 impressions. Um and so, this is really the vision that we have of GrokBot of, you know, multiplayer kind of folks collaborating with each other. We have this view in both Slack and soon Microsoft Teams. Um really really the vision that you want to be thinking about. And just want to kind of um showcase

6:13since we're not doing a full demo here. This is sort of the interface of GrokBot. Um on the right, you can, you know, create bots for different jobs, message bots like teammates. And the key here is that bots keep context and memory. And then I think the two things I really want to call out are um this top-right functionality, which is the ability to share bots with your teams. Uh so, you can create these idea of um bots that are shared across your team, shared context, shared memory, and everyone can kind of message them. And, you know, you as a leader or an admin can actually set

6:42them up. And then also setting up automations and routines. And to what I called out about these end-to-end workflows, this is really critical of, you know, being able to think about these end-to-end workflows and set up um automations and routines. Um cool. I'll pass it back to John to take us through some admin controls here. Yeah, cool. Um just before we dive into the the admin and security controls, uh I wanted to give a quick overview of the architecture of of GrokBot. So, uh there's three main parts there. Um you have the the client, uh which is either a desktop or a mobile app. That's it's

7:16on the left side here. This is a thin client. So, so GrokBot runs completely in the cloud. There is no uh nothing happening locally on the computer or phone here. Um as an admin here, you can define uh team policies uh uh that I'll walk through a bit more later. And if we look at what's happening on the cloud here, uh each user gets their own VM that the [clears throat] uh bots can use. So, that's a Linux computer that can browse the web. It can run applications and you can do work as

7:47you would on a regular computer. Um, this VM here is hosted in the in the US and you can define different network policies here to be able to control what destinations are reachable from the VM. So, this would typically be that that the the bots can use the computer to to to both reach

8:06websites but also SaaS apps on the web. Um, and then you have the the the agent harness here. The main agent loop. Um, this calls model providers. Here we provide zero-day data retention. Um, and this agent harness also can connect to MCP servers. So, this would typically be the tools that you're already using

8:32like Slack, Gmail, GitHub, and so on. The tool calls here you can define which tool calls are allowed. And we also have the concept of auto review that runs on every bot action. I'll come back more later on how how you define these rules as well. But I think the key thing to understand here is that Grokbot is cloud-only and you have a VM or a computer that the bots are able to to use in addition to the MCP servers

9:00here. Cool. The first thing I want to talk about is how you can set the guardrails for what the bots are able to do. This would typically be the first thing you do as you start setting up Grokbot for your organization. Um, and the main three things here is to decide what what plugins and MCPs the the bots should be able to to use, whether the bot should be able to do work on the user's local computer, and

9:29what network destinations are reachable. Um, for plugins and MCPs, um, Um, if you're already using Cursor, you probably have defined some uh plugins and uh and uh custom MCP servers for your team. These are um available in Grokbot as well, so you don't need to do any more setup here. Um, you do also have the option here to go into each uh MCP server and define which tools uh are allowed. So, for example, you can disable destructive tools in uh the

9:57Google Drive uh MCP. Um, for local uh execution here, um it can be useful uh for the bot to be able to access local files, um but you can uh uh disable it if if if you don't want uh users to do that. Um, and lastly here, uh the Grokbot computer has a browser and it has a terminal, so if you want to limit the uh destinations of this computer, you can set a network allow list, uh which means that all uh network calls from the computer and MCP tool calls uh will uh follow this

10:33network allow list. The uh second set of features that we have for uh for controlling bot actions is rules and hooks. Um, one key thing to mention here is that all sensitive actions uh do require user approval, uh and this is automatically evaluated by a review model that we have that we call auto review. So, this will always run on uh on uh on all uh bots, but you have the ability here to add more rules and

11:03configuration uh if you wish to do that. So, the first uh category here is what we call team rules. So, these are team-wide rules that apply to all the bots uh within your organization. Um, these are injected into the bot's system prompt, so they will always be present in the context. And an example here of a team rule could be that never send an email without uh user approval to an

11:27external company, for example. Um, the second set of rules here is the auto review rules. So, as an admin here, you can enforce auto review on all bot actions. So, this is both responses, MCP tool calls, and also things that the bot is doing on the computer. Um, and you have the option here to choose between some actions that should be allowed automatically. So, kind of like loosening the the the policy here, or you can enforce that the bot will always ask the user before they do a certain

11:59action. Um, yeah. And then lastly, we recently launched something we call Inference Hooks. Um, so this lets you connect Grokbot to your own endpoint. So, this would typically be a security software like Palo Alto Networks or Zscaler. And here, every user prompt is being sent to your destination, which either returns that it's allowed or denied before anything is sent to to to the model provider. Um, so this means that you can both apply deterministic controls like regex matching, for example, as well as having inference run on your endpoint to determine if if an action is is allowed

12:46or not. Um, this is currently in in early access, but we're happy to to enable this if you reach out to your account manager. Cool. The second thing here is how you can make sure to keep your data secure. Um, so the the bot computer here is a Linux VM, and we give you the ability to configure it with your own tools. So, this can be, for example, your company VPN or EDR

13:16software like CrowdStrike. So, you have the concept here of team setup scripts that you that will always run when the VM boots up. Um, typical use cases here is that install a VPN or EDR, and you can also define secrets here that should be used to configure these tools. Uh, these will never be made available

13:35to the bots. Um, and for the the bot computers here, you do have the ability to both uh force update or recreate or uh wipe this these VMs, and you can do that in bulk. So, for example, if you install a set of scripts here, you can uh you can force restart all the VMs in your organization. And as mentioned here, you also have full control over uh what destinations the computer can reach

14:00through network controls. Um, on the security side, we also provide monitoring uh of uh of bot activity here. Um, you have audit logs that contain all the human actions done in the product, but we also have the ability to set up auto exports, uh which will include all the bot actions, including prompts and responses, and MCP tool calls. And these would be exported to your OpenTelemetry

14:27collector. Um, as mentioned here, you can also easily manage uh the VMs in your organization, and uh you can also set up retention policies for the VMs, for example, to uh automatically delete the VMs after 30 days of inactivity. And lastly, we provide an uh an admin API here that lets you programmatically uh apply these actions. So, this opens up opportunity to, for example, have your auto export uh looking for certain uh actions, which then can trigger a uh

15:00cleanup uh action on on the VM side. Cool. Um, then um how do you roll out Grokbot to your team? Um, so for uh Team plan customers, Grokbot is enabled by default, uh while uh on the Enterprise side, uh you as an admin need to enable it manually. Um, we have onboarded a lot of customers over the past months now, and uh, what we're seeing, uh, works well is that you typically want to start with a smaller group as you set up the the guardrails and and the security policies. And here we offer the ability to, uh, give access only to certain member groups before, uh, before you roll it out to the the

15:46whole company. Here you can either use, uh, SCIM groups that already exist, or you can, uh, create a pilot group where you add, uh, uh, a smaller amount of people. Um, on the on the access side, uh, all users with a Coda receipt automatically has access to Grok Bot as well. Um, but if you want to invite, uh, a larger part of your org, you would need to, uh, provision, uh, seats for these people. Um, and after you've done that, um, you can invite the full team. So here you can, uh, both share a invite link that, uh, you can share on Slack, for example, that works for all users. Or you can

16:23send invite links by by email to to new users. [snorts] Uh, last thing I want to touch on, uh, when it comes to to, uh, introducing Grok Bot to your org is how you can use team bots, uh, to make it simpler to, uh, to scale automations and and bots in in your org. Um, and the way we think about this is that instead of having a hundred people in your org uh, setting up the same automation with their own bots, what if you can create a shared bot that all of these hundred people can use, and, uh, then you can have one person managing this and making sure that it's, uh, that

17:02it works great. So, um, what we're seeing here, uh, is that typically you would have a a, uh, functional leader, uh, that that we partner with to design these workflows. So, uh, Rohan mentioned here in in in support and also in sales uh, where we've seen a lot of interesting use cases. And here the the functional leader would

17:25build this team bot. So, this is a combination of access to tools, skills, and instructions. And after a team bot has built that has been built, you can share that with the whole org or certain members. And as an admin here, you can also select a team bot and say that this should be the default for the whole organization or a group of members. So, as an example, if you create a deal desk bot that is relevant for sales people, you can automatically share that with the sales org and make sure that every new user getting started with with Grokbot from the sales team has access to this bot

18:07right when they start using the product. Cool. Last thing to touch on here is how you can understand usage and cost. So, we provide two sets of capabilities here. The first one on usage is something we call usage analytics. Here we go through all the tasks that your team is is using Grokbot for and we categorize them within a set taxonomy. As an example here, you can see here that that internally here sales is actually the most active user

18:41of of Grokbot. Engineering and products come second here. We also have a measure here on the level of automation in in the tasks that Grokbot is using where it goes from not automated which is more like the chat bot. You ask something, the Grokbot responds. Semi-automated where the human approves something before the bot does an action. And lastly, fully automated which is a workflow that would run uh, On spend spend management, you can also track Grokbot spend by team, uh, group and user, and you can set spend limits both per team and user, um, including separate control controls for

19:22on-demand, uh, usage. And and one more thing on this is that we're exploring more ways to kind of, um, showcase usage and and uh, ROI as well. We're exploring ways to maybe show things like, you know, hours saved or spend. So, constantly trying to to improve this and make this better as it's a really really critical. For sure. We have a, uh, we have a prototype internally now that actually calculates the time saved using a Grokbot, and we can match that to your spend. So, we're working with some

19:49design partners on this currently now. Yep. Cool. Um, Awesome. Yeah, let's jump into Q&A. So, um, there are a couple that have come up, um, quite a bit. So, the first is, would be great to cover memory for Team Bot, um, just to kind of like help set expectations there. And, um, there was another one that was similar, uh, very similar question. Um, yeah, does a bot context, uh, window chat session compact uh, upon compact as bloated? Just like generally talking

20:18about memory and token burn there. Yeah. So, uh, generally, um, since there's a limited context window in in models now, um, we do have, uh, context compaction. And what we do here is that we turn, uh, um, user context, uh, as well as actions a bot has performed into, uh, uh, a memory. So, uh, you can imagine that the memory is a set of events and facts and things that the the bot has learned

20:46about you. Uh, and this will, uh, grow over time, but as a user, you don't really need to think about, uh, about how to manage context with Grokbot, which uh, is one of the the really nice things about using the the the product. Um, I have bots that I've I've used every day, uh, since early July. Uh, I have not had any problems with these bots remembering facts or or doing something different than what

21:12I've told them earlier. Um and there was one question about about team bots as well with memory. Yes. Yeah. That'd be great. Yeah. So with with team bots, there's two sets of memory here. So it's the shared memory and it's the per user memory. So a team bot can be used both in the context of Slack for example. You add this bot to a Slack channel and multiple users can talk with it. Um this would then go into the shared memory and you also have the option to directly DM your team bot or you can also talk to the to the team bot via the Crocbot app. This would also go into

21:52your personal memory. So there's there's a clear separation here between memories. There's a lot of questions just more generally about team and shared bots and just for for context we're making a bunch of changes here. So by next week we're going to have a some some some more updates. But maybe for now if you can talk a little bit about just like how does like permissioning work with team bots? Can can folks like add different people to it and specifically

22:17service accounts as well came up. Yeah. So you can you can configure team bots with with service accounts but by default a team bot will use the personal connectors of of the user that is interacting with the team bot. So that means that you don't lose control over what a user can can make the team bot

22:37do. Um when it comes to to sharing and access to the team bots, you have full control over which users are able to to talk with the team bot and you can also add managers to the team bot that in order to have multiple people for example setting up skills or changing instructions or setting up more tools

23:01that the team bot should use. Cool. Um lots of I already answered this one, but I'll just mention again, lots of questions about Microsoft Teams support. It's coming. We're working on it. We're beta testing it internally right now. Can't give a specific timeline, but it's coming out. I'll just say like next couple weeks is

23:22probably the the time I can give. Um let's see. Um lots of questions. Just trying to sift through some. Um yeah, maybe you can talk about like personal accounts and team accounts a little bit more. Yeah. With your team and bot thoughts. Yeah, so um you can So in the sidebar of Grokbot, you can have both your personal bots, but also the team bots. So in many ways, the only difference between a team bot and a personal bot is that the team bot is shared with

23:59multiple people. Cool. Um and then sorry, just getting going through these. Yeah, so can individual spend insights attribute which my Grokbot's are associated with the usage lens? Not currently now, but it's something that we're working on right now. So we should have more on this coming later. Cool. What will happen when limit for spending

24:25is reached? I can answer this one. You basically you will either have a slow pool for for Grokbot or you may not be able to use it. So you basically would have to go if you're an admin, you can just up your up the credits. If you're not an admin, you basically would go to your admin and would need to increase it. This is typically monthly spend limit. So the other thing to do would be to wait till the end of the month. Typically would just go to your your admin to to

24:53increase it. Yeah, it really depends if if the admin has enabled on demand spend or not. Uh and the admin can also control monthly spend limits. So, yeah, it depends a bit on on what the setup in your organization is. Cool. Um I know we don't have any answer for this, but I'm going to bring it up. Um I just cuz I know there's a lot of questions around it. Um a lot of folks are asking about kind of, you know, self-hosted private workers, just like

25:16running it on their own infrastructure. Um I know we're thinking about it. Uh don't have a timeline here, but maybe you can share more about like our thinking and philosophy here. Yeah. Um our thinking here is that Grok Bot should be possible to deploy at any company, uh including the most uh security conscious companies. So, we're actively working and talking with customers right now to figure out the the right shape of uh deployment options

25:39that we provide uh for Grok Bot. Cool. Um let's see. What is the What is the relationship between Grok Bot and Cursor? Uh so, uh Grok Bot and Cursor are both products of the the same company, uh Space X SI. Um there are some really cool integrations also between Grok Bot and and Cursor. So, if you're using uh cloud agents with Cursor, you can just ask your bot to to spin up a cloud agent. So, uh a very cool engineering use case area is to to actually use Grok Bot to to manage your your coding agents. And uh and Grok Bot can also give you links to

26:21the PRs as well as preview images or videos of the work that the cloud agents are doing. Yeah, and um I think like again, we've totally seen great use cases in Grok Bot for engineering. I think Grok Bot is especially good for for non-engineers. So, I want to make that point clear. Like the the top kind of use cases we've seen are around sales, marketing, operations, finance, um HR, IT, recruiting. Like these are really kind of the bread and butter of Grok Bot. Um we really see this like massive step change up from engineering to Sorry, from non-engineering in in Grokbot. A lot You have tools like

26:58Cursor and other companies that have launched these really, really great agenting agentic coding tools. And so, the jump from that to a Grokbot may not be as big. Still very valuable for for folks if they want to, you know, kick off Cursor Cloud Agents, for example. Uh but really, really encourage folks to to be thinking about, you know, these other domains, especially given that, you know, the tools that exist today are have existed are are not the same, you know, level of like agentic and and kind of AIT made

27:26behavior. Um cool. I know we're coming up on time. Let's see. There's a bunch of questions. So, I'm just kind of sifting through these. Give us 1 second.

27:47Yeah. So, the question here is it's possible to have team rules based on departments within the company. Um so, all of the controls that I talked about can also be configured on a group level. So, that means that you can have a scheme group that that includes members from a specific department, and you can set different rules for for that group compared to

28:09what you set to your team overall. Cool. I think we answered this one. Um let's see. That one. Sorry, I'm just looking through. Um

28:38Cool. Is there a library with the latest tutorials, videos, questions? Yes, so if you go to Grokbot, um, you can look up kind of our guides online that have a lot more information. You can also look up use cases. If you go to x.ai/bot, that's where you'll have kind of all this information. Highly recommend using those. We're going to keep launching more tutorials and guides, template bots as

29:03well to to help you get started. So that's that's going to be there. I think that's a good note to to end on. We're constantly shipping. We're shipping very fast. Um, so highly recommend just you know, continue to come to these workshops, but also just keep your eye out for for new launches and we're really excited for you all to both try out Grok Bot if you haven't and to bring it

29:27to enterprise. Cool. Thank you. Thank you for joining. Thanks so much.